Academics Need to Wake Up on AI, Part II

Disclosure for my Bluesky friends: This post is 100% human-written (assisted with human-verified AI summaries of your critiques of my earlier post). Human mistakes are possible.

Earlier this week I argued that academics need to wake up on AI and offered ten theses on how agentic AI is changing social science research. The post went viral, especially after I revealed it was fully generated and posted by AI, leading to news features, over a million views, and over a thousand (often angry but also enthusiastic) responses across various social media platforms.

In hindsight, I should have done a few things differently. First, revealing that Claude wrote the original post (even based on my earlier social media writing) as a cheeky follow-up was a mistake. It distracted from the substance and gave critics an easy reason to dismiss the arguments. Rather, I should have been upfront about my basic setup.1 Second, I should have been clear that current agentic AI tools are better at doing most social science research tasks than professors globally.2 This doesn’t mean you necessarily get replaced, but it does mean the nature of your work will change. Third, the AI-generated post had minor, somewhat weird stylistic errors that better human-augmented editing would have caught.

In this respect, I want to highlight Solomon Messing and Joshua Tucker’s Brookings piece, published shortly after mine, which makes a much more persuasive version of many of my arguments—with concrete use examples, no unnecessary provocation or AI-generated text, and a more constructive vision for the future. If my post was too abrasive for you, read theirs instead.

I was deliberately provocative, and I stand by that choice. It backfired in some sense but worked in another sense: dozens if not hundreds of academics are now trying agentic AI tools who would not have otherwise yet. After reading most responses, I certainly changed my mind on a few things, but I am still convinced of my core claim that, because of already existing AI tools, our research workflow will have to change whether you like it or not. Here are ten more theses that came from my reflections.

11. Qualitative research and novel data collection will increase in relative value.

The strongest substantive critique of Part I was that it conflated “research” with the specific tasks AI handles well—literature reviews, data analysis, conceptual synthesis. Several respondents rightly pointed out that AI cannot conduct ethnographic fieldwork, interview detainees in illegal prisons, or spend years building trust with a community. They are absolutely right. My theses were primarily about (the currently dominant) quantitative and conceptual work in social science, and I should have been clearer about that scope.

But the implication is not that qualitative researchers should relax. It is that the relative value of original data collection—fieldwork, interviews, archival work, participant observation—is about to rise. If AI can synthesize existing literature and run standard regressions, then the premium shifts to the things AI cannot do: generating new data that did not previously exist, especially from hard-to-reach contexts. Qualitative researchers and field experimentalists should see this as an opportunity to do more great work they have comparative advantage in instead of transcribing their interviews or compiling literature reviews.

12. Due to “jaggedness,” AI opinions are polarized by beliefs in their utility.

Ethan Mollick describes AI’s capabilities as a “jagged frontier”—superhuman at some tasks, embarrassingly bad at others, in ways that do not map to human intuition. AI can write a serviceable literature review but struggle with a basic visual puzzle. It can synthesize findings across 500 papers but hallucinate a co-author’s first name.

This jaggedness explains why the AI debate in academia is so polarized. Critics point to the troughs; enthusiasts point to the peaks. Both are right about their corner of the frontier. The overlap with the qualitative-quantitative divide in social science is hard to ignore: researchers whose work involves the tasks AI handles well (data analysis, literature synthesis, pattern recognition) tend to be more positive, while those whose work involves the tasks AI handles poorly (fieldwork, interviews, archival interpretation) tend to be more skeptical.

But I noticed something beyond mere disagreement. Bluesky users who despise AI viscerally were often the first to deny basic, easily verifiable facts—for instance, that it can produce slide decks well. Very few respondents acknowledged that AI capabilities for research are real but worried about their consequences. People either dislike AI and deny any productive use, or like it and exaggerate its utility. Some of this is motivated reasoning—the explicit existential threat of a computer doing things better than you. But I suspect even more of it is simply ignorance about “the other.”

Contact theory is real. If you believe that Claude Code is evil or incompetent, I dare you to install it and use it to organize your research folders or create slide decks for your upcoming conferences. Earlier I encouraged folks to “spend a week with Claude Code.” It should have been “spend a day” (which should be enough).

13. User expertise still vastly determines output quality.

Perhaps not surprisingly, much of the criticism on Bluesky still assumes that using AI means copying and pasting from a chatbot. That is just not how agentic AI works. Agentic AI operates autonomously within your file system, reads and writes code, consults documentation, and executes multi-step research workflows—all guided by detailed instructions you build over time.

One related, common form of AI denial also assumes that because the tool is accessible, anyone could produce the same output. That is like arguing that because everyone has access to a stove, everyone can cook a good meal. There are obvious differences in cooking skills, recipes, and the quality of ingredients.

But the question is not whether AI is better than most professors at doing most important research tasks (I still stand by the assertion that it is), but whether good researchers with AI are better than good researchers without AI (they absolutely are). Honestly, I would take well-prompted AI slop over Bluesky slop (hundreds of anonymous users responding ai/dr whenever they see the feared AI keyword regardless of any substance) any day of the week.

14. Publication lag makes AI capability critiques obsolete by the time they come out.

Here is a problem that almost nobody in the debate acknowledges: academic and book publication timelines are structurally incompatible with AI’s rate of improvement. When someone cites a 2025 paper (initiated in 2024) documenting GPT-4’s hallucination rate to argue against using AI in March 2026, they are citing evidence about a system that no longer exists. It is like citing a 2005 study on flip phone limitations to argue against smartphones. That’s probably why the new “AI Con” book is so bad—clearly outdated before it even hit shelves.

I am not dismissing all of this research itself. The studies are often methodologically sound. But the evidence base expires faster than it can be published, reviewed, and cited. Messing and Tucker’s Brookings piece, published in March 2026 (and reviewed “rapidly” in only two weeks), already documents capabilities that would have seemed speculative six months earlier. By the time a peer-reviewed paper on current AI limitations appears in a journal, the limitations it documents will likely be fixed. This is not a comfortable situation for academics who are trained to rely on published evidence. But it is the situation we are in.

15. Most papers are already mostly read by AI, not humans.

It is an open secret in academia that most published papers are never cited or read by anyone beyond the authors, reviewers, and sometimes the editor. With the coming proliferation of AI-written papers—whether complete slop or not—it will become impossible for researchers to keep up even with their own niche fields. I like to think I am aware of all the new literature on immigration attitudes, but I am probably missing 80% of what gets produced outside the US, Europe, and top disciplinary journals.

This means that academics should accept their primary audience is increasingly LLMs. Tyler Cowen has been talking about writing for LLMs for some time, but with the ascent of agentic tools, this applies to most academics too—including qualitative researchers whose work itself cannot be automated. I do not have a firm sense of what authors should do about it, but ensuring that a machine-readable version of your paper exists (ideally in .md format) seems like a good first step.

16. AI exposes what was already broken in academia and beyond.

Related, a large number of responses to Part I amounted to: “If AI can do your research, your research was never good.” I agree (LOL)—but that is an indictment of much of social science, not a defense against AI or a smart attack against me personally. The replication crisis, citation padding, p-hacking, and the production of papers no one reads were all pre-existing conditions.

Human-generated academic slop was always pervasive; AI just makes it visible. Nathan Smith put this more bluntly in his restack: academic institutions hoard human capital, the tenure system rewards collective navel-gazing over public impact, and most professors could be more useful doing something else. That is a harsh framing. But if only a small percentage of published papers have genuine value, the system AI is disrupting was not exactly thriving.

17. Skill atrophy is a real risk, especially for the future generation of scholars.

This brings us to what I consider another strong reaction to my initial post: that outsourcing cognitive processes like “evaluating sources” and “coding data” damages the researcher’s own understanding. Many folks rightly worry about “reducing complex, thought-driven processes to a series of discrete tasks to be outsourced, when there’s so much that goes on cognitively both between and after the steps.” Messing and Tucker flag the same risk under “skill atrophy.”

I take this seriously, and I concede the risk is real—especially for students and trainees who have not yet internalized the cognitive skills that AI might short-circuit. The researchers who worry about skill atrophy are right that something is lost. But they underestimate what is gained: the ability to operate at a higher level of abstraction, to test more hypotheses, to iterate faster. For established researchers, the risk of atrophy is low because the skills already exist. For students and future researchers, we urgently need to figure something out in updating our grad school curriculum.

18. AI writing detectors and disclosure norms do not work.

AI writing detection tools were bad, are still bad, and will probably remain bad. The original Claude-produced post passed every major AI detector as “100% human” without any elaborate prompting to avoid this on my side. Many critics of my initial post said they immediately “sensed” it was AI-written. But they said this after I revealed the workflow—a textbook case of confirmation bias. Before the reveal, nobody flagged it. In fact, someone even complained I didn’t use AI to write a post boosting AI.3

The more important point is about disclosure incentives. Messing and Tucker recommend standardizing AI usage declarations across fields. I respect their reasoning and the call for standardization (instead of the chaos that we have now), but I disagree that any expansive AI declaration standard can have any merits given the current incentive structure.

Do not get me wrong—people in positions of authority like journal editors should be transparent about their workflow. But for regular authors, voluntary disclosure creates a system where honest users get punished and dishonest users face no consequences. I disclosed my AI workflow and received threats, professional attacks, and calls to fire me. The rational incentive is to lie. “AI usage acknowledgments” sound reasonable, but they collapse on contact with the actual social dynamics of academic life in 2026. Until the professional costs of disclosure drop, mandatory acknowledgment norms will select for dishonesty.

There is also a deeper problem: disclosure norms get the accountability question backwards. For some, AI disclosure can even function as a cop-out—”I used AI, so it’s on you now to figure out if it is slop.” But authors should stand by the final product regardless of how it was produced. If AI introduces an error, that is the author’s responsibility. What matters is whether the work is correct and valuable, not whether a human or a machine typed the sentences.

19. Academic Bluesky is not a serious venue for this debate.

I have to address this because it colored everything that followed. Bluesky generated almost as many reactions as Twitter, but they were overwhelmingly hostile in the least productive way possible. The most common response was some version of “If you didn’t write it, why should I read it?” or “ai/dr.” Many included curses, accusations of being paid by AI companies (?), and calls to not cite my earlier published work (??) or even to fire me (???) with people tagging my employer to replace me with AI since I’m claiming it’s so good.

My original post was provocative. But I did not attack anyone personally. I made arguments about AI and academia, based on my own experience in the field, which you may agree or disagree with. For that, academics on Bluesky responded with professional threats, ad hominem, and coordinated pile-ons. I have thick skin and employment security. I can absorb this.

But most people who might share heterodox views on AI in academia do not have that luxury. They are graduate students, contingent faculty, and junior researchers (in fact, I was one myself just a couple of months ago!) who watch what happened to me and draw the obvious conclusion: keep your mouth shut. That is the real cost of pile-on culture—not to people like me, but to the open exchange of ideas that academia is supposed to protect. And while I appreciated all the sympathetic folks who reached out in DMs, I wish you would speak out publicly. That is the only way this unfortunate dynamic can change.

20. Research can lack “soul” and still serve the public.

Max Kagan articulated and addressed a common concern from Bluesky folks that resonates with me too: the idea that research produced by or with AI lacks something essential—call it soul, craft, or authentic intellectual engagement. The process of struggling with a question, sitting with ambiguity, and slowly building an argument is personally transformative for many scholars. There is a reason people pursue PhDs despite terrible labor market prospects: the work itself is meaningful. When AI compresses that process into hours, something genuinely valuable is lost.

I feel the pull of this. But I am not sure it survives contact with the question of who pays for it. Most academic research is publicly funded. Taxpayers do not fund universities so that professors can self-actualize. They fund universities to produce knowledge that benefits society. If AI-assisted research produces more and better knowledge faster, the public interest argument for embracing it is hard to resist—even if the private experience of research becomes less romantic.

***

All in all, the discourse around Part I was messy. But it was also productive. It probably encouraged a few dozen if not more academics to try agentic AI tools for the first time, so I take it as a win. The strongest substantive objections—hallucination, skill atrophy, qualitative research gaps—forced me to think more carefully about both the risks and the opportunities. The Bluesky pile-on artists just showed yet again that it is not a serious platform for an open exchange of ideas. But the intensity of the reaction only confirms the stakes. People do not argue this fiercely about things that do not matter. Academics are waking up—some enthusiastically, some kicking and screaming. Either way, they are waking up.