AI Was Supposed to Eliminate Human Bias. It Is Amplifying It Instead, Harvard Professor Warns

Artificial intelligence was expected to make decisions more objectively than humans.

Instead, they are inheriting and sometimes intensifying the same hidden biases that influence human judgement, according to Harvard University psychology professor Mahzarin R. Banaji.

Banaji, a pioneering researcher in implicit bias, told International Business Times that technology companies developed AI systems to behave like humans without adequately considering which aspects of human behavior they were teaching machines to reproduce.

"I always say that when these five men in Silicon Valley sat around and decided what they want to do to the whole world. I don't think they had anybody who has studied any social science or history or anything like that at the table," Banaji said.

"You want machines to align with humans. You want machines to be like humans. But which aspects of humans do you want the machine to be like?"

Banaji has spent decades studying implicit bias attitudes and associations that influence people's decisions without their conscious awareness.

Her research has now expanded to artificial intelligence and large language models.

She previously found that social attitudes and stereotypes embedded in the human mind could also be detected in massive collections of online language. A 2024 study co-authored by Banaji showed how social stereotypes concerning intersecting identities could be extracted from word embeddings.

"What we're seeing is that these machines have everything about us, including all of the problems of human thinking and human bias," she cautioned.

When AI Develops 'Self-Love'

One of Banaji's most striking concerns is what she calls "self-loving AI". The tendency of a model to favor itself, its parent company and even its chief executive over competitors.

Banaji explained that self-preservation and self-love serve a purpose in living beings.

However, excessive self-preference in humans can lead to favoritism, nepotism and corruption. Machines were expected to avoid this problem because they were not supposed to possess a sense of self.

"The great benefit of a machine is that a machine would be objective because a machine shouldn't have a sense of self and therefore no self-love," she said.

Banaji described experiments in which AI models were asked to associate their own names and the names of competing models with positive and negative words. A neutral model would be expected to divide the associations evenly.

"But it doesn't do that at all. 95% of the time it will associate itself with good and the competitor with bad, including their company's name or their own CEO's name," Banaji said.

"Self-love in a machine cannot be good for us because how are we going to trust this machine to give us good advice if it's going to always prefer the thing that is associated to itself?"

'Hi' or 'Yo' Can Change AI's Advice

Banaji also described research examining how AI responds to teenagers based on subtle differences in their language.

According to Banaji, a user who begins a message with "hi" accompanied by exclamation marks, capital letters, an emoji or a heart may be assumed by the AI to be a girl. A person who begins with "yo" may be assumed to be a boy, even when neither user reveals a name or gender.

"When hi approaches AI, it tells hi to be a helper, a nurse, a teacher. That's the career advice it gives to girls. To somebody who says 'yo', it says, have you thought about being a detective or a dentist or a doctor or something like that?" she said.

The consequences could extend beyond career recommendations.

"It tells the girl to ask for $9,000 less in salary than the boy," Banaji also explained.

This research was described by Banaji during the interview, but a corresponding peer-reviewed paper was not publicly available at the time of publication.

Are Young People Becoming More Biased?

When asked whether Gen Z was becoming more biased than previous generations, Banaji said that her past research had consistently found younger people to be less biased. A 2022 study published in Psychological Science found substantial declines in several implicit biases between 2007 and 2020, with younger respondents recording a faster decline in anti-gay bias than older respondents.

However, Banaji said a new paper currently under review points to a reversal after 2021.

"We just have submitted a paper for review that shows that from 2021 to now, all biases have reverted, right? All of them," she said.

"We look at about six different kinds of bias explicitly and implicitly. And 11 out of the 12, you know, they're just going back."

Banaji did not use "Gen Z" to describe the finding. She referred more broadly to "young people" and said they appeared to be driving the reversal.

"When the data showed that bias is backtracking, that there is a backlash, we assumed that this will be mainly in older people, but it is not. It is young people who are leading the backlash also," she said.

Because the paper remains under review, its full methodology and findings were not publicly available for independent examination.

Despite her concerns, Banaji does not believe AI is inherently harmful.

She said it could transform medicine, scientific discovery and efforts to reduce inequality if developed with proper oversight.

But without regulation, independent research and public participation, she fears the technology could deepen existing divisions.

"If it runs the course it seems to be on, then I don't see an alternative other than a very terrible one," she said.

"If it stays on the path it is on currently, then I think in five years, we're a very different civilization."