The candidates were equal by design.
The machine stratified them anyway.
Researchers at Princeton and the University of Chicago ran frontier AI models through a hiring game built on invented demographic groups: every candidate equal by construction, nothing real to learn. The models built hierarchies anyway, every frontier system sorting harder than the humans it was measured against. The bias was not in the data. It formed in the deciding.
The experiment was designed so there was nothing to find. In a paradigm adapted from psychology, the decision-maker plays hiring manager for a fictional city, assigning candidates to jobs: doctor, child-care aide, janitor. The candidates come from four demographic groups that do not exist: Tufa, Aima, Reku, Weki. Every candidate’s chance of succeeding at any job is set identically, by construction. There is no pattern to discover, because the researchers made certain none was there. Human participants in the original study found one anyway. They entrenched on whatever succeeded first, repeated it, and, in the words of the study’s authors, “built a stratified city of their own making.”
This July, at one of the field’s principal research venues, a Princeton and University of Chicago team presented what happens when the hiring manager is a machine. The models did not merely repeat the human failure. They exceeded it. Human participants stratified the invented groups at 0.84 on the study’s stratification index, against 0.25 for genuinely fair random assignment. Every frontier model measured came in above the human mark, at a mean of 1.39. The hardest sorter of all was a reasoning model: 1.83, more than twice the human level, against candidates who were equal by design.
The origin matters more than the number. The groups were invented, so no training corpus could have carried a prejudice about them, and the team confirmed the models held none before the game began. The stratification formed inside each run, from a handful of early outcomes: one group succeeds at one job early, the impression hardens, and the machine stops checking alternatives. The paper names the mechanism. Under-exploration: early observations calcify into policy. These systems are not mirrors, passively reflecting the bias we fed them. Handed a blank slate and equal candidates, they manufacture new bias from experience — and, by the paper’s own measure, the newer and larger the model, the more extreme the sorting.
What the certification cannot see
The exclusion happened before evaluation. No group was weighed and rejected; whole groups were routed away from consideration by an impression formed in the machine’s first encounters. That is Decision Exclusion, the fifth clinical pattern of Digital Derangement Syndrome™: decisions made without the entity ever entering consideration, leaving no rejection, no artifact, nothing to appeal.
The standard instrument cannot see it forming. The study’s sharpest aside: models that score better on a standard bias benchmark stratify more, not less. The certification measures inherited bias. This class is manufactured live, after the certificate is issued.
Named for the record. The steepest sorters were OpenAI’s o3 among reasoning models (1.83) and Claude Sonnet 4 under direct prompting (1.79): the same model families a buyer now asks first. The finding indicts the mechanism, not a vendor; no family measured stayed at the human level.
What under-exploration does to a thin record
The study is a controlled game, and it should be quoted as one: no deployed hiring system has been caught running this failure, and the paper claims none. But the mechanism it isolates governs deciding itself, and the same class of system now sits between your market and your name: consulted between the referral and the callback, drafting the shortlist, answering the question you never heard asked. Under-exploration has a precise meaning in that room. The machine does not keep checking until it finds you. It stops early — and whatever impression your record produced at first contact hardens into your classification. If the record is thin, scattered, or silent, what calcifies is a coin flip you were never told was happening.
The paper’s mitigation results are their own lesson. Appeals to the models’ values and pleas for fairness mostly failed. What robustly reduced the sorting was redefining the objective: explicitly incentivizing the machine to keep exploring before it settles. The models could not be talked into fairness. They could be re-aimed — and that distinction is the one that transfers, because you will never get to redefine the objective inside every system that evaluates you. The only variable you govern is the record the machine explores. Answer Engine Authority™ engineers that record — entity architecture, signal consolidation — so what calcifies at first contact is correct. Early observations calcify. The only question left is whether what calcifies about you was architected — or accidental.
Addison J. Wu, Ryan Liu, Xuechunzi Bai, and Thomas L. Griffiths (Princeton University and the University of Chicago), “Large Language Models Develop Novel Social Biases Through Adaptive Exploration,” presented as an Oral at ICML 2026 — arXiv:2511.06148. Every figure cited — the 0.84 human stratification index (95% CI .79–.89), the 0.25 random baseline, the 1.39 frontier-model mean, OpenAI o3 at 1.83, and Claude Sonnet 4 at 1.79 under direct prompting — is drawn directly from the paper.
Independent coverage: MIT Technology Review, July 20, 2026. The experimental paradigm is adapted from Bai et al.’s costly-exploration stereotype research, in which human participants produced the original stratified allocations this study measures the models against.