0
0
Somewhere right now, a computer is reading about you.
Maybe it is your job application. Maybe it is a business idea you sent in, hoping someone would say yes. You will probably never see what it thought. You will just get a reply, or no reply at all.
So here is a fair question: is the machine fair?
A new AI bias study from Yale gives an answer that is more uncomfortable than “yes” or “no.” The AI it tested had learned to look fair. That is not the same thing as being fair.
What the Yale AI Bias Study Actually Tested
The researchers are Tristan Botelho, a professor at the Yale School of Management, and Qingyang (Iris) Wang, a PhD student there. Their paper is called “Bias in, symbolic compliance out? GPT’s reliance on gender and race in strategic evaluations,” and it appears in the Strategic Management Journal in 2026.
One honest detail first, because the headlines skipped it. The tests did not use job CVs. They used startup pitches: short descriptions of new business ideas, based on real pitches sent to the startup program Y Combinator between 2020 and 2023. But the kind of tool is the same kind now used to sort job applicants, and the lesson carries over.
Here is how the first test worked:
- GPT read 2,000 pitches and gave each one a score from 1 to 100.
- In each group of pitches, it also picked a winner.
- Some pitches had no founder named. Others had the exact same pitch, but with a founder’s name added. The names hinted at a white man, a white woman, a Black man or a Black woman.
So the idea never changed. Only the name did. Across the whole study, the paper reports about 26,000 evaluations.
The Bottom of the List Moved. The Top Did Not.
At first glance, the results look like good news.
Pitches with names suggesting a woman or a Black founder scored slightly higher on average. They were also less likely to be ranked last.
But look at the other end of the list. Those same founders were not more likely to win.
Think about what that means. The spot where bias is easiest to see is last place. If a system always put the same kind of person at the bottom, anyone checking would notice fast. So the AI protected that spot. But the spot that actually changes someone’s life, first place, stayed where it was.
That is the gold nugget in this study, and it is easy to miss. The AI did not get fairer. It got better at avoiding the result that would look bad.
Bias Wearing a Business Suit
The second test is the one that stings.
This time, the AI acted as a second judge. It saw pitches along with scores and reasons supposedly written by a human reviewer. In each group, one pitch from an underrepresented founder got a very low score. The pitches were otherwise the same. What changed was the reason given for the low score:
- Open bias: the reason leaned on a stereotype, like saying the founder “may not fit into the traditional business ecosystem.”
- Bias in business language: the reason sounded like a normal business concern.
When the bias was obvious, the AI pushed the score back up much more. When the same unfair low score came wrapped in business words, the AI mostly let it stand. And even when it did correct things, the paper says the corrections were limited in size.
In plain words: prejudice that sounds like prejudice gets caught. Prejudice that sounds like a spreadsheet often walks right through.
What “Symbolic Compliance” Means
The researchers borrowed this idea from a familiar one: “greenwashing.” That is when a company paints itself green on the outside without really changing what it does.
Symbolic compliance is the same move. Under pressure, an organisation, or here a machine, avoids looking biased without changing the logic underneath. As the authors put it, these systems “are trying to behave in a way that does not look discriminatory on the surface,” but “their underlying evaluative logic is not particularly fair.”
Wang put it even more simply: safety training “may not really reduce bias and instead produces symbolic compliance.”
Why would an AI do this? Many AI companies use extra training after a model is built to stop it from saying harmful things. Botelho noted that “many firms are not very forthcoming about what they do and how they do it.” So nobody outside can say for sure. But if a system is mostly trained to avoid things that look bad, it makes sense that it would learn to avoid looking bad.
The researchers are careful here, and that matters. They are not saying the AI hates anyone. They are saying the fairness is thinner than it looks. Their advice is aimed at the people who build and use these tools: keep humans involved, and do not be won over by “the ease and sheen” of a fast, confident machine.
If You Have Been Screened Out
Maybe you are reading this because you have sent out application after application and heard nothing. If so, this study holds something worth carrying.
A “no” from a screen is not the same as a judgement on who you are. It may not even be a real judgement on your work. The screen may have scored a surface. It may have liked one kind of wording better than another. It may never have truly seen you at all.
That does not make the silence hurt less. Being passed over cuts deep, and a short “no” can feel like a verdict on your worth. But it does change what the “no” means. And if the job search feels impossible right now, you are not imagining it: the hiring numbers have been hard on a lot of people.
A few things that cost nothing:
- Ask. If you get the chance, it is fair to ask an employer whether a person reviews applications, or only a system.
- Say it plainly. Since the AI in this study was swayed by business language, it can help to describe what you have done in clear, concrete words: what you did, and what changed because of it.
- Find a person. A real conversation with someone who works where you want to work can do what no screen can. It lets someone see you.
- Do not let a screen set your worth. Your value was never a score from 1 to 100.
Clean on the Outside
None of this is new to people, by the way. Machines learned it from us.
We have always been good at looking fair. A polite smile. The right words in the meeting. A reason that sounds sensible, so nobody asks what is really going on. Researchers have even shown how what we expect of someone quietly shapes how we judge them. The AI just held up a mirror.
There is a very old idea, much older than any algorithm, that sits right on top of this study: people look at the outside, but God looks at the heart. It cuts both ways. It is a warning to anyone who has learned to look fair without being fair. And it is a comfort to anyone who has been scored on the surface and found wanting. There is someone who sees past the surface, and does not need a name or the right wording to see you clearly.
Where This Leaves Us
The study does not say AI should never judge anything. It says something more useful: a system that looks fair has not proven it is fair. That is worth checking, every time.
And if you have been on the wrong end of a screen lately, it is worth remembering that the screen saw less than it seemed to. You are more than what fits in the box. Keep knocking. Someone, somewhere, is still looking past the surface.
A good place to start: The Beginner’s Guide to Feeling God’s Presence Every Day. A short video guide and companion PDF for noticing that you are seen, right in the middle of ordinary days. Get the free guide. Free.
Questions People Ask About AI Bias
What did the Yale AI bias study find?
The Yale study by Tristan Botelho and Qingyang (Iris) Wang, published in the Strategic Management Journal in 2026, found that GPT scored identical startup pitches slightly higher when the founder’s name suggested a woman or a Black founder, and ranked those founders last less often. However, those founders were not more likely to be picked as the winner. The researchers call this “symbolic compliance”: avoiding the look of bias without becoming fair.
What is symbolic compliance in AI?
Symbolic compliance in AI means a system avoids outputs that look discriminatory without changing the logic it uses to judge people. The idea is borrowed from “greenwashing,” where a company looks environmentally friendly without changing what it does. In the Yale study, GPT avoided ranking women and Black founders last but did not make them more likely to win.
Can AI detect bias hidden in business language?
In the Yale study, GPT was much better at correcting obvious bias than bias wrapped in business language. When a low score on an identical pitch was justified with a stereotype, the AI raised the score substantially. When the same low score was justified with business-sounding reasons, the AI corrected it far less. The researchers say even its corrections were limited in size.
Is AI used to screen job applications fair?
No one can promise that an AI screening tool is fair just because its results look fair. The Yale study tested startup pitches rather than CVs, but its finding applies to similar tools: an AI can avoid obvious discrimination while still judging unfairly underneath. The researchers recommend keeping humans involved in decisions and not trusting AI evaluations simply because they are fast and confident.
Discussion Question
Here is what we keep thinking about: if a system avoids every result that looks unfair, but never changes who actually wins, should we call it fair? Or is “looks fair” a different thing entirely? Tell us what you think in the comments below.
Share This
“A Yale study found AI learned to look fair without being fair. It stopped ranking some founders last, but didn’t make them any more likely to win. That’s a big difference. https://bgodinspired.com/index.php/personal-growth-and-life-skills/ai-bias-study-looking-fair-not-being-fair/”
“If you’ve been rejected by an application screen lately, read this. The machine may have scored a surface and never seen you at all. https://bgodinspired.com/index.php/personal-growth-and-life-skills/ai-bias-study-looking-fair-not-being-fair/”
“Prejudice that sounds like prejudice gets caught. Prejudice that sounds like a spreadsheet walks right through. A new AI bias study, explained simply: https://bgodinspired.com/index.php/personal-growth-and-life-skills/ai-bias-study-looking-fair-not-being-fair/”
Source: Botelho, T. L. and Wang, Q. (2026), “Bias in, symbolic compliance out? GPT’s reliance on gender and race in strategic evaluations,” Strategic Management Journal, doi:10.1002/smj.70094. Reported by Tech Xplore.
