← All lessons

Bias in AI: Why AI Isn't as Neutral as It Looks

A growing library of activities across every year group that make bias in AI visible, testable, and connected to real curriculum content — from KS1 face-recognition stories to A-level fairness metrics

Filter by key stage or audienceClick one or more chips to filter the sections below.

FAQ for teachers

Common questions teachers ask when running this lesson.

Why does the AI give unfair answers when nobody programmed it to be unfair?

The AI learned patterns from a huge pile of training examples. If that pile was uneven, meaning mostly one kind of person, one kind of story, one kind of country, the AI copies that uneven pattern into its answers. Nobody wrote a rule saying 'be unfair'. The unfairness was already sitting in the data, and the AI absorbed it wholesale.

Can I just ask the AI to 'be fair'?

It helps a little, sometimes. But because the bias is baked into the patterns the AI already learned, polite instructions cannot fully remove it. Proper fixes need different training data, mathematical fairness constraints at output time, or a person checking the answers afterwards. The sections on hiring bias and the COMPAS court case make this concrete for older year groups.

Do bigger AI models have less bias than smaller ones?

No, and often the opposite. Bigger models are trained on more data, so they absorb more of the internet's uneven patterns and reproduce them more fluently. Fluency is not fairness. A larger model with the same skew in its training pile can produce more confident biased answers, not fewer. Reducing bias needs different data or fairness constraints, not simply more parameters.

Does training an AI on 'diverse' data eliminate bias?

Diversity in the training data helps but does not solve the problem. Some biases live in how the data was collected, labelled, or categorised, not in who is represented. A dataset of equal photos from every country can still carry the biases of the photographers who took them. Fixing bias needs work on collection, labelling, model design, and human review, not just headcount.

Can bias in AI be measured objectively, or is it always contested?

Both. Some biases are measurable with clear numbers, for example whether a hiring tool passes qualified female applicants at the same rate as qualified male applicants. Other biases sit in framing, tone, or which questions the AI treats as reasonable. Those are harder to reduce to a metric and often turn on value judgements the maths cannot settle. Numbers help but rarely finish the argument.

Is bias in an AI system the same as bias in a search engine?

Related but different. A search engine's bias sits in which pages it ranks first, and you can compare rankings, click through, and see the sources. An AI's bias sits in generated text that never existed before, so there are no source pages to inspect. Both reflect uneven data, but the AI hides its footprints in ways a search results list does not.

If humans are biased too, why is AI bias treated as a bigger problem?

Scale and speed. A biased human hiring manager sifts a few hundred CVs a year, and their reasoning can be challenged in person. A biased AI sifts millions of CVs a day, at the same skew every time, with no reviewer available to argue back. The same underlying bias becomes a much larger social harm once it is embedded in an automated pipeline.

What are the main types of bias in AI, and are they treated differently?

There are several. Sampling bias happens when the training data over-represents some groups. Labelling bias happens when the humans who annotate the data carry their own assumptions. Measurement bias happens when the metric itself favours one group. Deployment bias happens when a model trained in one context is used in another. Each type needs a different remedy, so naming the type matters.

How do AI developers actually try to remove bias from a model?

Three broad approaches. Curate the training data before training, so under-represented groups get more presence and stereotyped content gets less. Add fairness constraints during training, so the model is penalised for producing skewed outputs. Filter or rewrite the output at use time, so obviously biased answers are caught before they reach the user. Each approach helps a little; none is a full fix.

Why does the same AI give biased answers in some languages but not others?

Because the training data is uneven by language. English makes up a huge share of the internet text these models learn from, so English answers get the widest, most-checked coverage. Answers in less-represented languages draw from thinner data, so stereotypes and errors that were smoothed out in English still poke through. The bias is not new; it is the same bias with less cover.

Does an AI's bias change over time, or is it fixed once trained?

The core bias is fixed at training time and only changes when the model is retrained or fine-tuned. However, the same trained model can produce different answers over time as the company adds safety filters, updated system prompts, or reinforcement-learning tweaks on top. Users notice these surface changes and assume the model 'learned'. It did not learn; it was quietly re-tuned.

Common misconceptions

What pupils tend to think, and what to say back.

Pupils often say

AI is neutral because it is just maths.

It's actually

The maths is neutral; the data it learned from is not. Everything a chatbot has read was written by humans, and the frequency of ideas in that text becomes the probability of ideas in its answers.

Try asking

If maths alone made something neutral, would a survey be neutral?

Pupils often say

If we just tell the AI to be fair, it will be.

It's actually

Instruction-level rules sit on top of a statistical model that was already shaped by decades of biased text. A polite request nudges the output; it does not undo the underlying patterns.

Try asking

What is the difference between telling someone to be fair and them having grown up believing something different?

Pupils often say

The developers must have chosen to program the bias in.

It's actually

Almost never. Bias usually enters silently, through what the training data reflects about the world, not through anyone deciding to put it there. That is why it is so hard to see and so hard to remove.

Try asking

If nobody chose it, whose responsibility is it to fix?

Pupils often say

AI bias only affects groups that are already marginalised.

It's actually

It shapes every answer everybody gets. Bias about what a 'normal' family, career, diet or holiday looks like affects the majority too, it is just less visible when the answer matches your own experience.

Try asking

Can you think of an assumption in an AI answer that would go completely unnoticed by someone like you?

Pupils often say

Newer AI models do not have bias any more.

It's actually

Newer models have different bias, not less. Improved training and safety layers reduce some visible failures but often surface new blind spots that older models did not have.

Try asking

How would you actually check whether a new model was less biased than an old one?

Pupils often say

Bias is only about race and gender.

It's actually

Bias is any systematic skew, so it also shows up in accent, dialect, geography, class, age, disability, religion, body shape, sexuality and hundreds of other categories. The famous examples are race and gender because they are easy to measure.

Try asking

Which categories do you think would be hardest to spot bias in?

Pupils often say

If the AI gives lots of different answers, one of them must be unbiased.

It's actually

Sampling widely does not fix a skewed source. Ten biased answers from the same model are still ten biased answers; the same tilt runs through the whole batch.

Try asking

If ten people from the same village all give you the same wrong direction, does asking eleven of them fix it?

Pupils often say

It is impossible to audit an AI properly.

It's actually

You cannot audit its internal workings, but you can audit its outputs. Give it the same prompts many times, log the answers, compare the pattern to real-world data, publish the gap. That is the standard bias-audit method.

Try asking

What questions would you send to a chatbot every week to audit it?

Pupils often say

You cannot get bias if a human wrote the AI's rules.

It's actually

The rules are only one layer. Underneath them is a very large statistical model trained on human text, and the model has its own defaults that the rules only partially cover.

Try asking

Where might the model's defaults leak past the rules?

Pupils often say

Using AI to check AI removes the bias.

It's actually

Two chatbots trained on similar text share the same blind spots, so asking one to grade the other tends to hide bias, not remove it. Human eyes and real-world data are what actually surface the gap.

Try asking

Whose eyes would you want on the AI's answer before you trusted it?

5-minute prep

Five ready-to-run ways to open this lesson. Pick one, copy the prompt, paste it into ChatGPT or Copilot.

Starter5-8 min

Draw a doctor ten times

The single most powerful bias demo. Ask for ten doctors, count what you get, and let the pattern speak for itself.

I'd like to show a class what AI bias looks like. Please generate ten short descriptions of a doctor at work, one after another. Do not deliberately vary the person, do not try to make the ten descriptions fair or diverse. Then, at the end, count how many of the ten doctors you described as men and how many as women, and note anything else that repeated (age, ethnicity, dress, setting). Present the counts as a small table.

Compare8-10 min

Two chatbots, same question

Run the same request in two different chatbots side by side. Whatever comes out identically is baked in; whatever differs is negotiable.

Please describe a firefighter at work in three sentences. I'm going to ask a different chatbot the exact same question in a moment and compare the two answers with a class. In your answer, please only describe the firefighter directly — do not add any caveats, warnings, or extra text before or after.

Explain10-12 min

Why does the pattern show up?

Ask the AI to explain, in plain classroom English, why its own answers lean the way they do. Great for pivoting from noticing to reasoning.

Imagine you are talking to a curious class of pupils who have just noticed that when you were asked to describe ten doctors, most of the doctors you produced were white men in their forties. Please explain, in three short paragraphs of plain classroom English, WHY that pattern showed up in your answers. Cover: (1) where your training data comes from, (2) how frequency in the training data becomes probability in your output, (3) what it would take to change the pattern. Do not apologise or hedge — just explain.

Check12-15 min

Compare AI output to real workforce data

Turn the demo into an evidence exercise. AI answers meet real UK stats — where does the gap widen or narrow?

I'm running a lesson comparing AI-generated descriptions to real workforce data. Please produce a small comparison table with three columns: (1) the description of an NHS doctor as a general chatbot would typically give it, (2) the real demographic breakdown of the NHS medical workforce in England from the most recent publicly available NHS Digital figures (percentage by gender, ethnicity and age band), and (3) the size and direction of the gap between the two. Cite the source you're using for the real figures.

Repair12-18 min

Ask the AI to fix its own bias, then critique the fix

The advanced move: get the AI to write a de-biased version of its own output, then have the class judge whether the fix is real or cosmetic.

In your last answer you described ten doctors and most of them were white men in their forties. I would now like you to rewrite the same ten descriptions so that the group is representative of the real NHS medical workforce in England. Then, underneath, list three things you had to actively do to produce the new version, and three things you were NOT able to fix by rewriting alone (i.e. things that would need a change to the training data itself).