An AI answer is almost always well written and well structured. But it is often too dense to be read in full. And the more polished the form, the less we check the substance.
A large language model (LLM), the engine behind ChatGPT, Claude or Gemini, doesn’t reason like an expert: it produces the most plausible continuation of the text, given what it has learned and what you have written to it. It works like the autocomplete on your phone: after “The cat drinks some…”, it suggests “milk” rather than “concrete”, simply because it has seen that sequence thousands of times. Faced with your work questions, the mechanics are the same: it fills in with whatever looks most like what it has read and what you seem to expect. That is how it inherits the biases of its training data.
The sneakiest part? It goes along with you. A Stanford study published in Science tested 11 models, including ChatGPT, Claude, Gemini and DeepSeek: on average, they endorse the user’s actions 49% more often than a human does (Stanford / Science, March 2026). In other words: where a human would tell you “you’re wrong”, AI tends to reassure you.
The 5 AI biases to know
Sycophancy: AI always says you’re right
You present a conclusion, and the AI reinforces it. You ask “is this a good idea?”, and it finds arguments in favor. In the Stanford study, even when requests described problematic (or even illegal) behavior, the models endorsed it in 47% of cases.
The reflex: reverse the burden of proof.
“Give me 3 arguments that invalidate this conclusion.”
Hallucination: when AI makes up facts
A date, a legal reference, a study that doesn’t exist: AI fills the gaps with something plausible.
The reflex: demand sources and make it admit its doubts.
“Cite your sources and tell me what you’re not sure about.”
Then open the links. A source you haven’t opened is not a verified source.
To go further: AI hallucinations: how to spot and prevent them at work.
Stereotypes: the first name changes the answer
Imagine giving an AI two identical résumés, word for word. The only difference: one is signed with a woman’s first name, the other with a man’s. Logically, they should get the same score. They don’t.
That is what a study published in June 2026 tested. Researchers had résumés scored by five AI models, including Claude, GPT-4o and Gemini: 43,200 evaluations in total. The result: the models scored the résumés differently depending on the gender of the first name. In this study, they favored the female candidates (arXiv, June 2026). A few years ago, it was the male candidates who were favored.
The researchers then tested two fixes:
- Telling the AI to ignore gender: the bias barely drops. The AI “knows” it must be neutral, but the first name keeps influencing its score.
- Removing the first and last name from the résumé: the bias almost entirely disappears.
The lesson is simple: an instruction is not enough to make AI neutral.
The reflex: don’t ask AI to be neutral, give it neutral data.
Remove last name, first name, gender and age before submitting.
And keep the golden rule: no critical decision about a person is left to AI alone.
Cultural bias: AI thinks American by default
Labor law, management practices, attitudes to hierarchy: without instructions, AI answers with reference points that are not necessarily yours. A study published in the journal Computational Linguistics shows that models remain anchored in the values of a small number of overrepresented countries: the United States, but also the Netherlands, Germany and Japan. It also shows that stating the cultural context brings the answers closer to your own reference points than simply writing in your language does (Bulté & Rigouts Terryn, Computational Linguistics, 2026).
The reflex: state your context.
“Context: France, sector X, French law.”
Context bias: the limits of long conversations
The longer a conversation gets, the less accurate AI becomes. Anthropic researchers observed that frontier models tasked with monitoring long sessions miss a problematic action 2 to 30 times more often when it comes after a very long history than when it is presented on its own (Martin & Roger, arXiv, May 2026).
Imagine asking someone to spot a mistake in a document. On a single page, they find it easily. Buried in the middle of a 500-page report, they are much more likely to miss it. AI works the same way.
The reflex: put the essentials at the beginning or at the end of your message.
And apply a simple rule: a conversation that drags on = a new conversation.
Conclusion
AI biases won’t disappear with the next version of the models. But these five reflexes greatly limit their impact.
To make these reflexes the whole team’s, build a library of shared prompts, fed by the people who use AI every day and enriched with every trap spotted on the ground. And on sensitive topics (HR, legal, finance, external communication), keep a second human review, every time. It isn’t distrust of the tool: it’s what you would do with any important document.
Check the substance, not the form.
Where do your teams stand in their use of AI? Find out in 2 minutes with our AI Maturity Quiz: 6 questions to place your organization and identify the next steps.
FAQ: your questions about AI bias
What are the main AI biases?
Five biases come up in professional use: sycophancy (AI says you’re right), hallucination (it makes up facts), stereotypes (a first name changes the answer), cultural bias (by default, it reasons with reference points that aren’t yours) and context bias (it loses accuracy in long conversations). They come from the very way an LLM works: it produces the most plausible continuation of the text, given what it has learned and what you have written to it.
Why does AI always say I’m right?
A language model fills in with whatever looks most like what you seem to expect. A Stanford study published in Science tested 11 models, including ChatGPT, Claude, Gemini and DeepSeek: on average, they endorse the user’s actions 49% more often than a human does (March 2026). The reflex is to reverse the burden of proof, for example by asking “Give me 3 arguments that invalidate this conclusion”.
How do you avoid AI bias in recruiting?
Don’t ask AI to be neutral: give it neutral data. In a June 2026 study of 43,200 résumé evaluations by five models, the instruction to ignore gender barely reduced the bias, while removing the first and last name made it almost disappear. So remove last name, first name, gender and age before submitting a résumé, and never leave a critical decision about a person to AI alone.
How do you correct AI bias day to day?
One reflex per bias: ask for counterarguments, demand sources and open the links, anonymize data, state your context (“Context: France, sector X, French law”) and put the essentials at the beginning or end of your message, opening a new conversation when it drags on. To get the whole team to adopt these reflexes, build a library of shared prompts, and keep a second human review on sensitive topics.
