When AI Bias Becomes Our Bias

What happens to our own thinking when we start using AI to make decisions?

A few days ago, I shared an article about whether AI can have cognitive biases too. I looked at things like anchoring and framing, and at the ways seemingly irrelevant information or the way we phrase a question can change an AI's response.

While writing that article, I started thinking about the other side of the interaction.

What happens to us when we start using AI to make decisions?

We already know that AI can produce biased or inconsistent outputs. But once those outputs become part of our everyday thinking, something else starts happening. The answer becomes part of the way we see the problem.

And sometimes, we may carry it with us long after the conversation is over.

An AI answer can become our reference point

Imagine you're investigating a production issue. You ask AI to help you understand what might be happening, and it gives you a detailed explanation pointing toward a database problem. The explanation makes sense. It gives you several things to check, and you start investigating them.

A little later, you find a database query that looks suspicious. Now the original explanation feels even more convincing. But that suspicious query might have been there for months. It might have nothing to do with the performance problem you're investigating.

The AI answer has changed what you're paying attention to.

This is one way an AI-generated bias can become our own. The answer gives us a reference point, and everything we see afterwards gets interpreted partly in relation to it.

Research has found similar effects outside software engineering. In one set of experiments, participants were exposed to biased recommendations from an AI system and later reproduced similar errors themselves, even when they were making decisions without the AI. The interesting part is that the participants didn't need to consciously decide to trust the AI. The recommendation simply became part of the information they used to make sense of the problem.

A convincing explanation can change how we see the problem

There's another thing I find fascinating about AI conversations: fluency.

AI is very good at taking an idea and turning it into a coherent explanation. It gives you structure. It connects the dots. It explains why something might be happening and what you could do about it. That makes the information easy to work with.

It can also make a particular interpretation feel more reasonable. Research on AI-assisted decision-making has found that people can rely on AI advice even when other information points in a different direction. In a 2024 experiment, researchers found that participants over-relied on AI advice in a decision-making task, including situations where the advice conflicted with available contextual information and the participants' own assessment.

And again, people didn't need to believe AI was always right. The AI answer became another piece of information in the decision, and it carried enough weight to change people's judgments.

It is important because most of us aren't using AI while thinking, "I will now blindly trust this machine". We're usually doing something much more ordinary. We're thinking about a problem, we ask AI, and it gives us something useful to work with.

Then we continue from there.

What happens to what we remember?

This is my favorite part.

We often think about AI errors as a problem of getting the wrong answer. There's another possibility: the interaction itself can become part of what we remember.

In 2025, researchers from MIT and other institutions studied whether conversational AI could contribute to false memories. Participants first read an article and then either received no intervention, read a summary, or discussed the article with a chatbot. In some conditions, the AI conversation contained misleading information. The misleading chatbot condition produced significantly more false recollections than the control condition.

There's an important limitation here. The study deliberately introduced misleading information into the chatbot interaction, so it doesn't tell us how often ordinary AI conversations create false memories. It does show that information from a conversation with AI can become part of what we later remember.

And that's different from simply getting an answer wrong. If I ask AI about something and it gives me a detail that turns out to be incorrect, I might notice it immediately. Or I might remember the detail without remembering where it came from. Was that something I already knew? Something I read somewhere? Something the AI told me? Something I worked out myself?

The source can disappear while the information remains.

The AI becomes part of the information we use

Before generative AI, most of our information came through fairly obvious channels. A book, a person, a website, a search result, a meeting. We could usually identify where a particular piece of information came from.

A conversation with AI is different.

It sits somewhere between search, conversation, explanation, brainstorming, and advice. And because the conversation is personalised to us, it can feel surprisingly close to our own thinking.

I might start with an idea.
AI gives me another perspective.
I think about that perspective.
I change my mind slightly.
I ask another question.
The answer gives me something else to consider.

By the end of the conversation, the final idea may be genuinely useful, but it can become difficult to separate my original thinking from everything that happened during the conversation. The boundary gets blurry.

And that matters because our thinking is built partly from the information we encounter, the conversations we have, and the connections we make between them. If AI becomes something we interact with every day, its answers become part of that environment.

What should we do with that?

I'm not interested in avoiding AI. I use it too much for that.

I think it's more useful to become aware of how an AI answer can shape what we notice, what we remember, and how we interpret the information that comes afterwards.

Notice what the answer makes you look at.
When AI points you toward an explanation, pay attention to what suddenly seems important. Ask yourself what else could explain the same situation.

Keep some distance from the explanation.
A coherent answer can feel more convincing simply because it is coherent. Give yourself a chance to look for evidence that supports or challenges it.

Separate the conversation from the evidence.
An AI explanation can be useful for finding ideas and questions to investigate. A convincing explanation still needs evidence behind it.

Pay attention to what you remember.
If something important came from an AI conversation, checking the original source gives you a way to separate what you actually know from what you remember the AI telling you.

Ask for alternatives.
When an AI answer gives you one interpretation, ask what other explanations could fit the same situation. This can help keep one explanation from becoming the whole story.

The part I'm still thinking about

The strange thing about all of this is that the influence can happen without any obvious moment where we hand over a decision. There is no clear line between "the AI suggested this" and "this is what I think". The suggestion becomes a possibility. The possibility becomes something we investigate. Eventually, it can become part of the way we understand the situation.

That makes AI different from a tool that simply gives us information. It becomes part of the process through which we form our understanding.

And I think we're only beginning to figure out what that means.

Research

• Does AI Have Cognitive Biases Too? — My previous article exploring anchoring, framing, availability, representativeness, and other bias-like patterns in language models.

• Humans inherit artificial intelligence biases — Vicente and Matute, Scientific Reports (2023). Experiments examining how people can carry biases introduced by AI into later decisions.

• Confirmation bias in AI-assisted decision-making — A study of how psychologists respond to AI recommendations that agree or disagree with their initial judgments.

• Trust and reliance on AI: An experimental study on the extent and costs of overreliance on AI — Klingbeil, Grützner and Schreck, Computers in Human Behavior (2024). An experiment examining how people rely on AI advice in decision-making.

• Slip Through the Chat: Subtle Injection of False Information in LLM Chatbot Conversations Increases False Memory Formation — Pataranutaporn et al., IUI (2025). An experiment examining how misleading chatbot conversations can affect false memory formation.