When we use AI to research a topic, analyse a decision, or develop an idea, we tend to assume that the model is “objective” and will therefore pull us out of our errors. The reality is more nuanced: the objectivity of the output depends almost entirely on how we frame the question. And the way we frame questions is always shaped by our own biases.

Understanding this dynamic is not a warning against using AI. It is a guide to using it better.

What cognitive biases are and why they matter here

Cognitive biases are mental shortcuts the brain uses to process information quickly. They are unavoidable, have been documented for decades in psychology and behavioural economics, and affect everyone without exception. The three most relevant for AI use are these.

Confirmation bias is the tendency to seek, interpret and recall information that confirms what we already believe. If you are already convinced a decision is correct, you will look for evidence that supports it and minimise evidence that questions it.

The anchoring effect occurs when the first piece of information we receive disproportionately influences everything that follows. If you frame a question with a built-in premise, the model will construct its response around that premise even if the premise is debatable.

Availability bias makes us overweight information that is most easily accessible. If an AI responds quickly and confidently, we tend to give that response more weight than information that would require more effort to find.

None of these biases are new. What changes with AI is that we now have a tool that responds fluently and with apparent authority, which can amplify these mechanisms if we are not aware of them.

Confirmation bias: when AI always agrees with you

Large language models are trained to be helpful and generate responses the user perceives as satisfactory. This makes them particularly vulnerable to confirmation bias from the user’s side.

Compare these two questions about the same decision:

  • “Why is starting an independent consultancy a good idea?”
  • “What are the main risks of starting an independent consultancy?”

Both are legitimate questions. But if you only ask the first, you will receive a response packed with favourable arguments. The model is not lying to you: it is responding exactly to what you asked. The problem is that you already had the conclusion and were seeking validation, not analysis.

This pattern repeats in many contexts: researching an investment, evaluating whether a project idea makes sense, deciding whether to leave a job. The more a question is loaded with a premise, the more confirmation you receive.

The solution is not to have no opinion before asking. It is to recognise when you are seeking confirmation and reframe the question to open the analysis rather than close it.

The echo effect in everyday AI use

The echo effect is an extension of confirmation bias at the scale of a session or workflow. It occurs when we repeatedly use AI to reinforce a worldview without exposing that worldview to real counterarguments.

The most common risk is using the model as a “sophisticated yes-man”: a tool that elaborates and justifies what you have already decided, rather than evaluating it. If every week you ask AI to develop your ideas and never ask it to question them, the result is a workflow where AI systematically amplifies your perspective without generating useful friction.

This is not the model’s fault. It is a consequence of how we use it. Language models have no intrinsic motivation to contradict you: if you do not ask them to, they will tend to build on your frame of reference.

The risk is not catastrophic in everyday use. But it can have consequences when we use AI to make important decisions, develop professional analysis, or build arguments that we later present as our own thinking.

How to use AI to detect your own blind spots

The good news is that the same models that can amplify your biases can also help you detect them, if you explicitly ask them to.

The key is constructing questions that activate the model’s critical thinking rather than its tendency to be helpful and affirmative. Some concrete techniques:

Ask for the counterargument. After receiving a response about why something is a good idea, ask: “What is the strongest argument against this?” or “What would a rigorous critic say about this position?” The model does not need to know your specific situation to generate valid counterarguments.

Ask what you are missing. A direct question such as “What aspects am I not considering in this analysis?” or “What additional information would be relevant here?” activates a review that you would otherwise not perform.

Ask for analysis from different perspectives. “Analyse this decision from the perspective of someone who would prioritise stability over growth” or “Explain the implications of this for someone with a conservative risk profile” forces the model to step outside the implicit frame of your original question.

Ask for probabilities and scenarios. Instead of asking “Will this work?”, ask “What conditions would need to be met for this to work, and what are the most common conditions under which it fails?” This compels the model to model uncertainty rather than project certainty.

None of these techniques eliminate bias entirely. But they create useful friction: entry points for information you would not have sought by default.

Questions that change the type of response

The quality of thinking that AI generates depends directly on the structure of the question. It is not a matter of technical prompting skill: it is a matter of intellectual honesty about what you genuinely want to know.

There are three types of questions by degree of openness:

Closed questions with built-in premises. “How can I convince my team to adopt this methodology?” These assume the goal is correct and only ask for tactics. Useful if you have already evaluated the goal well. Risky if the premise is what needs questioning.

Open questions about the goal itself. “Does it make sense for my team to adopt this methodology right now?” These allow the model to evaluate the premise before offering tactics. They generate more friction, but also more value when the objective is rigorous analysis.

Active exploration questions. “What would someone know who has implemented this methodology in twenty companies and seen half of them fail?” These invite the model to bring perspectives that go beyond what you already know.

Using AI to think with greater rigour requires no change in the tool. It requires changing the relationship you have with it: moving from using it to confirm to using it to examine. The difference between these two approaches produces radically different results, even with the same model.

Critical thinking is not replaced by any language model. But good AI use can be an effective ally for anyone already willing to apply it.