Bias in, bias out: Is AI inheriting our human biases?

What is bias?

Think about how many decisions you make in a day. Your brain can't carefully weigh every single one, so it leans on noticeable patterns and past experience to speed things up. Most of the time, this works great. Sometimes, it leads you a little off course.

Human bias can be roughly broken down into two distinct categories:

• Cognitive bias: Mental shortcuts, like assuming something will happen again because it happened before.

• Social bias: Assumptions tied to group identity, like expecting certain things from someone based on their gender, age, or background.

Both of these types of bias show up in AI models,  just in different ways: 

• Cognitive bias tends to shape how a system reasons. 

• Social bias tends to shape who it treats fairly.

Psychologists have categorized these even further into specific biases, including a few you may have heard of before. Some common biases include:

• Confirmation bias: You notice and believe things that match what you already think.

• In-group bias: You're a little kinder to people in your own circle than to outsiders, often without meaning to.

• Availability bias: If something happened recently or was memorable, your brain assumes it happens more often than it really does.

• Halo effect: If someone seems good at one thing, you assume they're probably good at other things too.

• Automation bias: When a computer gives you an answer, you tend to trust it and stop double-checking, even if it's wrong.

The last one really sums up our relationships with machines. It's not really about how biased a machine is, but more about how quickly we hand over our own judgment the moment something looks automated.

The bigger idea here is that bias isn’t a glitch in the model, or even necessarily an evil thing. It’s a normal part of how our brains process information and it’s natural that these biases creep into how we train AI models. So the real question isn't "how do we get rid of bias completely." It's "how do we notice it before it snowballs."

How AI picks up our habits

These mental shortcuts don’t make you a bad thinker. They kept humans alive for thousands of years by helping us make fast judgement calls with limited information. The real issue is simply that we rarely notice them happening at the moment. So, what happens when we build machines that learn by analyzing millions of these very human, sometimes biased inputs?

AI systems learn by studying enormous amounts of human-generated data — words we've written, images we've tagged, decisions we've made. If that information includes some biased judgment (say, biased hiring decisions or biased labeling), the AI doesn't know to ignore it. It just sees it as a useful pattern and learns it like anything else.

But AI models don't just copy the bias, they can actually amplify them. That's because bias is often a handy shortcut for getting the "right" answer more often, so a system trying to be accurate has a reason to lean into the pattern.

Researchers at UCL saw this happen in a simple experiment. They asked people to look at photos of faces and decide if each one looked happy or sad. As a group, people leaned very slightly toward calling faces "sad" more often than "happy" — a tiny, everyday kind of bias. Then they trained an AI on those same judgments. 

The AI didn't just repeat the small lean toward "sad." It exaggerated it, becoming even more convinced that faces were sad than the humans who taught it in the first place. A small human habit went in; a bigger machine habit came out.

The feedback loop: Can AI make us more biased too?

The relationship between human bias and AI bias isn't one-directional. It loops back around.

In that same face study, a new group of people did the identical happy-or-sad task, except this time they could see what the AI had decided for each photo. After spending some time working alongside the AI, this group started calling faces "sad" even more often than before — more biased than the original group who had trained the AI in the first place. 

So the loop went: small human bias, bigger AI bias, even bigger human bias. People mostly had no idea how much the AI was shaping their thinking. Nobody sat there consciously deciding to become more biased. It just happened quietly in the background.

The next part of the study threw in a fun psychological twist: people were told they were talking to another person while actually talking to an AI. Surprisingly, they picked up the bias less. Researchers think this might be because we tend to expect AI to be more accurate than a person. 

Ironically, that trust in AI's “fairness” is exactly what makes us more likely to absorb its biases without questioning them. We tend to let our guard down around anything that looks neutral.

Can AI training actually fix this bias loop?

Good news: yes — at least partly. The goal isn't to eliminate bias forever (probably impossible). It's to build AI thoughtfully enough that it nudges us toward fairer thinking instead of away from it.

Notice your own first reaction before you rate or write

• Spot it: Your gut reaction to a prompt or response is often the fastest place bias hides. If a name, accent, or context nudges your judgment before you've really thought about it, that's worth pausing on.

• Mitigate it: Slow down for a beat on anything involving people, identity, or judgment calls. Ask yourself: would I rate this the same way if the name, gender, or background were different?

• Check it: Every so often, swap details in a prompt (a name, a pronoun, a location) and see if your instinct changes. If it does, that's useful information about your own blind spots, not a failure.

Watch for patterns in the examples you personally choose or write

• Spot it: If you're writing prompts, examples, or sample responses, it's easy to default to the same type of person, scenario, or voice without realizing it — usually whatever feels "neutral" to you.

• Mitigate it: Deliberately vary who and what shows up in your examples: different names, ages, genders, professions, and contexts, especially for anything meant to represent "a typical person."

• Check it: At the end of a batch of tasks, skim through what you produced and ask who's represented and who isn't. A quick pattern check like this takes minutes and catches a lot.

Don't let "sounds right" replace "is right"

• Spot it: When you're reviewing or rating AI output, a confident, well-written answer can feel more correct than it actually is. This is the same trap as automation bias — trusting something because it sounds authoritative.

• Mitigate it: Separate how something is written from whether it's actually accurate or fair. A well-written response and a correct response aren't the same thing.

• Check it: Before approving or rating something highly, ask yourself specifically: am I checking the facts and fairness here, or just the tone? If you can't answer that quickly, look again.

Flag stereotypes even when they're subtle or "common knowledge"

• Spot it: Bias in responses doesn't always look dramatic. It often shows up as a small, casual assumption. Maybe a certain job "usually" being done by a certain gender, or a certain group being described a certain way "by default."

• Mitigate it: Treat mild, common stereotypes as seriously as obvious ones. If a response leans on an assumption instead of stating a fact, that's worth flagging or correcting.

• Check it: Ask yourself if the same sentence would sound odd or unfair if you swapped the group being described. If yes, it's worth a second look.

Compare notes with other trainers when you can

• Spot it: Bias is hard to catch alone since it often feels invisible from the inside. What feels "normal" to you might not to someone with a different background.

• Mitigate it: When possible, talk through tricky or borderline cases with other people doing similar tasks rather than relying only on your own judgment.

• Check it: If you and a colleague rate the same tricky example very differently, treat that gap as useful data, not a disagreement to brush past. It usually means the example (or the guidance around it) needs a closer look.