AI is changing what my job is

Provocation · Damian Betebenner, Center for Assessment · AIME-Con 2026

One sentence

AI is not merely helping me do my old job faster. It is changing what my job is.

The frontier moved

One claim

Problems that once required a team of specialists—or remained unexplored because they were too expensive to pursue—can now be attacked by a single researcher working with AI.

What changed in my own work

In my workflow

Before AI With AI
Theory derivation is slow and locally constrained derivation and counterexample generation iterate continuously
Code implementation is a separate bottleneck code evolves alongside the mathematics
Simulation designs limited by time and implementation cost thousands of conditions can be explored and revised quickly
Critique review often arrives late adversarial checking can occur throughout the workflow
Documentation frequently downstream of the analysis documentation can evolve with the work
My role produce most of the technical artifacts specify, interrogate, verify, decide, and take responsibility

One concrete example

One example

A recent methodological problem in educational growth required:

  1. deriving properties of a new dependence-based model;
  2. translating the mathematics into working software;
  3. designing simulation studies;
  4. exploring thousands of conditions;
  5. diagnosing failures and constructing counterexamples; and
  6. repeatedly moving back from computational results to revised mathematics.

Previously: months or years, or a team with complementary specialties.
With AI: a tightly coupled research loop operating on the scale of weeks.

What became scarce?

Cheaper

  • mathematical derivation
  • code production
  • refactoring
  • simulation implementation
  • documentation
  • first-pass critique

More valuable

  • problem formulation
  • model specification
  • adversarial testing
  • recognizing nonsense
  • deciding what evidence is enough
  • accepting responsibility

The unresolved problem

One unresolved problem

The question is no longer whether I can do the analysis. It is whether I can warrant the analysis that AI has made possible.

If AI makes it possible to produce work whose complexity exceeds what its author—or a conventional peer reviewer—can independently reconstruct, what makes the result warranted?


As production becomes cheaper, does verification become the new scarce resource—and who is accountable for deciding that the evidence is sufficient?

Backup slides

Not part of the five minutes. For the dialogue, if useful.

A broader possibility

AI is collapsing the practical boundaries between mathematical theorist, statistician, programmer, simulation researcher, technical writer, and software developer.


The important change is not that every step is faster. It is that the entire intellectual loop can remain continuous inside a human–AI working unit.

Provocation for the field

If this is the emerging division of labor, are we training psychometricians for the work they will actually do?


What must a measurement professional be able to do personally in order to supervise, verify, and take responsibility for work that AI can execute faster (and, for some bounded tasks, better) than they can execute unaided?