Doctoral admissions at leading American research universities fell 15% this fall. The cause is not AI. It is the collapse of the departmental funding model. Federal grants are drying up. International students are looking elsewhere (we all know the story there). Universities are no longer sure they can support a student for the five to seven years a PhD takes. So this is not just about dumbing down students. Will they even have the opportunity to think?

Doctoral education has almost no revenue of its own. It is paid for by federal grants and cross-subsidized by undergraduate and Masters tuition: PhD students teach the undergraduates, Masters programs run at a margin, and the margin helps carry the doctoral program. When the grants collapse, the whole weight shifts onto tuition, and tuition was already stretched. Funding PhD students is a delicate balance among other revenues, the balance has always been fragile, and both sides are now failing at once. And a smaller pool is not just smaller. It changes who can afford to come, and what they choose to work on once they are here.

While the funding model was breaking, the work itself has changed. Frontier LLMs are now better reasoners than the median PhD student (and that is being generous in some fields). The middle of the research process becomes completely automatable. Not just the grunt work: the literature review, the coding, the analysis pipeline. The reasoning itself is commoditizing. Whatever remains scarce, it is not reasoning.

This inverts the economics of the grant proposal. A grant proposal is a promise to do work, written over months, funded maybe one time in five (very optimistically). Meanwhile, AI has collapsed the cost of doing the work itself, at least the computation and writing. When a study costs less than the time writing the proposal, the cost-benefit flips. It is more honest and efficient to do the work than to use AI to write the proposal to do it. It lightens the load on reviewers too.

So what is left, once the middle automates? Code runs or it doesn’t. A citation exists or it doesn’t. Errors get caught.

What remains is the two ends: choosing the question, and deciding what the result means. Both are subjective (and political) acts. When there is no checkable answer, reasoning does not do the deciding. Values do.¹ A model can out-reason the student and still be the wrong thing to lean on at the ends, because the ends were never reasoning problems. Worse, this is exactly where our own measurements find models least trustworthy. Ask questions with no right answer and models converge on the same defaults, and bend to how the question is asked.

So the question becomes how we train (and fund)² students to do the important parts of research now that the plumbing is out of the way. In short: identify the load-bearing decisions and own them.³ They are the two ends: the question and the interpretation.

On the question end there is real evidence for concern. People who ideate with AI produce more ideas that are more alike (Anderson; Doshi & Hauser). On the interpretation end, we don’t know much, other than that models can’t be trusted under ambiguity.

I am all in on using AI for research. This year I have run a research program with a model in the loop at nearly every stage. No grant paid for it. The studies cost tens of dollars each, and having literally no research money turned out to be a workable constraint. Any person with a scientific question can just find out, and publish what they find. The cost of curiosity has never been lower.

But being all in comes with a responsibility: to focus on the right kind of AI use in research, the part that doesn’t need critical thinking, so that we can focus on what does. Most of that conversation today is about the middle: disclosure, detection, what counts as cheating. I believe that the responsibility that matters has always been at the two ends.

---

Note about AI usage: This essay was written in the arrangement it argues for. I set the argument, the stance, and the order. Claude (Fable 5) proposed a skeleton; I cut it, moved it, added my own lines, and left instructions; it wrote the prose. I made the final pass and reviewed all the references. The machine did the middle. I kept the ends.

Footnotes

1. Haraway’s situated knowledges in STS, and Value Sensitive Design (Friedman & Hendry) in HCI. The values models express are measurable too: Anthropic’s Values in the Wild identified thousands in real conversations, and its follow-up found they differ by model and by language. ↩

2. The funding half deserves its own essay. It should also lead us to ask what counts as the “middle” in running a university, and how we could get rid of that (cough). ↩

3. See Eytan Adar’s AI-native PhD students. ↩

With the advent of AI-enabled “vibe coding”, I have been like the proverbial kid in a candy store. So many systems I have wanted to build over the years, are now possible at a fraction of the cost and effort!

One of these is a product called Fotomap. Building on Local Ground, our earlier open-source mapping platform, I wanted to build a dead simple mapping tool that allowed people to tell their own place-based stories using maps. There are alternatives, but they are clunky, expensive, or focused on data collection more than storytelling.

Fotomap is part storytelling platform, crowd-sourcing tool and qualitatively-minded GIS. It allows for the geo-referencing of all kinds of media, displaying it through maps and slideshows, and inviting people to participate by submitting their own media and stories. It’s open source and free to use.

I’ve been using this tool in my classes and research in a number of ways:

• Oral history - students in a 7th-grade social class shared stories of the “meaning of home”, by interviewing their parents about where they came from
• Photovoice - students in my Remaking the City workshop use Fotomap to share their experiences as new residents of New York
• Participatory mapping - students in a summer program in Oakland documented healthy food options (or the lack thereof) in their communities
• Storytelling - we created an interactive, multimedia cultural atlas for sharing stories — myths, poems, songs, and folk tales — from across India

Over the coming weeks I’ll share more about these experiences, and the features in Fotomap that made them possible.