AI Bias Checker

AI bias is a systematic, repeatable error that produces worse outcomes for particular groups of people. Systematic is the important word. A model that gets things wrong at random is inaccurate. A model that gets things wrong in the same direction about the same kind of person every time is biased.

Nobody has to intend it, and it almost never comes from a line of code you can point at. It arrives through the data the system learned from, the way somebody framed the problem, the proxy chosen for something that could not be measured directly, and the people who labelled the examples. Every one of those is a normal engineering decision. Together they are the most common source of harm in deployed AI.

This is a commercial problem as much as an ethical one. In the UK, a system that produces materially worse outcomes for people sharing a protected characteristic can amount to discrimination under the Equality Act 2010, whether a human or a model made the call. Automated decision-making that significantly affects people carries its own obligations under UK GDPR on top.

These are patterns rather than accounts of particular incidents. They are the shapes that keep turning up in real systems, and they are what the checker above looks for in text.

Text screening is the cheapest layer and the only one you can run in a browser. It catches linguistic bias, which is both the most common form in generated content and the easiest to fix, because fixing it means changing words.

It will not catch the rest. Detecting historical, representation or measurement bias needs access to the system: performance metrics broken out by group rather than one accuracy figure, counterfactual tests that change one attribute and hold everything else still, and an honest look at what the training data actually contains.

A proper evaluation runs all of those together and repeats them after every material change to the model, the prompt or the population it serves. Bias is a property that drifts, so testing for it once tells you about one afternoon.

The UK has no single AI act. The obligations come from instruments that already applied to you, now read across to automated systems.

The Equality Act 2010 prohibits direct and indirect discrimination across nine protected characteristics. Indirect discrimination is the one that catches most AI systems: a neutral-looking rule that puts a group at a particular disadvantage is unlawful unless it is a proportionate means of achieving a legitimate aim. A model is a rule.

UK GDPR and the Data Protection Act 2018 add duties around fairness, transparency and solely automated decisions that produce legal or similarly significant effects on people. The ICO publishes specific guidance on AI and data protection, including on bias and discrimination.

Sector regulators stack their own expectations on top. Financial services, healthcare and employment each have more to say. None of it is satisfied by running a text checker, which is why this page ends with an offer to look at the system rather than the sentences.

It reads text. It has no access to your model, your training data, your evaluation metrics or your decisions. It can tell you whether a passage of writing carries bias a reader would recognise. It cannot tell you whether a system is fair.

It is also a language model itself, so it has the same failure modes it is looking for. It will over-flag cautious writing, miss bias that depends on context it cannot see, and carry the norms of whatever it was trained on. Treat every flag as a prompt for a human judgement.

Run the checker and nothing you paste is stored. The text goes to the model, the result comes back, neither is written to a database. Submit the form to unlock more runs or to download the report and the text comes to us with your address, which is what the tick box on that form says. Either way, keep personal data, client-confidential material and anything under NDA out of it. There is no good reason to send any of that to a third-party model to check wording.