Goldilocks and the Three AIs: What We Owe Graduate Students About AI

 Our graduate students are training for professional and academic careers, but our curriculum was built for a world before AI.

This isn’t an indictment; curricula move slowly by design. They encode what a field has agreed on, and agreement takes time. AI adoption is uneven, along what Ethan Mollick calls the jagged edge: strong in some tasks, unreliable in others, and on an improvement trajectory that makes next year (nearly) unforeseeable. A program that rewrote itself around last year’s capabilities would be revising again this year. Caution is a reasonable response.

The cases for using AI in academic work are tricky. Authorship is contested; Data protection rules were written before anyone pasted an interview transcript into a chat window. Faculty attitudes range from enthusiasm to anxiety, fear, and even hatred, and behind these emotions there are real arguments. I don’t think anyone is wrong to be uneasy.

I do think we are getting the cost of waiting for clarity wrong.

Our graduates are entering fields where AI literacy is already assumed. They will submit (and referee) to journals with AI disclosure policies. They will sit on IRB panels. They will be asked by a district or a dean or a search committee what their position is, and “I wasn’t taught that” is not a position. Sending them out without practical skill and a defensible ethical stance is not neutrality. It is a decision, and the cost lands on them.

On our podcast recently, a graduate student walked through how he arrived at his own working relationship with these tools. Not refusal, not delegation. He described testing where the tool helped his thinking and where it replaced thinking he needed to do himself, then settling into a level he could defend. It is about finding the Goldilocks level.

That process is what I want to teach. He got there on his own, which is fine for some but useless as a program design. Some students will figure it out. Others will land on one extreme or the other and stay there, and neither extreme serves them. Guiding students through that calibration is what post-secondary education is for. We do it with statistics, with critical theory, with writing. There is no reason to treat AI differently.

So rather than complaining in my writing (an exercise I enjoy), I am adding something alongside it. Starting this term, I am offering a free weekly work session for graduate students. Not a lecture series. Work time, with structure. Students bring their own projects, and each session anchors on a cluster of questions with substantial time to try things and see what breaks.

My initial plan covers nine clusters:

1. Authorship and agency. COPE standards, journal policies, what disclosure actually requires, and where the field has not settled.

2. Privacy and data protection. What you can put in a model, what you cannot, and how IRB commitments constrain your workflow.

3. AI in the literature review process. Where it accelerates search and synthesis, and where it steers you wrong.

4. Building a researcher identity with your AI. Setting up context so the tool works from your frameworks and your voice rather than a generic average.

5. Qualitative analysis. Promises and pitfalls, coding assistance, and the risk of losing contact with your data.

6. Quantitative analysis. Code generation, assumption checking, and verification you still have to do yourself.

7. Designing research instruments. Item generation, revision, and validity questions that do not go away but may look different.

8. AI as a teaching assistant. For their teaching, and for their own learning.

9. Friction and learning. Information processing theory, and why removing effort from a task sometimes removes the learning with it.

That last cluster matters most to me, and I plan to keep returning to it. Ease is not the goal. Some cognitive work has to happen in the student’s head, or it does not happen at all, and knowing which work that is separates skilled use from outsourcing.

I expect the list to change once students tell me what they actually need. That is part of the point. The clusters are a starting structure.

The argument is simple. Our students will practice in a field that has already changed. We can hand them the skill and judgment to work in it, or we can leave them to assemble it alone, after they graduate, with the stakes considerably higher. Those are the two options. There is no version where the question goes away.

P.S. Sarah often says that she hates this timeline. I often agree but until we can join the Time Variance Authority- this is pretty much it.

AI is not Social Media: But somtimes it is

AI is not social media. I keep sensing that much of the angst comes from the comparison, and I worry it may lead us to the wrong lessons and wrong actions. The anger at the arrogance of the tech bros is obvious and justified, and I share some of it. But the technology underneath works differently in ways that educators and Ed policymakers need to understand.

Social media has a main use and a main abuse. It captures attention and does not let go. Facebook decides what shows up in my feed. When I open ChatGPT, Claude, or Gemini, I bring the question, and I pick the task. That structure leaves agency with the user, and agency is at the center of what we try to build in students. We can take charge (though it has its agency traps) of AI, and we can teach our students to do it. We spent two decades watching attention platforms pull kids away from reading, sleep, and each other. A tool that helps a teacher differentiate a text, or helps a student get unstuck on a draft, runs on a different logic. A tool that helps students make, inquire, and design has potential.

Made with ChatGPT

The cost of running these models has also kept companies from chasing engagement the way social platforms did, at least so far. But the danger is still lurking in developing relationships with bots. Much like social media replacing human connections with synthetic ones at critical moments in development can have very harmful effects on kids. But unlike social media, AI offers a complete set of affordances unrelated to synthetic relationships. Moreover, you can realize the benefits without a unique login.

I am not saying the tech bros turned good. I am saying the technology is different, and our response in education should reflect that. Schools that treat AI as another attention trap will ban it and give up the chance to teach students to use it well. At the same time, the real lessons from social media, about business models and about protecting kids, still apply, and we should carry them forward. Nuance takes more effort. In classrooms, we cannot afford to skip it.

My proposition for education is to avoid a blanket ban on screens (though no phones in school sounds reasonable to me). Teach students about their own cognition and about AI. Then use AI to augment and enrich their education while emphasizing agency.

What I Actually Want From AI in Peer Review

 It’s summer, and I finally have a bit more breathing room, some of which I use for peer review (the rest is for my granddaughter). The pressure never really lets up, though. I’ve written before about how AI is reshaping that pressure, on both the writing and reviewing sides of the equation.

Publishers are drawing a clear line. They’re asking reviewers not to run submitted manuscripts through AI tools, and they’re right to ask. A paper under review is unpublished work. Feeding it into a commercial AI system risks the author’s intellectual property in ways reviewers might not stop to consider.

I follow that line carefully. I use AI to help me write up my reviews, sharpen the language, organize my thoughts, and make the feedback easier for authors to act on. I never run a manuscript I’m reviewing through AI, though often I think it could help.

I know it could be helpful because with my own writing I do something different. When I finish a draft, I ask AI to act as a peer reviewer and pick apart my argument. Where is it weak? What can I cut when I’m over a word limit? AI has been remarkably good at both. It catches gaps I’ve stopped seeing because I’ve read my own draft too many times. I get back to work the same way I would after an actual peer review, provided I agree with the notes, which, much like human reviewers, I do not have to agree with.

I do this through institutional AI that doesn’t train on what I feed it. I’ll admit, honestly, that the training question isn’t the part that keeps me up at night, even if I did not have such access; the benefits definitely outweigh the risks. What strikes me instead is how good the feedback is and how much better the review process could be if this kind of access extended to reviewers too.

I’m not calling for AI-written reviews- in that case humans are not needed. I want to be clear that we (still?) need human reviewers. What I’d love to see is publishers giving reviewers access to no-training AI as a tool.

Peer Review made with ChatGPT 5.5

Picture a reviewer working through a manuscript that cites a document by name. Right now, chasing that down means opening another tab, another search, another interruption. AI access built into the review platform could pull it up on the spot. Or picture a paper built on a statistical test with a dozen parameters. A reviewer could ask what those parameters mean and whether they were applied correctly, instead of taking the authors’ word for it or quietly moving past a section they don’t fully trust but have no time to verify.

I recently reviewed a paper where the data looked suspect. It was disappointing, and in that case, the signs were obvious enough to catch without much effort. But it left me wondering how many more subtle cases slip through, the ones that look clean on the surface and only fall apart under closer inspection. That’s exactly the kind of pattern-checking AI does well, if reviewers had a legitimate, secure way to run it.

The bigger opportunity is replication. Right now, a reviewer can, in principle, rerun an analysis or test an alternative approach if the authors shared their data. In practice, very few do it, because it takes far more than any reviewer has the time and resources for. AI with real document and data access could shrink that gap. Reviewers could actually check the math (within approximation) instead of trusting that it holds. This may be critically important when people are using AI to generate research facsimiles, sometimes fabricating the data and the writing.

None of this changes what a peer reviewer is for. The job is still to weigh the argument, judge whether the evidence supports the claims, and tell the editor whether the work deserves to move forward. AI wouldn’t replace that judgment. It would just let reviewers exercise it with more precision, on more of the paper, in less time.

Faster reviews. Sharper reviews. Reviewers who can actually verify instead of just trusting. That seems like a use of AI worth building, and one publishers are well positioned to build safely, since they control the data and access.

The Tinkering Ways AI is Transforming Research

 I recently sat on a grant panel. As on many panels, I was asked not to discuss the details of the decisions, a request I always respect. The proposals spanned a wide variety of fields, so there was plenty to learn and chew on. What struck me most was the robust use of AI.

Well, you’d say, we live in the age of AI. Everybody’s got ChatGPT (or pick your favorite model), and people mumble something about how it can be used because generative AI is sexy and new. The cynic would point out that we’re saturated with AI, both as an economic and a cultural phenomenon, and that if you want funding, you have to at least nod toward some magical use of AI that will completely transform your field. The cynic would add that in most cases, the AI is described vaguely and hand-wavily. And they wouldn’t be entirely wrong.

But in many of these applications, I saw something else entirely. Most of them actually included well-thought-out machine learning, not generative AI. They used small models, built from scratch for a specific purpose, as opposed to large language models. In a way, these proposals were “unsexy”: they didn’t reach for the new models at all. They offered innovative solutions using the old ones.

Created with ChatGPT 5.5

The insight took a few days to land. AI was everywhere, but not in the cynical way I’d expected. Here’s how I read it: the cultural significance of generative AI, and the sheer accessibility of AI, has gotten researchers in every domain thinking, “Hey, maybe this AI thing can help me solve this really hard problem.” (Sorry, I can’t reveal much from the panel, so no details.) It’s fairly clear to me that generative AI has started a conversation inside research that will produce exciting new solutions to some of our most intractable problems.

The next challenge, oddly, will be having enough compute, coders, and AI experts to bring these projects to fruition. I saw this at one of our research gatherings earlier this year: everyone is looking for an AI research partner, while our AI scientists are flooded with collaboration requests. I think we could triple the number of people doing that work and still have demand to spare.

So where does that leave us? I believe research universities should invest in compute and AI know-how as a shared service that can support many research efforts at once. Some have already gone this direction. I hope more, including UNL, will follow.

Incoherent Thoughts from the Jagged Edge

[originally posted on my substack] 

I started this week sitting across from a long-time colleague, someone whose work I respect and whose judgment I trust. Let’s call him Eric. Eric is co-authoring something with a younger scholar, and he has been finding hallucinated references in the manuscript. He has, I think, tacitly accepted that AI was used in the writing. The references are the tell. He has not said so explicitly, but the evidence is sitting right there in the reference list.

What struck me about the conversation was not the hallucinations. Those are a real problem, and anyone using AI-assisted writing needs to know that. What struck me was that Eric is not using AI himself, which means his mental model of the technology is frozen at whatever he last heard about it. He does not know yet that frontier models (the current generation) are substantially less prone to fabricating citations than earlier versions. His experience of the technology is secondhand and dated, and that gap between perception and current reality is itself a kind of problem. The tool is a moving target, and the critique needs to move with it.

Later in the week, I had a completely different kind of conversation. Dave Fowler, a math education professor who retired a few years ago, shared with me the exchanges he has been having with ChatGPT about the nature of theories. Dave is curious, playful even about what this thing is and what it can do.

He shared some of the thinking he was doing with and about AI. We smirked together at the turns of phrase the model deployed in those conversations, including things like “You’re drawing a very precise and interesting analogy, Dave.” which my brain heard in the voice of HAL 9000. There is something both charming and slightly unnerving about that sentence. It flatters, it redirects, it sounds like a thoughtful interlocutor. Is it? Is it not? Dave was not sure, and neither was I, and that wondering felt like the right place to be.

I am still turning that conversation over in my head. What Dave is modeling is something I think we are not talking about enough in education: what it looks like to approach AI with curiosity rather than with a verdict already in hand. He retired, he has no institutional pressure either way, and for $20 a month he is just... exploring. There is something intellectually honest about that.

Uncertainty Saloon (Created with ChatGPT)

Meanwhile, the news cycle this week offered its own jagged edges.

A winner of the 2026 Commonwealth Short Story Prize has been accused of submitting an AI-generated story, and the situation quickly became a full-blown literary scandal. The accusations were amplified when Pangram Labs ran every Commonwealth Prize winner back to 2012 through their detection tool and found strong signals of AI-generated text in multiple entries, including the 2025 overall winner. The literary world, predictably, is having a moment.

And then Barnes & Noble CEO James Daunt appeared on the Today show and said he had no problem stocking AI-written books, as long as they were clearly labeled and not misrepresenting themselves. Cue the boycott calls. Cue the “all generative AI is ripping off someone else” counter-arguments on social media. Daunt has since clarified, repeatedly, that Barnes & Noble does not knowingly sell AI-generated books and takes active measures to exclude them. The clarification landed with about as much impact as you would expect, which is to say, not much. The outrage had already found its shape.

These two stories (the prize and the bookstore) are related. Both are really about thresholds. At what point does AI-assisted become AI-generated? Who gets to decide? What is the meaningful distinction between a human writer who uses AI as a tool and one who uses it as a ghostwriter? I do not have clean answers. What I am sure of is that any calls for no AI will just feed into shadow AI use.

On a more personal note: I have been noticing that when I write text on my own, without AI assistance, Grammarly flags it as AI-generated. (I still have a Grammarly subscription, though the controversy around the company has me reconsidering that). The flag itself made me stop and think. Am I absorbing AI patterns from all this use? Or is this a simpler and more uncomfortable explanation: that AI was trained on enormous quantities of mediocre writing (including mine), and is now reflecting the patterns of mediocre writing back at us, and Grammarly is simply recognizing them? I am a mediocre writer, so am I no better than AI?

I find this genuinely interesting and only slightly humbling.

Social media, that well-known venue for nuanced and measured discourse, has been full this week of a particular kind of certainty. The message, in various forms, is this: In five years you will look back and realize that AI in schools was a terrible mistake.

I do not know how to evaluate that claim, because I do not think anyone can know it. We are, as a society on the verge of something, but the outcomes are unclear. The people making this prediction with the most confidence tend to be the ones least encumbered by curiosity or evidence. As Ted Lasso misquoted “Be curious, not judgmental.” I think of Dave Fowler, sitting with his ChatGPT transcripts, genuinely wondering, and then I think of the social media certainty crowd, the UnDaves. The UnDave has an opinion. The UnDave does not let curiosity or facts complicate the opinion. The UnDave is very confident about what five years from now will look like.

I am not an UnDave. I hope I never become one. Here is where I am, for this week at least.

I love the exploration. I love what AI does for the scope and scale of what I can get done. And more than either of those things, I love that GenAI keeps opening doors to new questions, questions I would not have thought to ask, problems I would not have noticed, conversations like the one with Dave that I will be thinking about for weeks.

On schools specifically: I have said this before, and I will say it again. Caution is warranted. We do not have enough evidence yet about long-term impacts on learning, on writing development, and on productive struggle that builds capacity. Those concerns are legitimate and deserve to be taken seriously.

But on the other side of that, we must teach students about AI, what it is, how it works, and how it is changing the world they are growing up in. Not doing so is irresponsible. These students are going to live in a world shaped by this technology, whether we prepare them for it or not. Much like sex ed, if we don’t teach about it in schools, they will learn it elsewhere, with potential negative consequences.

I understand the desire to put the djinni back in the bottle. But we learned from the nuclear age that you cannot undo technologies. We must work as a society to reckon with the age of AI, and part of that is education.

Eric’s colleague used AI and did not know how to use it well. Dave used AI and understood enough to find the conversation genuinely interesting. The difference is not access. The difference is curiosity and intellectual engagement.

That is what education is (or at least should be).