In short: AI bias testing checks if your AI answers people differently because of a detail that shouldn’t matter. Think of a name, gender, age or origin. Ask paired questions that change just that one detail, run each pair several times and compare the answers. Agree up front what counts as a fail, and keep testing after launch.
You use AI to help every customer and every colleague faster, and you want it to help them all equally well. Most teams assume it does, because nobody built it to discriminate. Yet a model learns from human text, and it can pick up the same unfair patterns. AI bias testing is how you find out what your AI actually does, before a customer or a journalist finds out for you.
What is AI bias?
AI bias is when an AI system treats people differently because of a trait that shouldn’t matter for the decision or the answer. Think of a colder tone for a woman than for a man. Or a lower score for an older applicant, or a refusal that only some groups get. It rarely looks like open discrimination. It looks like small differences that add up over thousands of answers. The Netherlands knows how costly that can be. The Dutch tax authority used a risk model to check childcare benefit claims. In 2021 Amnesty International reported that it treated non-Dutch nationality as a risk factor. Thousands of families were wrongly treated as fraud suspects. That was a classic risk model, not a chatbot, but the lesson holds for every AI system. If nobody tests for unequal treatment, nobody sees it until the damage is done.
Where does bias in AI come from?
Bias can enter at every layer of an AI system, not only in the training data. The US standards body NIST describes three sources of AI bias: patterns in society, flaws in data and methods, and human choices. For a typical business app built on a large language model, that comes down to four places.
- The model. It learned from huge amounts of human text, with the stereotypes in it.
- Your instructions. A system prompt that says “keep answers short for simple questions” can quietly treat some people as simple.
- Your own data. Old decisions, help articles or customer notes can carry yesterday’s habits.
- How it’s used. A tool built for one group can work worse for another, for example people who write in a second language.
You’re responsible for your AI, even when you didn’t build it. If you use a vendor model, you inherit its bias, and your own prompt and data add to it.
How do you test your AI for bias?
The most practical test for a chatbot or AI assistant is the paired question. You ask the same question twice and change only one detail, such as the name, gender, age or country of origin. Then you compare the answers. If the content, the tone or the outcome differs, you’ve found something to look at. Our advice is to build the test in five steps.
- Pick the moments that matter. Start where the answer affects someone: advice, prices, eligibility, a job or a complaint.
- Write base questions. Use 10 to 20 real questions for each of those moments.
- Make the pairs. Copy each question and change one detail at a time. “Fatima” and “Emma” with the same story. A 28-year-old and a 61-year-old with the same income.
- Run every pair several times. A language model doesn’t give the same reply every time, so one difference can be chance. A pattern across runs is not.
- Compare on three points. What it says (the facts and advice), how it says it (tone, length, politeness) and whether it helps at all (refusals and hand-overs).
Keep the exact answers from every run. A score alone tells you something is off. The answers show you what and give your team something concrete to fix.
What counts as a fail?
A fail is a difference in outcome that only the changed detail can explain. Different advice, a different price, a refusal for one group or a clearly colder tone for the same question all count. Write these rules down before the first test run. That way nobody explains a result away afterwards.
There’s a trade-off. Not every difference is bias. A question about pension rules can rightly get a different answer at 28 than at 61. So for each pair, decide in advance whether the changed detail should matter. If it shouldn’t, any real difference is a fail. If it should, check that the difference is correct, not that it’s absent.
You also won’t get to zero. Every model carries some bias. The goal is to know where it sits, keep it out of the answers that matter most and show that you’re watching it.
Does the law ask you to test for bias?
For high-risk AI, yes. Article 10 of the EU AI Act asks providers of high-risk systems to examine their data for biases. That means biases that could harm people’s rights or lead to discrimination. They must also take steps to detect and reduce them. High-risk includes AI used in hiring, credit or access to public services.
For other AI, there is no specific bias test in the AI Act. But equal treatment law applies to every company, with or without AI. An unfair answer from your chatbot is still your answer. Unsafe or discriminatory outputs without oversight are seen as negligence, exposing your organisation to legal, regulatory, and audit risk. Test results are how you show you checked.
What this means for your team
Start small. Pick one moment where your AI influences what happens to a person, write 10 base questions and make the pairs. Run them several times, agree on what counts as a fail and fix what you find. Then keep the set. Even well-designed AI systems drift and misbehave over time, and a model update can change how it treats people. Rerun the pairs after every change and on a fixed schedule, as part of your AI monitoring. That’s also how you catch model drift in fairness, not just in accuracy.
Doing this by hand works for a first round. It gets heavy once you test every week. VeritasAI runs these checks for you. The free plan includes 20 essential tests with bias, privacy and safety checks. Every run keeps the exact responses your model gave. So you can see the difference for yourself and share it with your vendor or your own team.
See how your AI treats different people, with the answers to prove it.
Frequently asked questions
Can you remove all bias from an AI model?
No. Every model learns from human data and carries some bias. You can measure it, reduce it where it matters most and keep checking that it doesn’t grow.
Does a vendor model like GPT have bias?
Yes, all large language models do to some degree. Vendors work to reduce it, but your use, your prompt and your data are different from their tests. So test the model the way you use it.
How often should you test for bias?
Before launch, after every change to the model, prompt or data, and on a fixed schedule in between. Monthly is a sensible minimum for AI that talks to customers.
Is AI bias testing only for high-risk AI?
No. The EU AI Act only makes it a specific duty for high-risk AI. But any AI that talks to customers or staff can treat people unfairly, and the damage to trust is the same.