This headline feels surprisingly human to me. My own brain still runs on knowledge, habits, and assumptions I picked up in school during the 1990s. Big Tech might imagine the future replacing the past, but historians know better. The past is still running in the background, shaping what comes next.

As I learn more about autistic masking, I’m noticing how differently unmasking is experienced depending on race, class, gender, and other forms of privilege. Some can treat masking as something to calibrate in the pursuit of goals. Others face real danger or exclusion when they let the mask drop.

As Autism Awareness/Acceptance Month closes, I wanted to share a new paper. It introduces the concept of "visibility logic" and describes masking as a "double bind" for autistics. Masking can lead to burnout, and unmasking causes autistics to be seen as incapable: www.tandfonline.com/doi/full/10....

Thanks to @historians.org for distributing today a call for members in the "Neurodivergence and the Historical Profession" CARE Group, which I help facilitate. Click here for more info: www.historians.org/community-ca... If this sounds interesting, please reach out for more information and join us.

I wouldn’t read the book right before seeing the movie—you might focus on what’s missing. But this still got to me. Rocky and Grace’s friendship brought the film to life, even with the science dialed back a bit. The cinematography was breathtaking, at least for humans, and the music almost Eridian.

I’ve been adding salmon to my weekly routine. Today the waitress remembered my order — main, sides, and drink — without a single prompt, and without making it awkward. Small moments like that make me smile. #AutismJoy

Smart take in Dan Cohen’s newsletter. AI excels at well-defined problems and linking existing knowledge, but is less adept at expressing new ideas or posing innovative questions. As Hannah Fry notes, it’s strong at interpolation—less so at extrapolation beyond current boundaries (see 7:06–9:31).

Scanning old files, I found my high school interest survey. With hindsight, it’s striking how closely it matched my path—programming, libraries, a history major, and broad interests across the social sciences and humanities.

Shipped 1,000+ books from my library to be scanned. First PDF arrived today. Not the full “wearable memory” future (↓), but soon my shelves will be searchable. In the AI era, the challenge isn’t just access to info, but preserving and strengthening the human role in telling stories about the past.

AI debates right now feel split across three worlds: • Work: AI won’t replace you—but someone who uses it better might. • AI labs: We can reward correct answers in math. What’s the reward signal for a great novel? • Universities: Guard the essay, explore the research assistant.

Most sincere: “Whatever our souls are made of, his and mine are the same.” Heaviest: “I have not broken your heart — you have broken it; and in breaking it, you have broken mine.” Most fateful: “Kiss me, and let us both be damned.” Love as ontology, injury, and doom.

A former Google AI exec argues PhDs are becoming obsolete in the age of AI. His perspective is that college is for credentialing and the soft skills for the workplace can be obtained elsewhere. He might be underestimating how college is an ideal place for teaching critical thinking and creativity.

Daniela Amodei—a former literature major and now president of Anthropic—argues that as AI excels at STEM, the humanities matter more, not less (see ↓). That hope is worth holding onto, even as AI disrupts work and firms like Anthropic still profit from us. Interview: www.youtube.com/watch?v=qUDG...

This image in today’s Washington Post feels like a turning point. Libraries and bookstores long mediated how readers encountered knowledge. Digitization began tipping the scales. AI may further reshape how books are accessed and used, including in proprietary systems. That shift is worth noticing.

This article reminded me of something Ben Vison said in his presidential address to the AHA last week. He called it the "massification of higher education." I'm looking forward to the publication of his speech to explore this concept in more depth and what it might mean for the future of higher ed.

Ben Vinson, in #AHA26 presidential address, expresses his hope that academic historians, who specialize in understanding and explaining change over time, are the best positioned individuals to become agents of change in society, impacting their students, institutions, and ultimately the common good.

It’s been fascinating listening to the many #AHA26 sessions devoted to the upcoming 250th. I like the many ideas for how historians can join the conversation of commemoration. I hope get an overview and assessment at some point of the collective impact of these contributions by historians in 2026.

Same hotel, different scholarly languages. SSHA (Nov, Palmer House): PowerPoint + extemporaneous talk. #AHA26 today at Palmer House: essays read from the lectern. The differences in linguistic registers & discursive strategies has me curious about the construction of scholarly authority and culture.

I worry some AI advocates in history are at risk of functional amnesia. They talk about scale and extraction, and gravitate towards AI-friendly questions of fact, effectively operationalizing positivism. Will they recall postmodern ways of reading texts or let AI outputs shape their interpretations?

I’m back in Chicago, for the AHA, after attending SSHA two months ago. I’m walking into a session about AI rewriting the past, after delivering my own paper with a similar title in November. Is this what doubly Déjà vu feels like? In any event, I’m interested to hear what this panel has to say.

I’m getting excited. I’ve lined up recommendation letter writers. I’ve requested transcripts. I’ve drafted personal statements, updated my CV, and dusted off my master’s thesis. I’ve narrowed down possible doctoral programs to half a dozen. Here’s hoping that I land at one of them in Fall 2026.

This essay proposes using backprop to determine when content contributes to an LLM response and compensating creators accordingly. It reminded me of how many academics want credit & payment when their work is used to power chatbots. There would be challenges, but this sort of mechanism could emerge.

This is a great academic analysis of the settlement. It mentions the 7.5M title database of LibGen titles compiled by The Atlantic. I’ll add that, according to the settlement, the 7M pirated downloads by Anthropic from LibGen and PiLiMi yielded only 465k titles. Something does not add not add up.

I have questions (below is NYTimes today). Why did Turvey go to publishers vs research libraries? If Anthropic took 500k pirated books, did they buy millions from publishers? How does this impact Public Interest Corpus? Is $3k per pirated book insulting? What kind of response if 500k list released?

I can understand the accidental framing, given the discourse around LLMs and their hallucinations, but I think the odds are decent for the model encountering and recalling this set of facts from the more than 7,000 Victorian-era documents it was trained on. Perhaps we will see more historical LLMs.

This week I'm reading the new book _Artificial Historians_ and getting more ideas for the conference paper I'm working on about history and AI.

The AHA has just reported that it filed yesterday, along with ACLS and MLA, a motion for preliminary injunction on behalf of NEH, in response to the harms done by DOGE. I asked OpenAI’s o3 model to evaluate the motion and the model gives it a 70% chance of success. I hope we prevail in the courts.

A historian of science ponders the future of the humanities after feeding documents into AI and hearing an impressive podcast and then assigning students to have a conversation with AI on the history of attention and finds that chatbots have perhaps become better than humans at getting attention.

I just received this email that the AHA is exploring possible legal action to protect historical funding from the NEH and other agencies. With some help from ChatGPT, I found that similar advocacy at the federal level to protect the work of historians has been successful before. I hope we prevail.

I worked with ChatGPT on some estimates. About 60B words published in univ press books since 1969. And > 250T words avail for scraping on the web. That’s a 1 to 4k ratio. How would comp work for LLM training on academic books if they’re a drop in the bucket? What criteria would work for licensing?

I'm writing a paper for this year's Social Science History Association conference. I decided to read the call for papers against the grain. My working title is "Emergent History: When AI Models Rewrite the Past." If the paper is accepted, I'm hoping I don't get called Foolhardy during my session ;-)

It has been 16 years since I've attended an AHA meeting (www.historians.org/perspectives...). Today I dug up an old password, successfully logged into historians.org, and renewed my membership. I'll be in Bozeman next month at an AHA event to learn more about large-scale historical research with AI.

I read "Provocations from the Humanities for Generative AI Research" (arxiv.org/abs/2502.19190). This got me wondering about adding social sciences and focusing on history. I prompted o1 Deep Research to reimagine the paper (chatgpt.com/share/67c0e7...) and generate more provocations (see below):

I gave the new copyright guidance a quick scan and it appears there was little to no discussion of public domain or Creative Commons licensing. Will digital libraries built by AI that summarize and extract insights from primary sources have the option of online open access or become walled gardens?

This post (resobscura.substack.com/p/the-leadin..., HT @nic221.bsky.social) by @resobscura.bsky.social caught my attention. Perhaps taking a cue from the author, I enlisted the help of the o1 model in a close reading of the text. You can peruse the remaining response at chatgpt.com/share/6791ca....

This is about research presented at NeurIPS. I agree that LLMs need better coverage of history across time periods and regions. But are multiple-choice tests the best way to measure how LLMs understand history? When will we measure how well they reason when given primary sources & relevant context?

Yesterday, I saw a request for info on a county founder in a Facebook genealogy group. I asked Google Deep Search to generate a bio. It read nearly 40 websites, wrote an essay, and dropped 14 footnotes to cite its sources. I wonder how this capability will improve as more sources are digitized.

I’m enjoying today’s episode of the History in Focus podcast, which is a discussion with historians of what AI means for teaching and research:

Today I used Transkribus for a first pass of transcribing historical court records, which was handy because it provides an editor with line-by-line comparison to the image. I then gave the image and first transcription to ChatGPT and asked it to produce a final transcription, which it did well.

I presented a paper at SSHA recently. I tried to put critical data studies into conversation with U.S. census data record linkage 1850-1880. I was excited our outgoing president spoke about applying source criticism to census data analysis and that colleagues from IPUMS asked for copies of my paper.

I’m working on a history of Wetzel County. It was in the borderland region of the Upper Ohio River Valley and sent men to both Union and Confederate forces. I’m building a population database to match the genealogical profiles of these individuals with their military service and pension records.

I had an interesting chat with Claude today that started with this prompt: “If nineteenth-century American sources could be made machine readable through OCR and handwriting recognition, why isn’t there much of a coordinated effort to construct this sort of public domain corpora for training LLMs?”

I’ll be testing out some digital public history this month. I’m using public domain sources from Internet Archive, organizing their chapters by chronology and themes, generating podcasts via NotebookLM, creating historical images via ChatGPT, and assembling narrated slideshows for AI YouTube videos.

With white-collar jobs for recent college graduates declining this year, it appears ChatGPT is still bullish about what AI agents, which can further automate many tasks, will mean for the prospects of job seekers. I hope students and workers will get ready to utilize AI and not be left behind.

As someone who grew up with the Internet in the 1990s, I hope the Internet Archive isn’t wiped out by lawsuits. It reminds me of how Roy Rosenzweig anticipated the outcomes of either digital abundance or scarcity of historical sources in the 21st century: www.wired.com/story/intern...

What an interesting day. I became aware of _Literary Mathematics_ via a podcast and was intrigued enough to purchase a copy. And then my social media is blowing up about PhD-level reasoning from the o1 model that OpenAI made available today.

Today I searched the GPT Store for transcribing handwriting (historic). #1 OCR-GPT needed an authorization code and could not process the image. #2 Handwriting to Text GPT struggled with some parts. #5 Handwriting to text OCR (see link below) did the best. #7 Recognize Handwriting did second best.

I prepped some bots this weekend for the GPT store opening this week. I had in mind students as the semester starts. OpenAI didn’t require me to disclose my training data, so I included a parenthetical note with the source and the Creative Commons license that allows derivative commercial usage.