Then I sat down to actually do this and realized that it would be a giant pain in the ass. I knew this little side quest was technically possible; I just wasn’t sure the analytical payoff would be worth the investment of time and skill to actually execute it. So I abandoned my vision. I downloaded a few PDFs of these transit tables into a folder on my computer, where they sat for the next fourteen years.

Every historian has these kinds of orphaned sources and abandoned side quests. Maybe it’s a collection of petitions you never had time to transcribe or a city business directory that you never got around to organizing into a spreadsheet. They often get left behind because of a basic calculus: is the juice worth the squeeze? Will an uncertain analytical payoff be worth the hours it would take you to transcribe the source or the money it would take to pay a research assistant to do it for you? Even if you did manage to get that directory into a spreadsheet, what then? Maybe you’d want to look for spatial patterns by overlaying all those businesses onto a historical map of the city. But that would mean learning how to use ArcGIS or finding a way to hire or work with someone who already does – all with no guarantee that you’ll find anything of interest. So you move on.

These two bottlenecks – the time and cost of transcribing sources and the technical skill to analyze or visualize them – have traditionally closed off many avenues of historical research before they ever began. In 2026, that’s changed. Projects that would have once required hundreds of hours and some combination of technical training, research funds, collaborators, and research assistants are now within reach for an individual researcher with minimal funding and little to no technical training - and they can be done in a fraction of the time. The reason for this shift? An uptick in the “reasoning” capacities of Large Language Models (LLMs) and a new generation of AI coding agents.

To show what I mean, let’s return to my abandoned mail transit tables that had been sitting on my laptop for the past fourteen years. I recently pulled them up to see if generative AI could help me finally complete my side quest. Over the span of about a month, spread across a half-dozen 1-2 hour sessions, I used Claude Code and Gemini to go from a folder of PDFs all the way to a fully working, interactive web visualization: “How Fast Was the Mail?”

The two bottlenecks that had caused me to abandon this source all those years were no longer major barriers: I didn’t transcribe any data by hand or write a single line of code myself during this process. In the rest of this post, I want to walk through how generative AI tools handled these bottlenecks and some of the larger implications for historical research.

The Transcription Bottleneck

The first step in most digital projects is to take a historical source and turn it into machine-readable data. Traditionally this has required either transcribing it by hand or using Optical Character Recognition (OCR) tools to try to transcribe it automatically. Depending on the source, OCR can be quite challenging. Take my mail transit tables. Beyond the spotty typeface (missing or smudged characters, a 3 that looks like an 8, etc.), these kind of tables present particular difficulties: mess up the alignment for one row or column and that mistake can cascade across the entire document. Traditional OCR tools like ABBY FineReader or Tesseract might see the following:

…and understandably transcribe it as:

Even putting aside these inaccuracies, I didn’t actually want a full character-by-character transcription of the entire PDF. What I wanted was to extract and reformat the underlying data from this source into a usable spreadsheet. This required a certain level analytical reasoning – structuring columns and rows, deciding how to record the time values (hours? minutes?), handling missing values, etc.

This is where Vision Language Models (VLMs) come into play. Traditional OCR tools work through pattern recognition: detect where text appears on a page and then identify individual characters and words. VLMs take a different approach. Rather than processing characters and words in isolation, they integrate visual perception alongside a sophisticated understanding of language and the relationships between words. This allows VLMs to recognize that the above source is a table of cities organized alphabetically by state, that lontpeller falls under the Vermont heading, and that the intended city is therefore likely to be Montpelier.For more on VLMs, OCR, and historical transcription, see Jacob Carlson, Tom Bryan, and Melissa Dell, "Efficient OCR for Building a Diverse Digital History" (2024) and Noah Dasanaike's working paper, "Pre-Trained Language Models as Zero-Shot Tools for Social Science Research" (2026).

At least, that’s the idea. When I first pointed Claude Code to my folder of PDFs of transit tables and told it to extract the data from them, it failed miserably. Rather than using a VLM approach, it kept trying to use Tesseract to perform traditional OCR on the documents. So I pivoted to a different generative AI tool: Google’s Gemini 3.1 Pro. After some adjustments to my prompts, I was able to upload a four-page PDF of a single transit table and get back a fully usable, accurately transcribed dataset. Not only did Gemini handle things like missing or smudged letters, it was able to recognize irregularities, maintain table alignment, and reformat everything into a single usable spreadsheet for each PDF – all in just a few minutes.As of March 2026, Gemini seems to be the most capable frontier model for computer vision and historical transcription. Although I didn't do a rigorous analysis of its error rates, I used Claude Code and some spot-checking to identify potential mistakes in the data. This led me to fix around 12-15 incorrectly transcribed values (across nearly 8,000 records).

A sample from Gemini’s “model thoughts” shows the difference between traditional OCR vs. a VLM approach to transcription: I'm now focusing on formatting the data into the "long" format for the CSV, refining the output structure. I will rigorously check OCR outputs, specifically converting "H. M." values to decimal hours, and correcting any potential errors, such as "Little Kock" to "Little Rock." This isn’t pattern matching; this is using reasoning to translate a historical source into usable data – much like a human researcher would do.Take these "model thoughts" with a grain of salt; they're a somewhat imprecise take on the internal processes unfolding within an LLM as it generates its response.

In short, Gemini was able to turn this:

…into this:

The technical capacity for accurately transcribing messy historical sources is there, while the usability of these tools is still catching up. My experience with Claude Code – which defaulted to traditional OCR rather than using a VLM approach – shows how uneven the landscape still is. But given the pace of change, it feels like a gap that’s going to close very quickly (if it hasn’t already).Lots of historians are working on this - see, for instance, Mark Humphries on transcribing handwritten 18th-century storekeeper ledgers or Colin Greenstreet on transcribing handwritten Ottomon sources.

The Coding Bottleneck

The second bottleneck comes down to the technical skill of coding. Any historian can transcribe a transit table into a spreadsheet; the bottleneck is time and labor, not skill. Building a fully interactive, map-based web visualization that responds to user input and presents temporal shifts in spatial data? Vanishingly few historians have the skill and training to do this. If you are not one of those unicorn historians who knows how to define a CSS hover effect or implement a JavaScript event listener, then up until a few months ago your options were limited. You could find a technical collaborator, get funding to hire a developer, or accept the constraints and limitations of out-of-the-box software like Tableau or ArcGIS StoryMaps.

This changed in late 2025, when Anthropic released a major new update for the model behind its agentic coding system, Claude Code. Prior to this, vibe coding (ie. describing what you want in non-technical language and having an LLM write the code for you) necessitated at least some technical fluency, like knowing how to run Python in the command line. Today, that barrier has all but vanished. Someone with zero coding experience can install Claude or OpenAI’s Codex, write a few prompts, and get back a fully functioning website, visualization, or app a few minutes later.

When I sat down to test out Claude Code with my mail transit tables, I decided I was not going to write a single line of code myself. I wasn’t going to review code and approve it, I wasn’t going to tweak variables until something worked, I wasn’t going to lean on prior experience to debug or catch errors. Instead, I channeled a colleague with zero coding experience. I just described what I wanted, in plain language, and let Claude Code do its thing. This was my very first prompt:

There are two files in this folder: a PDF of mail transit times between major cities in 1882 and the transcribed data from that PDF. I want to use the transcribed data to build some kind of interactive app where a user can quickly see and visualize the data.

After some further back-and-forth to help clarify what I wanted, Claude Code worked for about 5-6 minutes. When it finished, I opened the HTML file it had created and immediately said “oh shit.” Here was a fully working and interactive prototype of that original vision I had fourteen years ago when I first stumbled onto this source:

Watching Claude Code build something like this from scratch felt like sorcery. But I would argue that the iterative capacities of agentic coding - the back-and-forth of refining, adjusting, and experimenting that turns a prototype into something you actually want to show people - is even more significant. Iteration has traditionally come with a lot of friction. If you’re writing the code yourself, iteration means trying to change something, breaking your entire project, and spending hours trying to debug it. If you’re working with a web developer, iteration means sending over your proposed changes, waiting for them to implement your request, realizing it made everything worse, and sheepishly asking them to walk it all back.

Vibe coding removes much of the friction of iteration. From my initial prototype, I could easily add additional transit table data from other years, build out more interactive functionality, and make a wide range of design and layout adjustments and tweaks. If something broke, Claude Code was able to fix it. Here is an example instruction I wrote during this process:

When you click or select a hub, have all cities show up in a scrollable table on the right. Add a hover function so that when you hover over a city on the map it subtly highlights it in the table and vice versa. Something is wonky with St. Louis when you click it, it turns grey not blue.

Above all, I was struck by how Claude Code reduced so many of the pain points of coding by hand. To take one example: a few sessions in, I remembered that the economic historian Jeremy Atack had compiled a dataset of historical railroad lines that would be perfect to add to my visualization. If I had been coding this myself, I would have had to wrestle with coordinate systems and map projections – my personal Achilles heel as a spatial historian that would have left me Googling “WGS 1984 shapefile” for the 1,284th time in my career before closing my laptop in frustration. Instead, I just downloaded Atack’s dataset and told Claude Code to add it to the map. And it did, without a single hiccup. As many others have noted: Claude Code take the drudgery of coding and makes it, well, fun.

Is the code behind my visualization especially elegant or well-structured? Probably not! I’m no web developer, but my hunch is that a gargantuan HTML file with 10,600 lines of code is not going to win any awards. But, for my purposes as a researcher, that doesn’t really matter: it just worked.As Jason Heppler has recently argued in "Vibing Digital History", the technical debt critique of vibe coding that applies to large software systems doesn't really hold for small, one-off visualizations like this one.

Side Quests

One obvious effect of the rise of agentic AI and reasoning models is that the field of digital history is now open to a much wider swath of researchers than ever before. A lot of the end products of digital history – online exhibits, project websites, maps, charts, network graphs, etc. – that used to require a certain degree of technical fluency can now be built with Generative AI.I agree with Lincoln Mullen, who notes in "Behind, ahead" that technical fluency is still a major advantage when it comes to using these tools. There were several times when my experience coding and working with data helped me write better prompts. However, the gap between a "digital historian" and a "non-digital historian" has closed substantially and will only continue to do so as these tools get better and better.

A less obvious effect of these changes is that many more research side quests are now worth pursuing even when you don’t know if they’ll pay off. Side quests are essential to the research process. Even the dead ends are often valuable: a chance for serendipity, learning new skills, reframing research questions, or stumbling into entirely new lines of inquiry. But side quests come with concrete costs. Historians are constantly making judgement calls about which sources to dig into and which to put aside. Ideally, these decisions are driven solely by the research questions themselves; in practice, they involve the kind of calculations I did as a graduate student when I came across that first mail transit table: was the juice worth the squeeze?

And guess what? It wasn’t! As far as I can tell, my snazzy transit visualization shows precisely the kind of unsurprising patterns you’d expect to see: an overall shrinking of informational time-space between 1882 and 1908, lengthy transit lags for the Deep South and patches of the American West, and big leaps in transit gains following the completion of major railroad lines. I could be wrong; maybe something new is hiding in there and some other historian or geographer or communication scholar will find it.Have at it! The raw data is available at: https://github.com/cblevins/mail-time/blob/main/data/transit-times-merged.csv. But at first pass, I didn’t see all that much potential in terms of publishable results.

If I had decided to execute my original vision back in 2012, I would have wasted a whole lot of time going down an analytical dead end. At best, I would have lost several weeks that I could have spent on other parts of my dissertation. At worst, I might have fallen prey to the sunk cost fallacy and tried to shoehorn this source into my dissertation just because I had invested so much time working on it.

Fast-forward to 2026. Using Gemini, I could transcribe that same source in a few minutes, at a cost of $0.34.I didn't actually pay this cost; I was using Google's experimental AI Studio, but it helpfully offers a token usage and cost estimate for prompts if you had been using their API. Transcribing each transit table took between 32,000 - 113,000 tokens, with a range of \$0.32 on the low end to \$1.07 on the high end. I could then use Claude Code to quickly visualize the data, start looking for patterns, and iterate back and forth depending on what I found. The same analytical dead end ends up costing several hours rather than several weeks. Moreover, even if the macro patterns from this data aren’t revelatory, agentic coding meant that I could use that data to easily build and deploy something that might end up being useful for other researchers: a genealogist curious how long it took a letter to travel between their relatives, a historian trying to gauge how isolated one of their subjects was from a major city, or a stamp collector trying to estimate the value of a particular postal cover.

Breaking some of the traditional bottlenecks to transcription and analysis changes the threshold for whether to pursue a historical research question at all. When the opportunity cost of transcribing a source drops to a few minutes and the cost of a candy bar, the challenge is no longer getting the data – it’s whether the questions you can ask with that data are interesting enough to pursue.

And this is where a different kind of expertise comes into play that has nothing to do with writing code or importing shapefiles: discernment. One feature of LLMs is that they’re eager to give you something. When working with Claude Code, it kept presenting me with what it claimed were interesting patterns in the data. If I were a graduate student, I might have excitedly leapt at these suggestions. But part of becoming an experienced researcher is learning to recognize not just what IS worth digging into but what ISN’T worth pursuing. So when Claude ran a clustering algorithm on cities based on their connectedness and declared “the typology shifts are particularly interesting,” I had the experience to say: “No, they’re not.”

The old bottlenecks were time and technical skill; the new bottlenecks are judgment, discernment, and taste. This is true for many different disciplines, but history seems especially well-positioned to benefit from this shift. Accessing and transcribing archival sources has always been a major sticking point for historical research, and the ongoing decimation of federal grant funding has made that bottleneck even more acute. Vision Language Models don’t just make transcription faster or cheaper - they will (hopefully) expand the kinds of sources we can access and work with.

Meanwhile, coding remains a vanishingly rare skill for historians. For colleagues in the social sciences and hard sciences, agentic coding dramatically speeds up something they could already do. For historians, agentic coding allows us to do things we couldn’t do before. Without training in statistical thinking or computational methods, there’s a risk that we start vibe coding our way to questionable results. But agentic coding in history doesn’t have to mean pumping out a bunch of faulty regression analyses; it can also mean spinning up an interactive web app to explore a set of sources that have been languishing on your computer for the past fourteen years. By breaking the bottlenecks of transcription and coding, these tools free up time and energy for historians to use the skills we actually excel at: close reading, contextual thinking, narrative, and interpretive nuance.

Are AI companies gobbling up vast quantities of natural resources? Yes. Are their models built on the creative work of people without consent or compensation? Yes. Are these models hollowing out human relationships, connections, and work? Yes. Are they toxic for learning? Yes. But are LLMs just “autocomplete on steroids”? No. In the understandable drive to hold the AI industry to account for its many harms, we can’t keep moving the goalposts when it comes to evaluating what their products can actually do. Even if you are deeply opposed to generative AI, I think it’s important to have a clear-eyed understanding of where this technology currently stands.

In this post, I want to review some of the big developments of 2025 to show just how quickly this technology has advanced. Again and again, I’ve found that what many of my colleagues knew (or thought they knew) about generative AI twelve months or even six months ago is already outdated. One reason for this is that there’s an overwhelming amount of stuff out there to keep up with - library guides, Substacks, op-eds, podcasts, webinars, etc. Not only is it hard to sift through everything, but many of the resources on generative AI operate at such a high level that it’s difficult to know what it means for your specific job or discipline. This post tries to make things more concrete by looking back at some recent developments in generative AI through the lens of a historian: what changed in 2025 and what are the implications for the historical profession?

Multimodal Generation and History Slop

Let’s start with three images of Abraham Lincoln. Two of them are real photographs and one was made with ChatGPT. Give yourself 5-10 seconds and see if you can spot the fake. How confident are you? And how much money would you wager on your choice?

Whether or not you guessed correctly, this exercise illustrates a larger point: generative AI has gotten really good, really fast, at generating multimodal content, or stuff that isn’t just text (i.e. images, video, and audio). Over the past twelve months, we’ve gone from bizarre AI-generated photos with melted faces and six-fingered hands to images and videos that are increasingly difficult to distinguish from human-made ones.

Historical “deep fakes” have received a lot of deserved attention, but I’ve seen a lot less discussion of a far more pervasive issue: advances in multimodal generation have obliterated the barriers to producing synthetic historical videos at scale. The problem isn’t how realistic these videos look (most won’t fool you into thinking they’re “real”); it’s that multimodal generation has gotten so good (and so cheap) that it has already created a tidal wave of AI history slop.

Try searching YouTube for a historical event or historical figure and then filter your results to the last year of uploads; you will soon find yourself scrolling past AI-generated videos posted by users like ALLUSHISTORY or AmericanHistoryBufff:

Any historian interested in public engagement needs to pay attention to this development. Video (particularly short-form video) has come to dominate the media landscape. Social and video networks are now the most popular source of news for Americans of all ages, outpacing television, print, and online news sites for the first time. This is especially true for our current and future students: 3/4 of teenagers use YouTube every day, with TikTok not far behind. You can write all the op-eds and publish all the trade-press books you want; if you want to reach the public where it’s at in 2025, it’s video, video, video all the way down.

Content creators have been monetizing shitty history videos on YouTube for years. But until recently, making a successful video required at least some time and effort to put together (if no actual expertise). Recent jumps in multimodal generation has eliminated those barriers. Anyone, anywhere, can now make a polished-looking video on any historical topic at a very low cost and in a matter of minutes. No scriptwriting, no piecing together images, no audio recoring and editing; just type a few prompts and you have yourself a video.

The YouTube creator Historian Sleepy illustrates the changes unleashed by 2025’s advances in multimodal generation. Historian Sleepy has amassed 182,000 subscribers by posting history-themed videos for, you guessed it, people to fall asleep to. Each video is one long stream of AI-generated images and video clips about some general historical topic narrated by a soothing male voiceover, ranging from “How FRONTIER FAMILIES Survived The Coldest Nights and more” to “Who Were the Sumerians? The People Who Started Civilization” to “2 Hours of Life in the Gilded Age : 1880s America”:

The scale of output is staggering: each video runs 2 to 3 hours long and they’ve posted 286 of these multi-hour videos in just seven months - the equivalent of churning out a lengthy PBS documentary every single day (and sometimes multiple times a day). Each video has amassed tens of thousands of views. In 2025, academic historians are no longer competing for the public’s attention with the History Channel or bestselling “pop” history books; we’re competing with an ever-growing deluge of AI-generated history slop.

Deep Research

Multimodal generation was only part of the generative AI story in 2025. This year also saw the rise of reasoning and agentic models. To oversimplify, reasoning models break their responses down into smaller sequential steps (called “chain of thought”) and spend more time and computing power on generating an answer. This significantly improves responses and decreases hallucinations. Today, all the major frontier models offer this “reasoning” (sometimes called “thinking”) option. AI agents, meanwhile, use this “reasoning” architecture to come up with a plan for how to complete a task before taking a series of steps (often adjusting those steps as they go) in order to achieve that goal on their own.

For historians, the most direct application of these new reasoning and agentic capabilities came with the release of a “Deep Research” feature for all major frontier models. Selecting this feature sends the model off to browse the internet, sift through different sources of information, and synthesize all this material into a lengthy written report. If you rewind to the end of 2024, the ability of LLMs to locate reliable and credible sources was extremely poor, often returning bogus links to fabricated sources that didn’t exist. Twelve months later, that’s changed.

In September 2025, I tested out ChatGPT’s “Deep Research” feature by in my own particular area of expertise: the history of the American state in the American West. I selected the most advanced model available at the time (ChatGPT 5.1), selected the “Thinking” and “Deep Research” options, and instructed it to write “a historiographical essay on the current landscape of scholarship related to the history of the American state in the American West.” Over the next six minutes, the model conducted research and came back with a 7,000-word report.

The report hit many of the major of the historiographical notes I would hope to see on this topic, from Frederick Jackson Turner’s “frontier thesis” to American Political Development (APD) literature to the “New Western History” of the 1980s and 1990s to the more recent framework of “Greater Reconstruction.” It referenced work from a wide range of credible historians (Stephen Aron, Brian Balogh, William Cronon, Gary Gerstle, Sarah Barringer Gordon, Kelly Lytle Hernández, Beth Lew-Williams, Patricia Limerick, Benjamin Madley, Noam Maggor, Karen Merrill, William Novak, Margaret O’Mara, Elliott West, Richard White). There were no glaring mistakes, no fabricated citations to non-existent sources, no “10 Amazing Facts About the Wild West” websites - all things you would have found in a similar AI-generated essay in 2024. Hell, it even located and inserted (with full attribution) a map of federal land in the West, providing a compelling visual example of the state’s presence in the region:

Was it perfect? No. If I were writing this essay, I would have emphasized some different strands of scholarship or characterized certain trends differently. And there were other limitations, most notably a tendency to access and cite scholarship indirectly. To take one example, in summarizing William J. Novak’s article, “The Myth of the ‘Weak’ American State,” the model cited a Legal History Blog post…that itself was an excerpt of an AHA Today post…that itself summarized a roundtable in the American Historical Review…that was itself a discussion of Novak’s article in the American Historical Review. The reason for this game of scholarly telephone is that these models can’t, as of now, easily or directly access paywalled articles or full books. Even with this limitation, it did a remarkably accurate job of summarizing this scholarship. And, given the growing ability of AI agents to, say, log into databases using a user’s own credentials, I suspect the paywall issue won’t be a barrier for long.

Again, it’s important not to move the goalposts when evaluating this kind of output. If a graduate student had submitted this essay to me, I would have said they need be more concise and cite scholarship more directly, but otherwise would have been impressed with their command of the scholarly literature and how much material they managed to pull together into a coherent whole. I know this because I went back and compared the ChatGPT report to my own work in graduate school. It was, in certain respects, better than any historiographical essay I wrote as a PhD student (maybe I was an especially poor student). There is, of course, one final caveat: this particular “Deep Research” report was not conducting new, original research; it was a compilation of existing scholarship. This is precisely the kind of synthesis that LLMs should excel at. But if you want to dismiss this kind of report as derivative, unoriginal “fancy autocomplete,” then you need to similarly dismiss the historiographical essay as a bedrock genre of the historical profession.

Handwriting Transcription

Arguably the most notable development in generative AI for historians came in the realm of handwriting transcription. First, some background: the mass digitization of archival material has had a profound impact on historical research practices, and much of that impact came from making sources searchable and discoverable at scale. This required Optical Character Recognition (OCR), or using a computer to automatically transcribe historical documents. The problem is that a messy handwritten letter is a lot harder for a computer program to transcribe than neatly typed text. Although there have been major advances (most notably with the Transkribus platform), generally speaking computers make far more errors than human transcribers when it comes to handwritten historical sources.

The release of Google’s newest frontier model (Gemini 3.0) in November 2025 seems to have ushered in a new era for AI handwriting transcription. To illustrate this, I’m going to go into the weeds a bit and draw heavily from the work of historian Mark Humphries (I would encourage you to read his posts in full: Part 1, Part 2, Part 3).

Mark tested out Gemini 3.0 with a handwritten accounting ledger kept by an Albany merchant in 1758. The document is a mishmash of numbers, dates, names, abbreviations, and columns - precisely the type of messy handwritten source that computers typically struggle to transcribe. Accurately deciphering this kind of document requires more than just visual accuracy (e.g. distinguishing a “4” from a “9”); you need to be able to analyze surrounding clues from the rest of the document and understand the historical context in which it was produced. In short, it requires reasoning, knowledge, and interpretation. Take this single entry from the ledger:

Gemini accurately transcribed the first line as Entd Samuel Stitt 27 albany Dr. This records one entry (Entd - entered) from March 27th, 1758 (27 - the month and year recorded at the top of ledger’s page) noting a purchase (Dr. - debit) made by an Albany-based merchant named Samuel Stitt.

The second line, however, is where the rubber hits the road. As a human reader, I would transcribe this as: 1 loff sugar 145 @1/4 0 19 1. However, unless you’re deeply versed in 18th-century British currencies, weights, and accounting practices (and are willing to do some math), you probably have no idea what this string of numbers actually means. I sure didn’t! But Gemini didn’t just transcribe these numbers. It went on to correctly interpret them as:

• 1 loaf of sugar
  
• weighing 14 lbs. 5 oz.
  
• at the rate of 1 shilling, 4 pence per pound
  
• for a total cost of 0 pounds, 19 shillings, and 1 penny

Gemini’s “reasoning trace,” where it explains how it reached this conclusion, is truly mind-blowing. Rather than interpreting 145 as, say, a running tally of how many loaves of sugar this merchant had bought over the past year or any number of other plausible but incorrect explanations, Gemini correctly identified it as a quantity of weight. It did so by conducting a series of calculations and conversions across multiple units of weight and currency to arrive at the result of 14 lbs. 5 oz. In Mark’s words, Gemini “was seemingly able to correctly recognize, convert, and manipulate hidden units of measurement in a difficult to decipher 18th century documents in ways that appear to require abstract, symbolic reasoning.” Again: we cannot keep pretending this is just “fancy autocomplete.”

Other historians have gotten similarly impressive results with Gemini 3.0. If it holds, this development has huge implications for historical research. Many digital collections have skewed towards printed, typewritten sources over handwritten ones in large part because of the challenges and costs of transcribing handwritten sources. But if LLMs like Gemini can now consistently rival the abilities of human typists, and at much lower costs across much larger scales, they will open up a new era for digital archives and, by extension, archival research.

Taking Stock

All three of the areas I’ve described here - multimodal generation, Deep Research, and handwriting transcription - have changed dramatically in the past twelve months. YouTube is now flooded with thousands of hours AI history slop that didn’t exist in 2024. With a single prompt, students can generate lengthy historiographical essays that would have been riddled with hallucinations in 2024. Archivists and researchers can use LLMs to transcribe handwritten sources with a degree of accuracy and sophistication that was impossible in 2024. Twelve months ago, all of this would have seemed far-fetched; twelve months from now, this post might be outdated. This is why I think it’s so important for historians to continue to keep pace with these developments.

Here is where I should offer some brilliant pronouncements about how to grapple with these changes. But I’ll be completely honest: I don’t have them. I don’t have a roadmap for responding to all of this. What I do know is that dismissing generative AI as just “autocomplete on steroids” is no longer tenable. We can’t reckon with this technology’s harms and benefits while clinging to outdated characterizations of its capacities. Understanding what generative AI can actually do - and how quickly those capacities are changing - is a necessary first step before we can sort through what it all means for our profession.

This is the first in a series of posts I hope to write about using common, off-the-shelf LLMs for historical research. I’m not writing this for folks who know how to, say, use OpenAI’s API to implement a RAG model for a historical corpus (if you know what that means, good for you - but this post isn’t for you). I’m writing for those of you who’ve read plenty of New York Times articles about ChatGPT but haven’t used it all that much themselves. You’re probably skeptical of the hype around these tools and wonder whether they can actually help you do your work. If this describes you, keep reading.

Today’s post focuses on using LLMs to transcribe and interpret primary sources. For most of my career as a digital historian, computational analysis has focused on “machine-readable” sources, or documents that can be easily transformed into data. A typed government report can be automatically processed with Optical Character Recognition (OCR), whereas a handwritten letter has to be transcribed by a human.Transkribus offers a specialized AI-based handwriting transcription tool. While effective, there is a higher barrier to entry than more general tools like ChatGPT - it requires a paid account, can take a long time to process documents, and there's a learning curve to understanding its interface. Other projects like From the Page are also starting to experiment with AI-assisted transcription.

But what counts as “machine-readable” is rapidly changing. To illustrate what I mean, I’ll use a concrete example from my own work. While researching my book, I spent a lot of time mapping large-scale spatial patterns in the 19th-century postal system as it spread across the American West; what was missing from this approach was a more qualitative, intimate look at how individual people and families actually used the US Post.

Enter the Curtis family: four orphans who used the mail to stay connected as they moved across the western United States. In 2015, I spent several days at the Huntington Library typing on my laptop to transcribe their correspondence. My favorite document from this collection was a letter written by Benjamin Curtis to his sister Delia Curtis. In it, Benjamin announced the birth of his daughter on a hot September day in 1886:

Take a peek through any historian’s hard drive and you will probably find thousands of similar photos form archive trips. Transcribing these documents has long been a bottleneck in the research process: a tedious but necessary step between photographing a primary source and actually using it in our work. When I transcribed Benjamin’s letter at the Huntington Library, it wasn’t particularly difficult work, but it was laborious - especially after three days of hunching over the same desk squinting at letter after letter.

Document transcription is exactly the sort of time-consuming, boring task we would want to hand off to a Large Language Model. But can it actually do the job? To find out, I chose (womewhat at random) a Custom GPT that was designed specifically for transcribing handwritten text. I uploaded my photos of Benjamin’s 1886 letter and, a few seconds later, had a shockingly accurate, well-formatted transcription. Here’s a sample of that transcription:

This is wild! If you had told me just a couple of years ago that a general-use tool could - in a matter of seconds - generate a largely error-free transcription of a handwritten letter from the 19th century? I would have laughed you out of the room! Just as impressively, ChatGPT was able to take its transcription and immediately generate an accurate, concise summary of Benjamin’s letter:

In a letter dated September 8, 1886, from Catalpa, Arizona Territory, Benjamin Wisner Curtis joyfully announces the birth of his daughter, Selia Etta, to his sister, L. Delia Augusta Curtis. He describes the baby’s healthy condition, with dark hair, and notes that both mother and child are doing well despite the hot weather. The letter highlights the involvement of multiple generations in the family’s care and the importance of familial bonds during this significant event.

This example points to the potential of using LLMs for historical research. Their growing ability to process both natural language (text) and “multimodal” sources (non-textual sources like photographs, video, or audio), makes them a powerful tool for working with archival material. And unlike other kinds of digital tools, ChatGPT doesn’t require a user to learn any sophisticated technical skills.

Now for the caveats. Benjamin’s letter is in some ways an ideal source for this kind of analysis. I’ve tried using ChatGPT to transcribe more challenging documents (ex. a 17th-century handwritten letter in Spanish) and the transcription quality degrades dramatically, sometimes to the point of gibberish. This plug-and-play approach won’t work for every type of source. It’s also not yet scalable. Don’t expect to be able to dump hundreds of documents into ChatGPT and receive flawless, neatly formatted transcriptions.

Moreover, the tool will make mistakes, even with an “ideal” source like Benjamin Curtis’s letter. To take one example: towards the end of his letter, Benjamin revealed that he and his wife Mary had decided to name their newborn daughter “Delia Etta”. ChatGPT, however, transcribed this as “Selia Etta”. It’s a minor and understandable error - a handwritten “D” looks an awful lot like a handwritten “S”. But this mistake had implications for ChatGPT’s ability to interpret the source.

Benjamin named his daughter Delia Etta in honor of his two older sisters, Delia and Henrietta (who Benjamin affectionately called Etta). The choice had particular resonance for the Curtis siblings, who were orphaned as young children. Naming Benjamin’s daughter after his two sisters was a way to strengthen the family bonds that had been so badly frayed by the loss of their parents. Of course, Benjamin never said any of this explicitly in his letter; it would have been obvious to his sister. It was obvious to me, too. After all, I’d spent many hours at the Huntington Library getting to know the Curtis family across dozens of letters just like this one. But without that context, a tiny transcription mistake - switching an “S” for a “D” - meant that ChatGPT missed an important part of this letter and its larger significance.

The Curtis letter demonstrates why using LLMs for historical research requires careful attention and disciplinary expertise. These tools are potentially quite powerful for aiding historians with some of the mechanics of primary source analysis, but they work best when we actively guide and verify their output. Which is why I would encourage you to take the plunge and start experimenting with these tools for your own research.

If you’re just getting started, here is some general advice:

• Think of an LLM as a research assistant. An extremely knowledgeable, well-read, and over-caffeinated research assistant. They can accomplish quite remarkable things but they will also make mistakes - like mixing up “Selia” and “Delia”. Unlike most research assistants, LLMs will occasionally hallucinate (ie. make up) things that sound quite convincing. Don’t take their output at face value and always check their work.
• Start small and be specific. Begin with shorter, discrete sources. These are easier for the model to work with and easier for you to evaluate. Just as you would with a research assistant, provide detailed guidance to the LLM about what, exactly, you want it to do. Generally speaking, it helps to assign it a role (ie. “You are a historical research assistant…”) before carefully spelling out the task that you want it to do.I tend to think prompt engineering is an overrated skill. But starting with something like the CREATE framework (or any of the seemingly hundreds of similar acronyms out there) can help you get the hang of writing effective prompts.
• Choose the right tool. This is tricky, given how quickly things change in Generative AI. As of October 2024, if I had to choose one off-the-shelf model for working with primary sources I would probably go with the paid version of GPT-4o (which requires a $20/month subscription). This lets you analyze “multimodal” sources (like photographs) along with using Custom GPTs tailored for specific tasks (like handwriting transcription).Ethan Mollick wrote a a nice breakdown of the major "frontier" models as of September 2024.
• Keep the human in the loop. LLMs work best when you actively work alongside them as a form of “co-intelligence”. Remember: you are the expert. Iterate back and forth. If the model misses something, point out what it got wrong and ask it to revise. I’ve found that this process of iteration (the “chat” part of ChatGPT) isn’t just about fixing mistakes; it’s actually a useful way for me to refine my own thoughts and interpretations.

What does it mean for historians to “keep the human in the loop”? It underscores the fact that when we read a primary source we aren’t just trying to accurately decode individual words on a page. We’re looking for subtle details and connections that aren’t always obvious. We’re listening for the silences of who’s left out and what’s left unsaid. We’re recognizing what is meaningful or surprising. We’re situating a source within the broader context of its time. All of this requires a certain degree of specialized expertise and experience that, for all of their massive training sets and sophisticated language capacities, Large Language Models can’t fully grasp on their own. At least, not yet.

• Can Generative AI accurately grade student writing? In other words, can it reliably sort out the A’s from the B’s from the C’s?
• Can Generative AI give detailed, high-quality feedback on student writing?

To find out, I looked on OpenAI’s directory of Custom GPT’s and chose one called Essay Grader Professor, developed by Stephanie Shasteen and described as “an academic GPT specializing in essay grading and tailored feedback.” Let’s put it to the test!

🧪 Testing It Out 🧪

What kind of measurements should we use to evaluate a Custom GPT like Essay Grader Professor? I figured the best way to test it out would be to pit its grading abilities against my own.A Custom GPT is a more specialized version of OpenAI’s more general-use Large Language Model GPT's (such as ChatGPT) and are tailored for specific tasks or applicataions. For the sake of documentation and transparency (and because the landscape of Generative AI moves so quickly), I did this on May 19, 2024 using Essay Grader Professor GPT powered by OpenAI's GPT-4.

In my Fall 2019 semester of History of the Western U.S. I had my students write a Pop Culture Review in which they had to review a piece of pop culture loosely within the “Western” genre, summarizing it and evaluating its historical context in 4-5 pages.I chose this assignment because I suspect you'd find a similar type and length of essay in hundreds (thousands?) of college courses each semester. I then went back to my gradebook and selected three student essays to use as test cases (carefully stripping out any personally identifying information): the essay that received the lowest score in the class, the essay that received the median score, and the essay that received the highest score.

I then uploaded the rubric I had used to grade the essays across five categories (Writing Mechanics, Analysis, Evidence, Structure, and Historical Perspective) and wrote the following opening prompt:

Me:

Please read the assignment description below. I am also attaching the writing rubric to use. The students are not required to use a particular citation style and do not deduct points for citation formatting (but you can evaluate their sources and how they use them). Focus more on their writing mechanics and argumentation, and less on their research/sources.

Essay Grader Professor GPT:

Great! I have reviewed the assignment description and the grading rubric. The rubric evaluates essays based on Writing Mechanics, Analysis, Evidence, Structure, and Historical Perspective. Now, please upload the first essay for evaluation. I will use the provided rubric to assess the essay and provide detailed feedback.

📝 How Did It Do? 📝

Can Generative AI accurately grade student writing?

Here’s a snapshot of the Custom GPT’s essay scores across five rubric categories compared to my own. Scores in green fell within the same range for the given rubric category - ex. under Writing Mechanics the Custom GPT might have given a 16 and gave a 20, but we were both within the same range of 16-21 points: writing lacks clarity and contains many errors and/or deficiencies in grammar, syntax, or phrasing). Scores in red fell within a different range on the rubric.

How should we interpret these scores? Let’s start with a thought experiment: what if I wasn’t comparing my grades to a Custom GPT, but rather to another professor grading the same essay? I’m confident we would still rank the essays in the same order (Low, Middle, and High), but I suspect that our rubric scores would have quite a bit of variance.

This is exactly what I found with the Custom GPT. It was able to accurately grade the essays using the rubric I provided and ranked the essays in same order as I did. Although it assigned lower scores across the board, none of them struck me as unreasonable. If anything, the results highlighted my own tendency to fudge student scores upwards a bit. If I had to choose which overall grade was more “accurate” according to the strict letter of the rubric, I’d give the nod to the Custom GPT (ie. Essay #1 was probably closer to the Custom GPT’s D grade than my own C+ grade).

Can Generative AI give detailed, high-quality feedback on student writing?

In short, yes. The Custom GPT’s feedback was extremely detailed, writing short paragraphs for each of the five categories that included specific examples from the essay along with a summary paragraph at the end (around 500-600 words of total feedback for a 1,500-word assignment). Based on conversations with students over the years, this is a lot more than they receive from most of their professors. For instance, here’s its feedback on Writing Mechanics for Essay #2:

Writing Mechanics (25 pts)
Score: 21/25
The essay is generally clear, but there are some grammatical errors and awkward phrasings. For example, “The movie was written and directed by Michael Crichton a Science Fiction heavyweight who brought us classics such as The Great Train Robbery Jurassic Park and Westworld,” lacks necessary commas for clarity. Another instance is, “The parallels between the guests’ attitude and that of settlers during the 1870s can be seen: treating everything as something they had a right to,” where the colon usage is incorrect and affects readability.

But quantity doesn’t mean quality. How good was this feedback? Here we get more of a mixed bag. With a few exceptions, I thought the Custom GPT did well on the Writing Mechanics, Analysis, and Structure categories. Compare, for instance, our respective feedback on the Structure category for Essay #3:

Custom GPT: “Some paragraphs could be broken down to enhance readability, such as the long paragraph detailing the pole’s origin story and legal aftermath.”

Me, in the margins of the same paragraph: “This is all good, but needs to be condensed down. You can tell the same story without getting bogged down in quite so many details.”

The Historical Perspective category was where the Custom GPT lagged behind. Essay #1 had several big historical inaccuracies that I used as a teachable moment to reinforce some of the course’s big themes that the student seemed to be struggling with. The Custom GPT missed these entirely. In Essays #2 and #3, meanwhile, the Custom GPT said they could have “delved deeper” despite both essays excelling in this category.This felt like a critique for the sake of a critique without any regard for the constraints of the assignment (akin to "Reviewer #2" who asks you to add three completely unrelated sections to your article). Moreover, the Custom GPT tended to critique passages in the essays that I didn’t find especially glaring. Most of its critiques weren’t wrong, necessarily, but they weren’t the ones that I would have chosen.

At my best, I still think I write higher-quality feedback than the Custom GPT - especially in the context of a history course. But I’m not always at my best. When I’m exhausted and in the depths of grading purgatory, my feedback is probably less concrete, actionable, or consistent from essay to essay. When judged against this bar, I think the Custom GPT did quite well.

Full Feedback

I’m including the full feedback generated by the Custom GPT for each essay - I’d encourage you to take a look for yourself to gauge its quality.

💭 Big Takeaways 💭

Overall, the Custom GPT’s performance far exceeded my expectations. In 2023, I tested out the free version of ChatGPT (GPT-3.5) to critique a piece of sample writing. It was borderline useless: hallucinating entire passages, misidentifying sections, etc. One year later, a Custom GPT powered by GPT-4 was able to follow the rubric, assign accurate scores for each category, identify strengths and weaknesses, and even pull out direct quotes from the essays to use as examples. Although it missed some of the subtleties and specifics of the assignment, if you told me a teacher wrote this feedback I wouldn’t bat an eye. If anything, I’d compliment their thoroughness and attention to detail.This is yet another reminder of how the GPT-4 class models are _vastly_ more capable than their predecessors. I stress this difference all the time when talking to colleagues, many of whom are basing their impressions of ChatGPT on the free version powered by GPT-3.5.

Even so, I’d want to tinker with the Custom GPT before I would consider using it in an actual course. This might involve writing more detailed prompts, providing examples of my own feedback to use as a model, or building my own version from scratch that was explicitly geared towards a specific course or assignment.If I was able to customize it enough, it might be able to offer _better_ feedback than me because it can scale up its feedback so easily (both in terms of how much feedback it's giving each student and the number of essays it can grade without losing its sanity). Above all, I’d want to be completely transparent with my students about how and why I was using this tool. And I suspect they would have serious reservations about it.

As tends to happen with Generative AI, this entire experience brought up bigger issues than the narrow technical capabilities of a Custom GPT.

Academic Labor

There is no question that a Custom GPT can “automate the boring” when it comes to grading. It takes me about 15-20 minutes to grade one student essay (leaving comments in the margins, assigning rubric scores, and writing a two-paragraph summary of my feedback). Using a Custom GPT could cut this down to 2-3 minutes per essay (stripping out identifying information, double-checking its output, etc.). With 20 students in a class, that would save me something like 5-6 hours of tedious work. Multiply this across several assignments per semester, and it quickly adds up.

In an ideal world, this kind of tool would free up teachers to spend their time on more meaningful pedagogical work. But we don’t live in an ideal world. Instead, I worry that widespread adoption would only accelerate the devaluing of academic labor. Administrators could easily use it as justification to hire fewer instructors while loading up existing ones with more classes, larger sections, and fewer teaching assistants.

Grades

Watching a Custom GPT capably mimic the grading process has only reinforced how hollow grades can actually be. I like to think that all the hours and thought I put into my feedback is actually helping students learn; that they reflect on my comments and incorporate suggestions into future assignments. And maybe they do. Most of the time, however, I suspect that they anxiously skip to the overall grade, skim through a couple comments, and that’s about it.

Grades tend to incentivize students in all the wrong ways. The goal for most students on any given assignment is not to learn new things or develop new skills; it’s to get an A (or, rather, to NOT get a B, or a C, or a D). This means that they shy away from taking intellectual risks, experimenting with unfamiliar approaches, or pushing themselves outside their comfort zone. Why risk it?

All of this leads to a dynamic that feels transactional and punitive; “You did X, Y, and Z, here is your reward. You didn’t X, Y, and Z, here is your punishment.” Guess what? This dynamic will only get worse if students know that a chatbot, rather than a human, is grading their work.I'm guilty of this dynamic myself, where I'll flag something in order to justify my rubric score - as if my intended audience is a student who is going to dispute their grade, not one who will reflect on my feedback and learn something from it.

Human Connection

Consider the dystopian and entirely plausible scenario where a student uses ChatGPT to write their essay and a professor uses a Custom GPT to grade it - a (non-)human centipede of utterly meaningless writing and feedback. To demonstrate, I asked ChatGPT (GPT-4) to write an essay based on the above assignment, which I then uploaded to the Essay Grader Professor GPT. It lauded the essay as “well-written and insightful,” giving it an A grade. 🤮

If this scenario makes your skin crawl, you can understand my unease about the loss of human connection in teaching and learning. I see this as part of a wider shift towards impersonalized education. Instructors turn into a disembodied series of pre-recorded lecture videos. Classmates become faceless figures on a Canvas discussion board. Teaching gets reduced to a mechanized content delivery system.

Maybe I’m shouting into the void, but I still think that the most effective teaching comes through empathy and personal connection. A history major applying to law school might need different kind of feedback than an engineering student taking the class to fulfill their humanities requirement. A flurry of harsh critiques might motivate one student to put in more effort and cause another student to disengage from the class entirely. A Custom GPT doesn’t understand any of this context, instead reducing the relationship between a teacher and a student into a series of automated transactions devoid of meaningful human engagement.

In November 2022, OpenAI’s release of ChatGPT triggered an earthquake in higher education. ChatGPT is perhaps the most famous example of Generative Artificial Intelligence (AI), or a larger umbrella category of computer models that ingest massive quantities of human-created data in order to learn patterns and generate new material that mimics that same content, whether images, video, music, or, in ChatGPT’s case, written text. A user can ask ChatGPT to write all sorts of things via a simple chat-based interface, from poems, recipes, and short stories, to (more problematically for teachers) entire essays, reading summaries, and answers to multiple-choice questions. Eighteen months later, we’re still trying to find our collective footing.

As a ThinqStudio faculty fellow and the Director of Digital Initiatives for CLAS, I’ve spent a lot of time over the past several semesters wrestling with these changes. I co-organized workshops to help instructors learn about Generative AI, from the basics of Large Language Models to convening a panel of faculty and their students who shared how they used these tools in their classes. I was part of a working group that drafted sample syllabus policies and other teaching resources around Generative AI for faculty to use in their classes. And I’ve taught students about some of these tools and topics in a course, Introduction to Digital Studies. So: what have I learned?

First, this is a new era for teaching. I’m typically a curmudgeon when it comes to the Silicon Valley hype machine. But we’ve reached an inflection point. Generative AI is not going anywhere; if anything, it will only get more ubiquitous, more powerful, and more deeply woven into our daily lives. As teachers, we can’t ignore this. In a way that is different from other tech developments, educators need to not only familiarize themselves with this new landscape, but to keep tabs on how it is changing. And it is changing quickly.

If you tried out ChatGPT after its release in November of 2022 - perhaps copying and pasting one of your essay prompts to see what it wrote in response - chances are good that you used the free version of ChatGPT based on the underlying model of GPT-3.5. And its response might have been mediocre to below-average: a tendency to hallucinate (ie. fabricate) citations and facts, an inability to retrieve information about current events, or a blandly generic writing style. But that’s changed. GPT-4 was released in March 2023 and powers the current subscription-based plan for ChatGPT (along with Microsoft’s Copilot tool). Many of the shortcomings you might have noticed in the free version of ChatGPT based on GPT-3.5 have been drastically reduced or even eliminated in GPT-4.

Take the following example. I teach HIST 1362: US History Since 1876, an introductory-level core humanities course for CLAS. During one class, I have students study America’s booming post-World War II consumer economy by looking at advertisements from the 1950s. The goal is to familiarize students with the topic while also teaching them how to closely read and interpret the imagery and symbols of consumer culture, such as this heavily gendered advertisement for a Frigidaire oven:

So what happens if I ask ChatGPT to analyze the above advertisement as if it were one of my students? When ChatGPT was first released in November of 2022, it wouldn’t have been able to do so. At the time, it could only handle text-based prompts that a user typed into the chat. Today? GPT-4 can handle not just typed prompts, but also entire PDF documents, photographs, or historical images like the above advertisement. And it can perform remarkably in-depth analysis. When given the Frigidaire advertisement, it correctly identified “the basics” of the source, unpacked some of its visual imagery, placed the source within a larger historical context, and even provided references and brief summaries of relevant articles and books related to the topic. Here are the instructions I gave to ChatGPT along with its full response; I hope it drives home just how rapidly this technology has advanced:

Instructors don’t need to become experts on Generative AI. But they do need to develop more familiarity with some of the main applications in this field and any major new developments. I’ve found the work of Ethan Mollick, a business school professor who writes about AI in higher education, to be a good starting point. He and Lilach Mollick have produced a short video series, Practical AI for Instructors and Students, that offers an accessible introduction geared towards educators. Ethan Mollick’s Substack newsletter One Useful Thing, meanwhile, is a useful way to keep up with new developments. But far and away the most effective way to get up to speed is to actually use these tools. If you don’t want to pay $20 a month for the paid version of ChatGPT, then you can access the underlying GPT-4 model through Microsoft’s Copilot tool. Set aside several hours to simply experiment with the tools and you’ll get a better sense for their capabilities and limitations than you would from reading dozens of articles about them.

This brings me to my second lesson learned: we need to actively adapt our teaching to this new landscape. For awhile, I told colleagues that it was not the responsibility of any one instructor to revamp their entire curriculum overnight, instead reassuring them that they could wait for the dust to settle around Generative AI and see where we ended up. Well, the dust has settled. We need to start making changes to our teaching.

At a minimum, everyone needs to decide what kind of policies you want for specific courses regarding the allowable use of Generative AI. Students need clarity. I would start with the pros and cons table of these three sample syllabus policies. Choose the policy that most closely aligns with your individual course and then tailor and tweak it accordingly. Once again, it’s imperative to become at least passably familiar with what the current tools can and cannot do. Think through the major assignments and goals of a course. Does it make sense to allow students to use ChatGPT for, say, brainstorming ideas for a project? Drafting an outline of a paper? Locating potential sources? Then think about enforcement. What happens if you receive an essay that you suspect was written by ChatGPT? How are you going to talk to the student?

The issue of enforcement is a bugbear for many instructors, who continue to focus on Generative AI’s implications academic integrity, plagiarism, and cheating. And it’s true that students have and will continue to use these tools to submit work that they did not produce themselves. There are no silver bullets coming from educational institutions. Nor should we expect the giant, profit-driven companies building these tools to come up with solutions. So-called “AI detection” software is, at best, fool’s gold - unreliable, easy to bypass, and prone to false positives for non-English speakers. Let me repeat: there is no software that can reliably “prove” that a piece of text was written by ChatGPT.

Rather than focusing so much on academic integrity, I would encourage educators to instead think more constructively about how you can integrate Generative AI into your teaching in ways that actually help students learn. One strategy is to use Large Language Models as tutors or coaches. During one of our workshops this past fall, Professor Dawn Gregg shared how she had started using ChatGPT in her online course ISMG 6020: Python Programming Fundamentals. In that class, she had her students use ChatGPT as a tutor to help document and debug their Python code, provide examples of programming concepts, and generate practice questions. Professor Gregg reported some remarkable results. Compared to the previous semester she taught the class, student attrition rate went down and the quality of their work went up. Moreover, Professor Gregg reported a dramatic decrease in the number of student emails she received, as ChatGPT helped them get unstuck on a problem before immediately asking the professor for help.

This example can apply to other disciplines. Large Language Models like ChatGPT have the potential to serve as a pseudo-teaching assistant or brainstorming partner. Precisely because the underlying model is built by establishing connections between different words, phrases, and concepts across massive quantities of text, they are often capable of arriving at surprisingly creative ideas. They can help students get unstuck, work through tricky concepts, or provide basic feedback. Rather than replacing an instructor, these tools can take on some of the drudgery of teaching, such as writing “passive voice” yet again in the margins of a student essay. Ideally, using Large Language Models in this way can free up our time to focus on higher-order concepts and thinking skills.

The third lesson I’ve learned is that we need to answer an overriding question from students: why am I doing this? By “this,” I’m referring to assignments, activities, skills, and approaches that educators have long taken for granted. We need to take the question of “why am I doing this?” more seriously and address it more directly. We need to drive home to students that it’s not about getting a grade for a finished product, but about honing the skills that go into creating that product. We then need to explain precisely and concretely what those skills are and why they are relevant and worthwhile. We can’t just gesture to things like “critical thinking,” “communication,” or “research” without making them relatable and applicable to students’ lives.

Or, we can use this as a chance to rethink some of our assignments entirely. Take a classic assignment in my own discipline of history: the essay. In HIST 1362: US History Since 1876, I might give students the following essay prompt: “Did the 1968 Democratic National Convention have a largely positive or a largely negative impact on modern American political history? Write a five-page essay that makes your claim, supported with both primary and secondary sources.” For years, the output for that assignment - the written essay - was a decent way to evaluate students on skills at the heart of historical inquiry: research, synthesis, critical thinking, and, of course, written communication and rhetoric. Today, GPT-4 can write that same essay better than 80-90% of my students. It can provide context, locate relevant scholarship, synthesize material into a clear argument, and use impeccable writing mechanics to make its claim.

Given this reality, I think it is completely reasonable for an overworked and overstretched student in my class to ask: why am I doing this? Why should I spend hours and hours on an assignment that a chatbot can do better than me and in a fraction of the time? After all, the vast majority of my students will never have to sit down and write a five-page essay as part of their job - something that will be even more true as Generative AI tools get further integrated into writing platforms.

Accordingly, I’ve started moving away from essays and towards other kinds of experiential assignments in the class. Instead of an essay prompt about the 1968 Democratic National Convention, I now have students engage in a multi-day roleplaying simulation in which they each take on the role of a delegate, protestor, or journalist attending the convention. The assignment helps them understand the historical significance of the convention, but in a way that is far more engaging, memorable, and effective than having them slog their way through yet another five-page essay.

If Generative AI has shaken up long-standing teaching practices in higher education, then ThinqStudio is exactly the kind of community to help us adapt to these changes. Teaching in an era of Generative AI will require experimentation, creativity, and a willingness to challenge the status quo - in other words, the guiding principles of ThinqStudio itself. From “playful pedagogy” to ungrading, ThinqStudio has long been helping instructors rethink some of the core tenets of higher education. Perhaps most importantly, ThinqStudio provides a community of people to think through these changes collectively. After all, none of us should have to navigate this unfamiliar landscape alone.

This might be obvious, but I should note that the following was not intended to be read on a webpage. This is intentional: one of the other reasons for posting this was that I want to pull back the curtain and show a real-life example of lecture notes. When students give presentations, I often find myself giving them the same sort of feedback about the differences between writing for someone to read versus writing for someone to listen (a paper versus a presentation). What follows is a very lightly edited version of my lecture notes from class on Wednesday, along with accompanying slides. Although it’s more formal and closer to written prose than many of my lectures, it’s still a challenge to scroll through online. There are repeated words and passages, sentence fragments, and course-specific references that are hard to follow for someone who hasn’t sat in my class for the past sixteen weeks. That’s kind of the point. I’m hoping all of this illustrates that difference between writing to read versus writing to listen.

One Last Lecture…

• A few of you are history majors, but I know for the majority of you, this is probably the last history class you’ll ever take.

• Which is a lot of pressure for me.
• This is it!
  • the last shot I’ve got for you to remember me.

• Now I should clarify that I don’t actually care if you remember ME, specifically,
  • me as a person or even me as a professor.
• Feel free to immediately forget every detail about me.

• But what worries me is that this class becomes one of those classes you immediately forget

  • a class that just kind of runs together with the dozens of other classes you took at Northeastern.
• So on that note, I want you to take that index card and write down one thing you’re going to remember from this class.

• Five years from now, what are you going to remember? This can be anything you want: - a big theme or topic - an assignment or activity - A lecture or a story - A factoid, person, event, whatever you want …

• I’m doing this partly because I’m curious what you find most memorable.
• But I’m also doing this because I want you to, in fact, remember this class.
• That could be the content of the course, and I’m certainly hoping you learned new things about the western United States.
• But in another sense, I don’t actually care that much about the nitty-gritty details of the content, either.
  • I don’t care whether five years from now you remember the Battle of Glorietta Pass
  • Or, to use an example of something I messed up earlier this semester - whether it was James Marshall or John Marshall who discovered gold at Sutter’s Creek
• I’m guessing none of you are going to become a historian.
• But whatever career or life path you end up choosing, I hope that you remember some of the core principles of historical thinking that we’ve talked about over the past couple of months
• I hope you come away from this class with a curiosity about the past
• The relationship between geography and history
• A thoughtfulness about myth, memory, and history
• And above all, an appreciation for how the past continues shapes the present …
• So why do I think you should you remember this class?
• The answer has very little to do with what happened a hundred years ago in the West
• And everything to do with the present world we live in today
• To try and drive home what I mean by that, I’m going to look at an event from the present-day
  • this particular event happened five days ago
  • right across the Charles River, just a few miles away
• I want to look at this event to show you why history is so important

Boston Globe

• Last Friday night, several Cambridge Police officers responded to multiple calls
  • that a naked man was wandering down the middle of Mass Ave
  • he was acting erratically, and had thrown his clothes at a woman
• After confronting the man and talking to him for several minutes,
  • eventually the man in question approached one of the officers
  • Which led a second officer to tackle him to the ground,
  • He struggled, they try to restrain him, and
  • During this struggle one of the officers punched the man four or five times in the stomach
• They subsequently arrested him and charged him with “indecent exposure, disorderly conduct, assault, resisting arrest, and assault and battery on ambulance personnel”
• It turned out this man, Selorm Ohene
  • is a black mathematics major at Harvard,
  • and news outlets across the country have picked up the story
• So how does studying history help us understand what happened last Friday night?

• I know you’re sick and tired of seeing this slide by now,
• But I show it so many times for a reason
• History is an act of interpretation
• I mean this on two different levels
• First, on a basic level, we need to interpret what actually happened that night
• If we treat this as a historical event, we need to get the details of what happened
  • the who, what, when, and where
• Which, as you’ve hopefully learned in this class,
  • Is never as easy as it might seem
• It requires you to look at as many different sources as you can

Cambridge Police Department Report

• For this particular event,
  • we might read the police report filed by one of the officers
  • It says how they interviewed two of the student’s companions, who said that Ohene might have ingested hallucinogens, which could explain his erratic behavior
  • And goes on to say:
  • “As Ohene [the name of the student] and I conversed, I observed him clinch both his fists. Ohene started to take steps towards officers in a an aggressive manner. I perceived this as a threat and thought an attack was imminent.”
  • It goes on to say, “We gave Ohene verbal commands to give us his hands which he did not. Unable to pry Ohene’s hands from underneath his body, I delivered approximately 5 strikes with a closed fists to the area of his stomach.”

Boston Globe

• But as we’ve learned, we can’t rely on just one source
  • Especially one that is narrating an event from a single perspective
  • Fortunately for future historians studying 2018
  • Cell phones are everywhere
  • Which meant that onlookers actually took several videos of the event
  • In this case, we’d want to look at these videos to try and piece together the details of what happened
• This is really no different from the kind of analysis you’ve done in this class
• A cell phone video or a police report - these are primary sources
  • Think about the maps of New Spain that you looked at, that had no idea what was in the interior of the West
  • Or anti-immigrant political cartoons
  • Or, perhaps most relevant, first-hand accounts of the events at Sand Creek
• Historical thinking forces you to look at these kinds of sources critically
  • to try and take into account not only what it says,
  • but also who made it
  • under what conditions
  • from what perspective
  • and for what audience
  • What are these sources missing?
  • What are their limits?
• Again, these are crucial skills not just for understanding something that happened a hundred years ago, but for understanding something that happened RIGHT NOW, a few miles away, in 2018.

• The second kind of historical interpretation involves getting beyond this basic level of what happened
• Because understanding the past is not just collecting facts, dates, and names
• You have to piece them together into a larger narrative or interpretation in order to understand the significance of these kinds of events, why they matter and how they fit together.

• In the last couple of days, two competing narratives or interpretations have emerged about last Friday’s arrest
• On Sunday, the president of Harvard, again, the school where this man was a student
  • Called the incident “profoundly disturbing”
  • Activists have taken it up as an example of police brutality and the kind of racism that infects our criminal justice system
• On the other side of the ledger,
  • the Cambridge police commissioner held a press conference on Monday
  • where he expressed his full-throated support for the officers in question
  • saying they were simply taking actions to protect bystanders, themselves, and the man himself
• Again, two conflicting interpretations not necessarily of the details of what happened,
  • but of the meaning and significance behind it
• If we think of this event as historians, we need to evaluate both of these interpretations

• To do that, you have to use that other skill that I keep yammering on about: historical empathy
  • the ability to put yourself in the shoes of the people who lived in the past and see the world through their eyes
• I would argue that that if you can pull off historical empathy
  • of understanding how the world looked through the eyes of a person who lived a hundred years ago
  • a world that is was so incredibly, radically different from our own
• If you can do THAT, then you’ll be better equipped to put yourself in the shoes of someone living today

Boston Globe

• Let’s go back to the interpretation put forward by the Cambridge police commissioner
• Empathy forces us to take his interpretation of the events seriously.
• So let’s think about it from the perspective of one of his officers who was there
  • There’s a naked man wandering in the middle of a busy street
  • He’s incoherent and behaving erratically,
  • He’s thrown his clothes at someone
  • And you’ve received six different 911 calls
• So on a basic level, you can understand why they might have approached the situation with a certain degree of caution and suspicion

• More broadly, though, that skill of empathy involves understanding the CONTEXT of an event
• In the case of last Friday’s arrest, there might have been other reasons why the officers decided so quickly to resort to force

Boston Globe

• In fact, just one day before the event a police officer in Yarmouth, Massachusetts, was killed while serving an arrest warrant
• I can’t know for sure, but these Cambridge police officers were probably aware of this
• Because the wider law enforcement community across Massachusetts had been offering their condolences to the officer’s family for the past 24 hours
• Perhaps they had this event on their minds,
  • perhaps they were thinking about their own families
  • maybe all of this put them on edge when they arrived at the scene last Friday night
• This is what I mean by context

• But historical empathy does not mean that all narratives and all interpretations are somehow equally valid
• It’s not just trying too see both sides and leaving it there
• At some point, you’ve got to choose an interpretation based on the information, the context, and your own analysis

Boston Globe

• In this case, my own interpretation falls a lot closer to the other side of the ledger
• The one that sees this event as “profoundly disturbing”
• The reason for this is that this event is not taking place in a vacuum

• When police officers use violence against a black man
• It is deeply entangled within a national conversation that has emerged over the past few years
• The Black Lives Matter movement
• That has called attention to a recurrent pattern of police violence against people of color

Eicardo Anthony

• It’s a conversation that was sparked by the killing of Michael Brown at the hands of a white police officer in Ferguson, Missouri in 2014
• Followed in short order

NY Daily News

• …by the killing of Eric Garner in New York City

AP / Los Angeles Times

• …and Tamir Rice in Ohio

Los Angeles Times

• This isn’t some recent phenomenon
• It’s part of a much longer history
• Last Friday’s events call to mind Rodney King,
  • a black taxi driver in Los Angeles who was pulled over in 1991
  • and severely beaten by members of the LAPD
• Despite the fact that it was captured on video…
• A jury acquitted the officers in question
• All of which ties into an even longer pattern in this country
• Of Americans seeing black men
  • As violent
  • As dangerous criminals
  • or even animals

New York Times

• You can see this if you start using another one of those historian’s skills
  • the close reading and analysis of sources
• And apply those skills to the grand jury testimony of the white officer who killed Michael Brown
• You can see how he described the eighteen-year-old in precisely these kinds of animalistic, violent terms
• Testifying that:
  • “I’ve never seen anybody look that, for lack of a better word, crazy. I’ve never seen that. I mean, it was very aggravated…aggressive, hostile.”
  • He “had the most aggressive face. That’s the only way I can describe it, it looks like a demon, that’s how angry he looked.”
  • And finally, when he explained why he shot Michael Brown no fewer than twelve times
  • he explained it was because the teenager seemed unstoppable, that “it looked like he was almost bulking up to run through the shots”

Cambridge Police Department Report

• Again, applying those close reading skills to another document
• The police report from this past weekend
• You can see parallels in how the officer describes the Harvard student Selorm Ohene
  • Aggressive, hostile, intimidating
  • Calling attention to the student’s physicality, that he was muscular
  • And how he didn’t even seem to feel pain
• This language fits within a larger pattern in which Americans of all stripes and political persuasions see black men as threatening and dangerous
• This is partly what I mean by the need to understand a larger historical context

• But to understand events, you also have to understand the local history of individual places
• Remember that one of the themes of this class was the relationship between history and geography
• How local context shapes the history of particular places
• Think back to that exercise you did near the beginning of class,
  • using crayons to draw that caricatured View of the World from Northeastern University
• Part of the reason why I had you do this
• Was that it forces you to think about the perspective of how the world looks from a particular place
• And if you look at the world from the perspective of Boston
  • you start to understand the importance of this local context behind last Friday’s event

Boston Globe

• Just this year, a Boston police officer shot an African immigrant driving an ATV
• It turned out that the officer had a long history of making racist posts on online message boards under the screen name, and I swear I’m not making this up, “Big Irish”

Boston Globe

• Or the case of another Boston police officer who was convicted of racially motivated violence in 2015
  • when he drunkenly attacked his Uber driver
  • shouting racial insults as he did so

New York Times

• Or go back a few more years, to 2009, and to Cambridge itself, not all that far from where this student was arrested last Friday
• In 2009 a black Harvard professor named Henry Louis Gates Jr.
  • had forgotten his keys and accidentally locked himself out of his house
  • When he tried to get back in
  • a neighbor called the police because she saw a black man robbing a house
  • A Cambridge police officer arrived at the scene, got into a confrontation with Gates, and led him off in handcuffs
  • Again, this is a nationally respected professor at the most famous university in the country
  • Arrested for breaking and entering into his own home

CNN

• Of course, after this event happened,
  • yet another police officer in Boston sent an incredibly racist email about Professor Gates
  • Before claiming that, quote, “I am not a racist”

TIME

• Or we can go back even further, to the 1970s
• When white Bostonians became enraged after the city instituted bussing to bring black students to largely white schools
  • to try and desegregate the city’s school system
• Leading to this famous photograph of a white Boston man attacking a black civil rights leader with an American flag, near the steps of city hall

• All of this calls to mind that William Faulkner quote, “The past is never dead. It’s not even past.”
• Everything has a history
  • And while not all of that history is equally important for understanding the present
  • Part of why we study history is to figure out which parts of it matter and how they all fit together
• In the case of the police violence last Friday against the black student in Cambridge
• I would say that all of this history that I just showed you, it does matter.
  • Applying historical thinking to this event is what allows us to understand it better
  • To interpret the event, not only in terms of what happened but how it fits into a larger context
  • To see it not as some isolated, present-day event
  • But as part of a larger historical narrative
  • a larger historical structure
• If you remember nothing else from this class, I want you to remember this
• That history is not neutral
  • it’s not simply stuff that happened
  • it’s not dead
• History is a living, breathing thing
• The past exercises enormous power over the present, over our world today
• To understand this present world
  • it’s absolutely crucial that we understand that power
• And THAT is why I want you to remember my class.

On November 5th, Kimberly Latta wrote a Facebook post describing how the co-founder of the Literary Lab, Franco Moretti, “sexually stalked, pressured and raped me” while Latta was a graduate student and Moretti was a visiting professor at UC Berkeley in 1984-1985. Moretti, who recently retired from Stanford, denies the accusation. It’s the first time that someone I’ve worked with has been accused of sexual assault, at least that I know of - a sad and necessary caveat. Latta’s post made me shudder. You should read it. Although I cannot speak on behalf of the Literary Lab, I can speak for myself: Kimberly Latta, I am sorry. I believe you, I support you, and I am sorry.

I don’t have anything profound to add to all of this, but the past two days has driven home some of the ways that power and patronage operate in academia. Because along with the anger and the sadness and sympathy, I also found myself wrestling with what to do or say. Or, if I’m being truly honest, what not to say. Do I risk estranging colleagues or burning professional bridges at the Literary Lab? Do I really want to call attention to my own personal connections to Franco Moretti? Wouldn’t it be safer to just say nothing and let it all pass by? Take a moment to think about how absurd that is. Here I am, a white male professor with a tenure-track job at a private research university - the walking, talking embodiment of academic privilege - fretting over the risks of publicly expressing support for a rape survivor. So why the hesitation? Patronage.

However superficial my relationship with Franco Moretti may have been, it has nevertheless benefited my career. My affiliation with his lab allowed me do exciting interdisciplinary research, talk about that work at conferences and in job applications, and meet influential people from across the field. In short: patronage. At this point it’s fairly obvious how the power wielded by men can silence survivors. But a less visible web of patronage knits together the wider culture of silence. As a beneficiary of that patronage, I want to say this again: Kimberly Latta, I am sorry. I believe you, I support you, and I am sorry.

I’m not going to venture into the dark, twisted swamp of how Donald Trump and his supporters abuse history. I doubt any of them will read this and I doubt even more that anything I write would change their minds. Instead, I’m writing this for those on the left who are marshaling “history” as a tool of resistance. Two core ideas have been articulated again and again in the last two weeks: “we cannot repeat the mistakes of the past” and “history will judge you.” Broadly speaking, I share these sentiments. And to the degree that they help spur action, let’s continue to use them. But we also need to understand their limitations and the ways in which they can actually be counter-productive.

The first idea is a variation on classic “those who do not understand the past are doomed to repeat it.” You can see this, for instance, in the parallels being drawn between Trump’s immigration ban and when the United States turned away Jewish refugees in the 1930s, hundreds of whom later died in the Holocaust. There are some surface similarities between the two episodes, which is part of what makes it an effective rhetorical tool. This parallel nevertheless implies that these sorts of historical episodes were “mistakes” - momentary, if calamitous, fuck-ups from an otherwise virtuous norm. In the case of immigration, it assumes that closing off our borders is a deviation from our true historical identity as a melting-pot of immigrants. The history of the United States is, indeed, a history of immigrants, but it is also a history of immigrant-haters.

The Jewish refugee episode, as reprehensible as it was, did not represent some momentary lapse. It was firmly anchored in the xenophobia, isolationism, and anti-semitism of the era. Prejudice, fear-mongering, and the exclusion of particular religious or racial groups from entering the United States have been part of national politics for a long, long time. Just because a Muslim ban fits within the longer tapestry of American history, however, doesn’t make it justified - any more than a history of denying women a political voice justifies repealing the nineteenth amendment. A ban on Muslims is wrong because it violates the moral and legal standards of the United States today. We shouldn’t have to turn to history to make that case.

Perhaps most broadly, the “doomed to repeat history” line can also lead to false equivalencies between the present and the past. I’m currently living in Munich, Germany. As I walk past the very same plazas where Hitler held rallies and gave speeches, it’s hard not to hear those echoes in today’s political climate. Listening too closely to them, however, is often counter-productive. Donald Trump is not Adolf Hitler, as much as they might share disturbingly similar political tactics. If you start to seriously compare the two, Trump - for all of his terribleness - is always going to fall short of the man who engineered the most destructive global war in history and the calculated genocide of millions of people. In my mind, this lets Donald Trump off the hook. We shouldn’t be evaluating his actions primarily in terms of their similarities to past fascist regimes. That’s an incredibly low bar.

The problem with historical parallels is that they can also blind us to actions that don’t have easy historical precedents. If we’re only on the lookout for the symptoms of Hitler-esque fascism, we’re going to miss all the other danger signs of, say, a more modern autocracy. In fact, given the sheer intensity and speed with which this administration has assaulted so many things I care about, I think that constantly trying to draw lessons from the past can actually serve as a distraction. It’s well worth debating which of the Andrews (Jackson or Johnson) serves as a better parallel for Donald Trump; in the meantime, he’s installing a climate-change denier as head of the Environmental Protection Agency or making a racist ethno-nationalist one of the most powerful people in the world.

The second line that I’ve heard again and again is that “history will judge you.” This is most often aimed at Republicans in an attempt to force them to weigh long-term legacies against short-term agendas. Yes, Trump might help you repeal Obamacare. But do you want to be the next George Wallace? Are future generations going to laud you for standing up for basic decency or deride you as someone who helped pave the way for hatred and bigotry? I happen to firmly believe our grandchildren will look back on Donald Trump in horror, but predicting the future is a slippery business. There is no archive that I can draw from to “prove” where our country will end up and how future Americans will think about the Trump presidency. Someone on the other side of the political spectrum can just as easily believe that history will look back on Donald Trump as a savior, and that anyone who supported him will have helped steer the United States off the dangerous path it had been hurtling down. When we say to Paul Ryan, “history will judge you,” it’s pretty easy to meet that accusation with “not-uh.” Ultimately, the only way to tell who’s right is to wait and see.

“History will judge you” rests on the assumption that a) our country inherently follows a particular kind of trajectory, and that b) it will continue to do so. The “arc of history” is one of Barack Obama’s favorite lines, itself an adaptation from (most famously) Martin Luther King Jr.’s quote that “the arc of the moral universe is long, but it bends toward justice.” This might be comforting for those of us hoping that our progressive values are destined to win out in the end, but it also offers false comfort. Historical trajectories look very different depending on whose trajectory you consider and what you decide to use as a end point.

The Reconstruction Era is a useful example. For many, many years the dominant understanding was that the North’s attempt to “reconstruct” the former Confederate states was needlessly punitive and a tyrannical overreach of federal power. Violent efforts by paramilitary groups to suppress freed slaves were, in fact, honorable attempts to restore the proper, natural social order of the South. If you were to look backward from, say, the 1940s, “history” had indeed judged these groups and found them not only innocent, but heroic. For African-Americans, meanwhile, the moral arc of history had in many ways bent backwards - from the dramatic gains made during Reconstruction to the crushing boot-heel retrenchment of the Jim Crow era. History does not follow one single trajectory that moves inexorably upwards.

If history is an arc, then it is an immensely pliable one that can be bent in any direction. Over the next four years, Donald Trump and Republican politicians are going to bend that arc in disturbing ways. There is no way around it. Whatever built-in resistance that arc might have - institutions, political norms, democratic checks and balances - have proven much weaker than we originally thought. It’s going to bend, and it’s going to bend badly. History is not self-correcting. There’s no guarantee that it will magically straighten itself out somewhere down the road. The thing we need to do is grab on as tight as we possibly can and keep it from bending in ways that prove permanent. I do not want to leave it in the hands of “history” to judge Donald Trump. We - the American public - need to judge him right the fuck now.

Part II. The Overland Trails, 1840-1860

Between 1840 and 1860 several hundred thousand people traveled westward across the United States, most of them ending up in California, Oregon, and Utah. Their migration has become a foundational element of American history, conjuring up visions of covered wagons and hardy pioneers. Or, if you grew up playing the educational computer game The Oregon Trail: floppy disks, pixelated oxen, and exciting new words like “dysentery.” The topic has been exhaustively studied by genealogists, historians, and millions of schoolchildren over the years. American Panorama attempts to break new ground on what is, like the trail itself, well-trodden soil.

The Overland Trails follows a similar visual layout as The Forced Migration of Enslaved People, with multiple panes showing a map, a timeline, aggregated data, and the expandable text from twenty-two trail diaries. Far more so than The Forced Migration of Enslaved People, however, it puts these written narratives into the spotlight. The visualization includes the full text of each diary rather than brief excerpts. Clicking on a specific diarist allows you to read all of their entries, with a linked footnote to the original source. As you scroll through the entries, clusters of dots track the progress of the emigrant’s journey on the map as they pass between landmarks like Courthouse Rock or Fort Laramie.

Two other panes provide context for that particular year: a short summary of trail activity and a small map breaking down the estimated annual migration to California, Oregon, and Utah. The timeline uses small multiples for each year that plot the seasonal progression of emigrant journeys on its x-axis and, somewhat confusingly, the (horizontal) longitude coordinates of these journeys on its vertical axis. Timeline aside, the overall reading experience is both intuitive and seamless. More importantly, the visualization strikes a balance between detail and context, weaving the full text of individual sources within a larger spatial and historical tapestry. In many ways, this is digital design at its best. But why does this elegant design matter? What is the historical payoff? The Overland Trails makes two contributions to the topic of westward migration - one archival and the other interpretive.

First, The Overland Trails gives us not just a new, but a better platform for reading and understanding the topic’s source base. The trail diary was a genre unto itself during the mid-nineteenth century. They were often written to serve as a kind of guide to help family or friends follow them westward, recording daily mileage, landmarks, trail quality, and the availability of water and grass. These details made the diaries immensely helpful for future emigrants, but immensely boring for future historians. Take an entry written by James Bennett on July 12th, 1850:

Friday 12th-After ten miles travel this day over a heavy, sandy and barren road, we reached Sweet Water river, where we took dinner. Here we found the grass very short and as our cattle were nearly exhausted by hard work and scant feed, we drove off the road five miles to the right, where we found excellent grass and a good spring.

Now imagine reading thousands of entries exactly like this one. You start to get hungry for anything that breaks the monotony of the trail: white-knuckled river crossings, exchanges with passing Indians, or fiery arguments about whether or not to travel on the Sabbath. Moreover, as a reader we often don’t care all that much about where these juicy episodes took place - does it really matter if they occurred in western Nebraska, northern Utah, or eastern Oregon? The nebulous space of “The Trail Experience” serves as a stand-in for specific geography of where things happened. But the loss of geographic context risks distorting the lived reality of nineteenth-century emigrants. For them, trail life was overwhelmingly defined by geography: boring, repetitive, grinding travel along an established trail itinerary, with mileage tallies or landmark notations acting as a means of marking their progress through that geography. American Panorama captures the experience of overland travel far more effectively than simply reading trail diaries on their own. As simple as it sounds, linking individual entries to their location on a map illustrates the small-scale, incremental geography that made up this massive, large-scale migration.

The second historical contribution of The Overland Trails involves a broader spatial reinterpretation of westward expansion. The phrase itself - “western expansion” conjures up the image of a wave of Anglo-American settlers washing over the continent. This was the geography embedded in Manifest Destiny iconography and Frederick Jackson Turner’s famous frontier thesis.

Source: Wikimedia Commons

American Panorama presents a much different geography. Western migration was not a wave; it was a narrow river. Hundreds of thousands of people may have traveled across the western interior between the 1840 and 1860, but they did so along a severely restricted corridor of travel. This might seem obvious; the Overland Trail was, after all, a trail. But the trail’s meaning has come to embody a certain idea of mobility, not just in terms of traveling westward to Oregon or California, but of experiencing and claiming the vast swath of land that lay in between. When mapped, however, the journeys of twenty-two emigrants resemble tightly braided cords that only gradually fray as they approach the Pacific Coast. Overland travelers operated in a tightly constrained space.

To take one example: although emigrants technically traversed from one side of Nebraska Territory to the other, most travelers didn’t see very much of it. The grinding necessity of daily travel kept them pinned along the Platte River. American Panorama illustrates just how narrow this pathway was and how infrequently emigrants deviated from it.

In the mid-nineteenth century, the interior of the western United States was seen as a region to pass through as quickly as possible, an area that had long been labeled “The Great American Desert,” or in historian Elliott West’s words, “a threatening void.” (The Contested Plains, 122) Much of the western interior was made up of territory that was ostensibly claimed by the United States but that remained largely ungoverned and unsettled by Anglo-Americans. American Panorama effectively recreates this geography through visual design: bright, sharp lines track the emigrants’ journeys along the trail, interspersed with landmarks and forts shown in equally bright colors. This tightly demarcated trail geography pops out from the map as it snakes across a minimalist base layer entirely devoid of the familiar political boundaries of states or territories. Instead, the underlying map consists of terrain, sparse water features, and the locations of Indian groups such as the Cheyenne in the central plains or the Goshute near Great Salt Lake. The Overland Trails manages to capture the experience of traversing a semi-arid, mountainous region still occupied by native people, one that was seen as largely off-limits for Anglo-American settlement.

The project’s cartographic achievement comes with a cost, however. The presence of native groups played a crucial role in shaping mid-century views of the interior. As historian Susan Schulten notes, “erasing Native Americans from both mental and actual maps” (29) was a central process in the eventual shift from seeing the western interior as an inviting area to settle rather than a forbidding area to traverse. To their credit, the designers of The Overland Trails put native people back on the map. The problem comes from the way in which they do so. The mapmakers label Indian groups using a muted gray color that is nearly identical to the map’s base terrain. Moreover, changing the zoom level causes some labels to shift locations or disappear entirely in order to avoid overlapping with the trail and its landmarks. The overall effect is to weave native groups into the natural landscape, making them visually analogous to the map’s rivers or mountains. This cartographic design ends up conflating native people and the environment - a deeply problematic notion that remains stubbornly lodged in the popular imagination. The visualization builds a marvelous stage for overland emigrants, but its set design turns Indians into a backdrop.

I don’t mean to quibble over (literal) shades of gray. After all, the map’s creators made a concerted effort to include Indian groups - the same can’t be said of other many other historical projects, digital or otherwise. But the project’s cartography highlights a common tension between digital design and historiography. From a design standpoint, the creators of The Overland Trails make all the right decisions. Brightly colored overland routes are foregrounded against a muted base map, including unobtrusive gray labels of Indian groups that give readers contextual information while keeping their attention firmly focused on the emigrant journeys themselves. When those same labels disappear or change locations depending on the zoom level, it helps avoid visual clutter. The problem is that effective digital design can run headlong into fraught historiographical issues, including the contentious idea of the “ecological Indian” and a longstanding cartographic tradition of using maps to marginalize and erase native claims to territory in the West.

Visual design is not the only sticking point for The Overland Trails and its place within western historiography. The visualization is, at its core, a digital archive of primary sources. As I’ve already noted, its interface contributes a new and fascinating way of reading and understanding these sources. What troubles me is the privileging of this particular archive. To be blunt: do we really need a new way of reading and understanding the experience of mostly white, mostly male pioneers whose stories already occupy such a central place in American mythology?

The historical commemoration of overland emigrants began almost as soon as their wagons reached the Pacific Coast. Western pioneer associations held annual conventions and published nostalgic reminiscences that romanticized their journeys. Historians, meanwhile, largely followed the blueprint of Frederick Jackson Turner, who immortalized the march of pioneer-farmers carrying the mantle of civilization westward. Nearly a century passed before historians began to reassess this framework, from uncovering the ways that gender shaped life on the trail to, more recently, interpreting overland migration as a “sonic conquest.” (to use Sarah Keyes’s formulation).

More often than not, however, historical treatments of the Overland Trail still tend to resemble book titles like Wagons West: The Epic Story of America’s Overland Trails, or quotes like, “An army of nearly half a million ragged, sunburned civilians marched up the Platte in the vanguard of empire…they emerge from their collective obscurity to illuminate a heroic age in American history.” (Merrill Mattes, Platte River Road Narratives, xiv) The Overland Trails doesn’t explicitly advance this viewpoint, but nor does it move away from it in any substantive way. The informational text accompanying the visualization’s timeline can, at times, read like a “greatest hits” of western lore: the Donner Party, the Gold Rush, Indian fighting, and the Pony Express (its freshest material centers on Mormon migration). The visualization’s space constraints leave precious little room for important historical nuance, leading to generalizations such as “White settlement in the West was disastrous for Indians everywhere.”

To reiterate a point I made in the first part of my review of American Panorama: prioritizing user exploration over authorial interpretation comes with risks. I don’t want to minimize the significance of The Overland Trails, because it contributes a truly valuable new interface for conceptualizing nineteenth-century historical geography and the experience of overland travel. But the project uses a novel framework to deliver largely tired content. My guess is that its selection of content was based on the fact that these particular diaries were already digitized. This kind of pragmatism is a necessary part of digital history. But explaining the interpretive implications of these decisions, not just the nitty-gritty methodological details, often requires a more robust and explicit authorial voice than many digital history projects seem willing to provide.

My hope is that The Overland Trails will serve as a prototype for visualizing other movement-driven sources. To that end, American Panorama has given outside researchers the ability to build on this framework by making the project’s source code available on Github.  The Github repository highlights the open-ended nature of the project, as its creators continue to improve its visualizations. In a similar vein, American Panorama’s team has several new visualizations to come that examine redlining, urban renewal, and presidential voting.  I have high expectations, and I hope that other historians will join me in giving them the substantive engagement they deserve.

American Panorama: An Atlas of United States History was released in December 2015 by the University of Richmond’s Digital Scholarship Lab. It is a collection of four map-based visualizations focusing on different topics in American history: slave migration, immigration to the U.S., canal construction, and the Overland Trails. Each of these visualizations revolve around an interactive map, with surrounding panes of charts, timelines, contextual data, and primary sources related to the topic. If I could summarize the project’s historical contributions in a single sentence, it would be this one: American Panorama incorporates movement into the history of the United States. To be even more specific, the project shines a new light on the historical movement of people. Its three most compelling visualizations (foreign immigration, slave migration, and the Overland Trails) illustrate some of the most monumental shifts of people in American history. There are certainly other episodes of travel and migration worth studying - Indian Removal or the Great Migration immediately jump to mind - but those selected by American Panorama are certainly three of the most consequential.

Like most digital history projects, American Panorama is a collaboration. Unlike most digital history projects, it’s a collaboration between academic historians and a private company. The Digital Scholarship Lab’s Robert Nelson, Ed Ayers, Scott Nesbit (now at the University of Georgia), Justin Madron, and Nathaniel Ayers make up the academic half of the project. The private half of the partnership is Stamen Design, a renowned data visualization and design studio that has worked with clients ranging from Toyota and AirBnB to the U.S. Agency for International Development. Stamen is also, in the words of tech journalist Alexis Madrigal, “perhaps the leading creator of cool-looking maps.” Stamen’s fingerprints are all over American Panorama. The visualizations are beautifully structured, deeply immersive, and packed with information. In fact, data depth and data density are the hallmarks of these visualizations - I don’t think I’ve ever seen this much historical content visualized in this many different ways, all within a single browser window. Furthermore, the project’s visual interface presents a new and valuable framework to understand the scale of people movements in a way that written narratives can struggle to convey. Writing about thousands or even millions of people moving around over the course of years and decades can often devolve into an abstract swirl of numbers, states, regions, and dates. American Panorama makes that swirl intelligible.

The project encapsulates many of the current hallmarks of digital history. It is aimed at a broad public audience and was “designed for anyone with an interest in American history or a love of maps.” Relatedly, the project is exploratory and descriptive rather than explicitly interpretive, and offers only hints at how the reader should understand and interpret patterns. Outside of brief and rather modest textual asides, readers are largely left to make their own discoveries, construct their own narratives, and draw their own conclusions. The common justification for creating exploratory visualizations rather than argumentative or narrative-driven ones is that they encourage participatory engagement. Empowering readers to control how they interact with a visualization nudges them to delve deeper into the project and emerge with a richer understanding of the topic. But an exploratory framework hinges on a reader’s abilities and willingness to discover, narrate, and interpret the project for themselves.

To take one example, American Panorama’s Foreign-Born Population, 1850-2010 offers by far the strongest interpretive stance out of the project’s four visualizations: “American history can never be understood by just looking within its borders.” Even so, the creators consign their interpretation to a short, solitary paragraph in the About This Map section, leaving readers to draw their own conclusions about the meaning and implications of this message. The tech blog Gizmodo, for instance, covered the project’s release under the headline: “See The US Welcome Millions Of Immigrants Over 150 Years In This Interactive Map.” Internet headlines have never exactly been a bastion of nuance, but to say that the U.S. “welcomed” immigrants is, well, not very accurate. It’s also an example of the kind of historical mischaracterization that can arise when projects push authorial interpretation into the background.

Full disclosure: I know and deeply admire the work of Rob Nelson, Scott Nesbit, and Ed Ayers. They are very, very smart historians, which is why I found myself wanting to hear more of their voices. What new patterns have they discovered? What stories and interpretations have they drawn from these patterns? How has the project changed their understanding of these topics? The creators of American Panorama do not answer these questions explicitly. Instead, they allow patterns, stories, and interpretations to swim just beneath the surface. This was likely a deliberate choice, and I don’t want to critique the project for failing to accomplish something that it never set out to do in the first place. American Panorama is not an academic monograph and it shouldn’t be treated as one. Nevertheless, the project left me hungry for a more explicit discussion of how it interpretation and historical literature.

I’d like to offer my own take on American Panorama using equal parts review and riff, one that combines an evaluation of the project’s strengths and weaknesses with a discussion of how it fits into themes and topics in U.S. history. To do so, I’ve focused on two visualizations: The Forced Migration of Enslaved People, 1810-1860 and The Overland Trails. Fair warning: in true academic fashion, I had far too much to say about the two visualizations, so I split the piece into two separate posts. The first is below, and the second will follow soon. (Update: you can read Part II here.)

Part I. The Forced Migration of Enslaved People, 1810-1860

In some ways, Americans remember slavery through the lens of movement. This begins with The Middle Passage, the horrifying transportation of millions of human beings from Africa to the Americas. The focus on movement then shifts to escape, perhaps best embodied in the Underground Railroad and its stirring biblical exodus from bondage to freedom. But there was a much darker, and less familiar, counterweight to the Underground Railroad: being “sold down the river” to new planting frontiers in the Deep South. The sheer volume of this movement dwarfed the far smaller trickle of runaways: between 1810 and 1860 southern planters and slave traders forced nearly one million enslaved people to move southward and westward. The Forced Migration of Enslaved People, 1810-1860 helps us understand the scale and trajectory of this mass movement of human beings.

The visualization uses a map and timeline to illustrate a clear decade-by-decade pattern: enslaved people streaming out of the Upper South and the eastern seaboard and into the cotton-growing regions of the Black Belt (western Georgia, Alabama, and Mississippi), the Mississippi River Valley, and eastern Texas and Arkansas. It shows that this shift was not uninterrupted, but came in fits and starts. The reverberations of the 1837 financial panic, for instance, dampened and diffused this movement during the 1840s. An accompanying data pane charts the in-migration and out-migration on a state and county level: during the 1830s more than 120,000 slaves left Virginia, even as 108,000 slaves streamed into Alabama. None of these findings are especially new for historians of the period, but The Forced Migration of Enslaved People brings them into sharp focus.

On an interpretive level, The Forced Migration of Enslaved People helps reorient the locus of American slavery away from The Plantation and towards The Slave Market. This is part of a larger historiographical pivot, one that can be seen in Walter Johnson’s book Soul by Soul (1999). Johnson reminds us that American slavery depended not just on the coerced labor of black bodies, but on the commodification of those same bodies. It wasn’t enough to force people to work; the system depended first and foremost on the ability to buy and sell human beings. Because of this, Johnson argues that the primary sites of American slavery were slave markets in places like Charleston, Natchez, and New Orleans. Soul by Soul was an early landmark in the now flourishing body of literature exploring the relationship between slavery and capitalism. The book’s argument rested in large part on the underlying mass movement of black men, women, and children, both through slave markets and into the expanding planter frontier of the Southwest. American Panorama lays bare the full geography of this movement in all of its spatial and temporal detail.

There is a certain irony in using Walter Johnson’s Soul by Soul to discuss The Forced Migration of Enslaved People. After all, Johnson’s book includes a critique that might as well have been addressed directly to the project’s creators. He bluntly asserts that the use of maps and charts to illustrate the slave trade hides the lives and experience of the individuals that made up these aggregated patterns. Instead, Johnson calls for the kind of history “where broad trends and abstract totalities thickened into human shape.” (8) His critique echoes the debates that swirled around Robert Fogel and Stanley Engerman’s Time on the Cross (1974) and continue to swirl around the digital project Voyages: The Trans-Atlantic Slave Trade Database.

The creators of The Forced Migration of Enslaved People gesture towards the larger historiographical divide between quantification and dehumanization in an accompanying text: “Enslaved people’s accounts of the slave trade powerfully testify to experiences that cannot be represented on a map or in a chart.” Instead, they attempt to bring these two modes of history together by incorporating excerpted slave narratives alongside its maps and charts. Clicking on icons embedded in the map or the timeline reveals quotes from individual accounts that mention some dimension of the slave trade. This interface allows the reader to shift back and forth between the visual language of bars, dots, and hexbins, and the written words of formerly enslaved people themselves. The Forced Migration of Enslaved People uses a digital medium to present both the “broad trends and abstract totalities” and the “human shape” of individual lives. One of the analytical and narrative payoffs of an interactive interface is the ability to seamlessly move between vastly different scales of reading. The Forced Migration of Enslaved People breaks important new ground in this regard by blending the macro scale of demographics with the micro scale of individuals.

Ultimately, however, the project’s attempt to combine narrative accounts and quantitative data falls short of its potential. On the whole, the scale of the individuals recedes under the scale of the data. The problem lies in the way in which the project presents its excerpted quotes. Flurries of names, places, events, and emotions appear divorced from the broader context of a particular narrative. Reading these text fragments can often feel like driving past a crash on the side of a highway. You might glimpse the faces of some passengers or the severity of the wreck, but you don’t know how they got there or what happens to them next. Then you pass another crash. And another. And another. The cumulative weight of all these dozens of wrecks is undeniable, and part of what makes the visualization effective. But it’s also numbing. Human stories begin to resemble data points, presented in chronological, bulleted lists and physically collapsed into two-line previews. The very features that make narratives by enslaved people such powerful historical sources - detail, depth, emotional connection - fade away within this interface. Narratives give voice to the millions of individuals whose stories we’ll never hear; The Forced Migration of Enslaved People helps us to hear some of those voices, but only briefly, and only in passing.

Historians characterize the years leading up to the Civil War as a period defined by sectional conflict between North and South. The abolition of slavery was not the major flashpoint for this conflict; rather, the expansion of slavery into western states and territories was the primary wedge between the two sides. The issue would come to define national politics by pitting two competing visions of the nation against one another. The Forced Migration of Enslaved People reminds us that this was not just an ideological or political issue, but a spatial issue rooted in the physical movement of hundreds of thousands of people into areas like the Black Belt and the Mississippi River Vally. By the 1850s, many northerners feared that this great heave of slaveholders and enslaved people would continue onwards into the Far West. The Forced Migration of Enslaved People forces us to take those fears seriously. What if the visualization’s red hexbins didn’t stop in the cotton fields of eastern Texas? What if its timeline didn’t end in 1860? Southern slavery did not stand still during the antebellum era and its demise was far from inevitable. This visualization gives us a framework with which to understand that trajectory.

I doubt that most Americans would put slave traders and shackled black bodies within the historical pantheon of great national migrations, but American Panorama injects this vast movement of people into the history of the antebellum United States. In the second part of my discussion, I’ll turn my attention to a much more familiar historical migration unfolding at the same time: The Overland Trails.