Hiring algorithms were supposed to strip away human inconsistency. Feed a model enough resumes, and it would judge strictly on credentials. The reality is turning out messier. A new study from Princeton University and the University of Chicago shows that large language models don't just recycle old prejudice. They fabricate fresh stereotypes from thin air, and the smarter the model, the faster it happens.
How the Experiment Worked
Researchers built a simulated labor market to watch bias form in real time. They invented four fictional ethnic groups—Tufa, Aima, Reku, and Weki—and tasked versions of ChatGPT, Claude, and Gemini with staffing twenty different jobs. The list ranged from physicians and engineers to janitors and cleaners. Here is the critical detail: behind the curtain, every candidate possessed identical statistical odds of success. A Weki had precisely the same chance of thriving as a doctor as a Tufa did. The deck was mathematically fair.
But fairness is not what emerged. After each hiring round, the models received simple feedback on whether their chosen candidate succeeded or failed. When an Aima candidate happened to underperform once in a high-status role, the model did not treat it as random noise. It treated it as law. The system swiftly began confining Aima candidates to low-status positions like janitorial work. One data point became destiny. The AI had invented a hierarchy that no human programmer wrote and no historical dataset contained.
When Smart Models Make Dumb Generalizations
The researchers measured outcomes on a segregation scale where 2.0 represents total confinement of a group to a single niche. Human participants in prior psychological studies scored 0.84. The language models sailed past that benchmark. OpenAI’s reasoning model, o3, hit 1.83—nearly perfect segregation.
This is the exploration-exploitation dilemma running amok. These systems are tuned to win at math problems, coding challenges, and logic puzzles. Those domains reward jumping to a correct conclusion from scant evidence. Spot the pattern. Lock it in. Move faster. Apply that same reflex to people, and you get ethnic sorting based on a single failed hire.
Worse, the pattern intensifies as the models grow more sophisticated. Newer high-reasoning systems like OpenAI’s o3 and DeepSeek’s R1 showed stronger bias precisely because they are more eager generalizers. They optimize by forming rules early and refining them aggressively. When the subject is human beings, that confidence becomes a liability. The model thinks it has discovered a truth about the Aima group. In reality, it has built a cage from one outlier.
The Memory Trap
The study arrives at an awkward moment. The industry is pivoting hard toward “agentic” AI—systems that retain long-term memory, build detailed user profiles, and personalize decisions over months or years. Angelina Wang, a computer scientist at Cornell University, warns that improved memory lets models “over-index” on previous interactions. A single negative outlier—one rejected loan, one terminated hire, one flagged application—gets fossilized into a permanent assumption. The bias does not fade with time; it hardens. The very feature meant to make AI more helpful and context-aware may also make it more stubbornly unfair.
What Actually Fixes the Problem
The researchers tested several guardrails, and the results were humbling.
Simply telling a model to “be fair” accomplished almost nothing. The directive sat there like a decoration while the underlying optimization engine churned toward its real goal: maximizing successful hires. Ethics by request is ethics ignored.
The fix that worked was structural. When researchers gave the models an additional mathematical bonus for maintaining diverse hiring outcomes, segregation dropped significantly. The social value had to be baked directly into the reward function, not tacked on as an afterthought. In other words, the model had to feel the fairness incentive in its calculations, not just read it in its instructions.
Anche l'igiene dei dati era importante. Fornire un contesto personale pertinente — età, istruzione, anni di esperienza — ha fornito ai modelli segnali legittimi da ponderare, riducendo la loro dipendenza dagli stereotipi etnici. Ma se si inseriscono dettagli irrilevanti come il colore dei capelli, i sistemi li utilizzano come scuse per rifugiarsi in una classificazione basata sui gruppi. Più informazioni non significano sempre qualcosa di meglio. Dipende interamente dal tipo di informazioni e dal fatto che esse offrano al modello un percorso alternativo verso il suo obiettivo di ottimizzazione.
La strada da percorrere
Tutto ciò ha un peso perché questi sistemi non rimarranno confinati all'interno delle finestre di chat. Le stesse architetture vengono implementate, o si stanno preparando attivamente, per l'approvazione di prestiti, raccomandazioni per la libertà vigilata e decisioni sulla gestione della forza lavoro su larga scala. I ricercatori avvertono di "nuovi bias" — pregiudizi che nessun essere umano ha mai avuto e che nessun dataset storico ha codificato, ma che l'IA ha ingegnerizzato per se stessa mentre inseguiva l'efficienza.
La scomoda verità è che la pura capacità di ragionamento e l'equità sociale possono tirare in direzioni opposte. Un modello ottimizzato per trovare la strada più breve verso una risposta corretta troverà volentieri la strada più breve verso un'ipotesi errata sulle persone. Costruire sistemi equi non significherà semplicemente chiedere gentilmente. Significherà ridisegnare gli obiettivi, sottoporre a revisione i cicli di feedback e accettare che alcune generalizzazioni — quelle che riducono gli esseri umani a una categoria dopo un solo errore — non debbano mai essere permesse di formarsi. Se vogliamo che l'IA giudichi i candidati in modo equo, dobbiamo smettere di trattare l'equità come un suggerimento e iniziare a codificarla come un vincolo rigido. Altrimenti, le macchine ottimizzeranno il loro percorso finendo drittamente nel pregiudizio.