TB: Some Thoughts on AI and Research (Andrews, MIT)

Précis

Isaiah Andrews, the MIT econometrician, wrote these notes originally for his PhD advisees and circulated them more widely. He frames the AI question in three cases — models surpass humans on all intellectual tasks, models improve substantially in some areas but stay weak in others, or model progress stalls. Under cases 2 and 3 the marginal returns to learning-to-use these tools are very high; case 1 makes human capital investment irrelevant so it shouldn't drive current decisions. The piece is a calibration document — calm, case-by-case, written by a tenured professor who is openly using AI more aggressively than his students are.

Key Takeaways

The three cases

  • Case 1: Models exceed humans on all intellectual tasks. Human capital investment seems irrelevant. Conditional on still wanting to do research, this case shouldn't drive current decisions because all investments have low return in this state. The non-Case-1 response some advocate: switch from human capital to maximum near-term earning.
  • Case 2: Models become much better at things current models are OK at, but stay below "good" human performance in other areas. Returns are high for skills that are both scarce and things models are bad at. Anything complementary to those scarce skills (e.g. next-generation model access) is also valuable. Andrews' best guess for "what they stay bad at": taste, judgment, problem selection.
  • Case 3: Model capabilities level off modestly above current level. Even with no further progress, current capabilities already shift the research production possibility frontier; "failing to figure out how to take full advantage of these tools would leave you at a major disadvantage." Andrews finds Case 3 implausible given the pace of progress.

Recommendations (valid across Cases 2 and 3)

  • You should be actively thinking about and planning around this. A PhD is an expensive long-term human capital investment; a change in research production possibilities should change what you invest in.
  • You should be learning to use these tools effectively. "Hiring, publication, and tenure standards are set in equilibrium. If the production possibility frontier expands and you do not keep up, you will lose out in an absolute sense." "No one uses these models for research" is not an equilibrium.

Three dimensions of "use the tools effectively"

  • Experimentation. Returns are high; nobody has "solved" how to use them. Andrews observes he is using these tools "more aggressively than many PhD students are" — and given he is "(a) a tenured professor and (b) kinda old," this suggests students are under-investing. Better models (institutional/individual subscriptions) are much better than free versions.
  • Verification. Output looks plausible but can be wrong in ways that require domain expertise to detect. Auditing model output is a core skill. If Case 2 holds, returns to taste and problem framing go up not down.
  • Division of labor. AI changes the production function not just for solo work but for collaboration — substituting for some RA tasks, changing optimal effort allocation across coauthors, and making feasible projects that weren't before.

Concrete career advice

  • "From year 3 on you should be attending at least 1 applied lunch + 1 applied seminar every week" — in addition to econometrics seminars. Empirical research is changing rapidly in both what gets done and how, and clear communication with applied researchers will be high-return under Cases 2 and 3.

Telling concrete detail

  • "I've found that GPT 5.4 Pro is substantially better at convex analysis proofs than I am" — from a senior econometrician. The capability claim is specific, not vague.

Notable Quotes

  • "Hiring, publication, and tenure standards are set in equilibrium. If the production possibility frontier expands and you do not keep up, you will lose out in an absolute sense (i.e. having worse outcomes than had the new technology not been available)."
  • "No one uses these models for research" is not an equilibrium."
  • "Learning when and how to audit model output is a core skill for working with these tools."
  • "I've found that GPT 5.4 Pro is substantially better at convex analysis proofs than I am."

Why this matters / Connections

The note is the cleanest case-based framework I have seen from an inside-academia perspective for how a PhD student should adjust skill investment given AI. It is also one of the most credible specific capability claims (a tenured econometrician on convex-analysis proofs) — citeable in arguments about whether frontier models are "actually" useful for research-grade math. The piece converges with Tyler Cowen, Andy Hall, Alexander Kustov, and Jason Fletcher on the practical conclusion (use them, daily, verify them, redesign workflows) but arrives there from a more cautious, case-analytic stance — which is rhetorically more transportable into a faculty meeting.

See also

Source

Some Thoughts on AI and Research — Isaiah Andrews, MIT Economics, 2026-04-03

Backlinks