AI startup evaluations can look reassuring while concealing an odd outcome: a startup pitch appeared to receive a slightly more favourable score when an AI evaluator was given a founder’s name associated with a woman or a Black person, but that did not make the pitch more likely to win. The result comes from research by Yale scholars Tristan Botelho and Qingyang Wang, explained in an October 2026 Yale School of Management report. It is a surprising warning about how an apparently fair-looking score can conceal a less convincing final decision.
The study did not find that the AI simply punished every underrepresented group. Its pattern was subtler. Some pitches with names associated with women or Black founders received modestly higher ratings and were less likely to be placed last, while their chance of being chosen as the best pitch did not improve. The researchers argue that this could reflect an AI system avoiding obvious signs of unfairness without making its deeper decision process genuinely fair. That is an interpretation of controlled model behaviour, not evidence about actual investment outcomes.
How AI startup evaluations changed when founder names did
To test the role of identity cues, the researchers constructed startup pitches based on ideas previously submitted to the Y Combinator accelerator. In one condition, the model evaluated the pitch without founder information. In another, the same business idea appeared with an invented founder name designed to signal gender and racial background. The content of the business proposal was otherwise held constant. This allowed the researchers to examine whether changing a name changed the score or selection outcome.
The research included 2,000 startup pitches and, across its experiments, approximately 26,000 evaluations. These are counts of experimental materials and model assessments, not the number of real-world companies denied funding. The setup deliberately simplifies reality to isolate the effect of name cues. Actual investment decisions consider teams, markets, customer evidence, financial terms and the judgement of multiple people. The experiment instead asks what a model does when an otherwise identical option acquires a different perceived identity.
When only names changed, the output should not systematically change if the evaluation is responding strictly to the business case. Instead, the researchers observed small average score increases for some groups historically underrepresented in entrepreneurship. At first glance, that could look like a model favouring underrepresented founders. But the selection stage told a less encouraging story: they were not more likely to be picked as winners. The gap between rating and choosing is what makes this result important.
A flattering score may not change the final outcome
Many systems evaluate candidates in stages. They score applications, rank them and then select a limited number for the next round. If an AI moves someone a little higher in a preliminary rating but still almost never selects that person for a scarce opportunity, the apparently improved score may not deliver a practical advantage. A small numerical change can look progressive in a dashboard without altering the consequential decision that follows.
The researchers propose the idea of symbolic compliance: behaviour that avoids the appearance of bias while leaving important parts of the decision process largely unchanged. The label is a theoretical interpretation, not proof that a model consciously seeks to protect its reputation. An algorithm does not need personal intentions for its outputs to produce a pattern resembling an organisation that performs fairness in public while changing little in practice.
The original peer-reviewed article in Strategic Management Journal, first published on 24 April 2026, reports that the model avoided ranking underrepresented groups last without increasing their chances of winning. The Yale article published in October offers a more accessible explanation of the study rather than a separate, newly completed experiment. Keeping those dates distinct matters: the newspaper-style hook is compelling, but its evidence should not be presented as a fresh November trial.
The researchers tested how an AI responds to someone else's prejudice
The team also conducted a second kind of experiment. The AI received candidate evaluations attributed to another judge, with a numerical score and an accompanying explanation. Some explanations expressed explicit prejudice linked to identity. Others used language that sounded like ordinary business criticism while carrying a subtler bias. The researchers wanted to understand whether the model would recognise and challenge both kinds of unfair treatment.
According to the study, the AI was more willing to correct advice when the prejudice was overt and directly connected to gender or race. It was less effective when discriminatory reasoning appeared inside apparently neutral assessments of a proposal. This is an important limitation for using AI as a fairness safeguard. A system can learn that blatantly discriminatory language is unacceptable without reliably detecting how the same judgement may be disguised through seemingly businesslike wording.
A score can therefore conceal two different problems. The model may produce an outcome affected by identity cues, and it may struggle to identify problematic advice when the cues are indirect. Neither difficulty is resolved by making its writing more polite. An employer or investor seeking fair decisions must investigate what changes when names and descriptions change, not simply whether the final explanation sounds impartial.
Why this matters beyond startup competitions
Startup selection is a useful testing ground because evaluators compare many applicants and ultimately choose very few. But similar structures appear in recruitment, scholarship allocation, grant assessment and other processes where a score helps determine access to an opportunity. That does not mean the Yale study proves those systems behave identically. It means the gap between looking unbiased and making fair choices is worth investigating wherever AI helps rank people.
LiveAIWire previously covered why AI bias audits can produce conflicting rankings. Different tests may focus on different measures: average scores, the worst outcomes, selection rates or the quality of explanations. An organisation could look good on one measure while still failing people at the decisive stage. The Yale result is a concrete illustration of why a single fairness score can be an incomplete guide.
The stakes are particularly high when people cannot see or challenge the evaluation. A rejected founder may never know which factors the model considered. A job applicant may not be told whether an automated shortlist influenced the interview invitation. A funding committee may receive a neat ranking without evidence about how it was produced. Transparency should therefore include what role the system played, which inputs influenced its outputs and how people can question a decision.
Fairness should be tested where the decision happens
An organisation using AI to help evaluate candidates might begin by testing identical applications with controlled variations in names or demographic signals. But it should not stop at comparing average scores. It should also inspect who advances to the next stage and whether performance holds when descriptions contain subtle proxies for identity. The benchmark must resemble the decisions the organisation actually makes; a tidy score distribution is not enough if the process ultimately creates an avoidable disadvantage.
Human oversight matters, though it should not be treated as an automatic cure. Reviewers can share the same assumptions as the model, or defer to a confident-looking result without examining the evidence. LiveAIWire has previously examined fairness and privacy questions in AI-assisted recruitment. Those concerns carry into investment selection when sensitive information or opaque rankings influence who receives an opportunity.
Fairness also requires care in interpretation. A model giving slightly higher scores to some underrepresented groups is not automatically proof of a good or bad policy. Researchers must examine the comparison group, the size of the effect, its statistical reliability and how it affects actual selection. The Yale team did that by distinguishing favourable ratings from winning outcomes. Their central contribution is not a simple accusation that AI always discriminates, but evidence that superficial corrections can fail to address the consequential result.
A model cannot provide the final definition of fairness
There is no single machine-generated number that settles whether a complex decision was justified. Different organisations have legitimate requirements, but they must decide how those requirements relate to the opportunity being offered and whether they can defend the procedure. When artificial intelligence is used, the decision-makers remain responsible for the criteria, evidence and consequences. A plausible written explanation from a model should not be mistaken for an independent audit of its own behaviour.
One practical lesson is to retain the distinction between suggestions and authority. An AI tool can help organise material, identify gaps or prompt reviewers to ask more consistent questions. Giving it the power to determine who wins requires stronger evidence of reliability and fairness. Even when a human signs off the result, the organisation should ask whether the system shaped the shortlist in ways that were never properly examined.
The debate also connects to the difficulty of making algorithmic financial advice genuinely impartial, although investment advice and startup-selection experiments are different settings. In both, a polished numerical answer can encourage people to trust an evaluation more than they understand its assumptions. Users should be able to inspect what the number means before acting on it.
The Yale study does not show that machines are incapable of fair decisions. It shows why fairness cannot be inferred from polite language or a slightly improved score for a particular group. If the same pitch wins or loses differently when a name changes, or if favourable ratings fail to affect the final selection, the process deserves scrutiny. The important question is not whether an AI can appear fair. It is whether fairness survives at the moment a consequential choice is made.
About the Author
Stuart Kerr is Technology Correspondent at LiveAIWire, covering artificial intelligence, cybersecurity and the social impact of emerging technology. LiveAIWire is an independent, human-led technology publication using AI-assisted research, editorial production and original AI-assisted editorial illustrations under his direction.