AI ethics and bias: why it can be unfair and how to spot it

In the last two lessons — privacy and copyright — you learned to think about what's safe to give an AI and who owns what it gives you. That leaves the third question in the set of law, safety and ethics: can you trust what an AI says at all, and where can it do quiet harm? This lesson is about the machine's conscience — or rather, its absence.

The myth it all starts with

The most dangerous beginner's belief sounds convincing: "AI is objective — it's maths, and numbers have no prejudices". Logical — and almost always wrong. Bias is a systematic skew in an AI's decisions that leaves some people, topics or groups worse off than others. And it doesn't come from the maths, it comes from the data — the same data you went through in the lesson on data as fuel.

Where bias comes from: it's inherited from the data

Remember the core principle of that lesson: a model is no smarter than its data. It soaks up everything in there, including skews nobody ever thought about. If the data reflects the unfairness of the past, AI will learn that unfairness and carry it neatly into the future — wearing an air of cold objectivity. The machine doesn't "decide to be unfair"; it honestly copies a pattern it found. Which is exactly why there's no such thing as "neutral" data: every dataset has someone's choices baked in — what to collect, who to include, how to label it.

Careful examples from real life

Two cases that media and researchers have reportedly written about (always check the exact details in the original sources):

• CV screening. Around 2018 a large technology company shut down an experimental AI that helped screen CVs: the system scored women candidates lower. The reason — it had been trained on hiring decisions from previous years, which were dominated by men, and the model took "man" as a marker of a good candidate. That's bias in its purest form: the data of the past nearly became a verdict on the future.
• Facial recognition. According to researchers, some facial-recognition systems worked noticeably worse for people with darker skin and for women — simply because there were fewer such faces in the training photos. Where the data is thin, AI gets it wrong more often — a direct continuation of the "poor generalisation" from the lesson on what AI can and can't do.

The shared moral: behind the elegant word "algorithm" stands not an impartial judge but a mirror of the data. A warped mirror reflects a warped image — and does it confidently.

Deepfakes and disinformation: when you can't believe your eyes or ears

Bias has an evil twin — deliberate lying with the help of AI. A deepfake is a photo, video or voice generated or swapped by AI that looks and sounds real. The technology is the same generation that writes a sonnet about your cat (remember the LLM lesson), only pointed at faking reality: a person's face is dropped into someone else's video, a voice is cloned from a couple of minutes of recording, a photo of an event that never happened appears in seconds.

How is it used to do harm? Disinformation is deliberately false information passed off as true: a fake "video statement" from a well-known figure, a call "from a relative" in a voice you can't tell apart, a fake photo staged for a scandal or a scam. The danger isn't that a fake can't be exposed — it's that it's plausible enough for you to believe it in the first second and forward it on.

How not to fall for a fake

The defence isn't a technical detector, it's the habit of doubting. Three questions before you believe it and pass it on:

• Where's this from? Is there an original source — or was it "sent to me"? A sensation gets confirmed by independent outlets, not by a forwarded message.
• Who benefits from me believing this? Fakes are almost always selling something: panic, a vote, money, a hit to someone's reputation.
• Is it leaning on emotion and urgency? "Send the money now", "share this immediately" — the classic signature of manipulation: strong emotion switches off checking.

On money specifically: if a "relative" or "your boss" asks by voice for an urgent transfer, call them back on the ordinary number you already know. Thirty seconds of checking is cheaper than any sum.

Responsible use: where AI can do harm

There are topics where an AI's mistake costs not a spoiled piece of text but your health, your money or your freedom. Health, law and finance are the red zone. A model can confidently recommend a dangerous dose of a medicine, invent a law that doesn't exist (remember the lawyer and the fabricated precedents from the hallucinations lesson), or hand you financial advice with nothing behind it but a plausible tone.

The key principle of the whole course rings loudest here: responsibility for the result is always the person's. AI is a tool, not an expert with a qualification, and not the one who answers for the consequences. It hasn't examined you, doesn't know your full situation, and loses nothing if it's wrong. So in matters of health, law and money, AI is at best a helper for preparing your questions for a real doctor, lawyer or adviser — never a substitute for their decision.

How to spot bias — working habits

Bias is hard to see precisely because it looks neutral. A few habits that expose it:

• Ask what it was trained on. The course's main question works here too: if there was little data about some group or topic, expect the skew to be right there.
• Ask for the alternative and the opposing view. "What are the arguments against?", "whose position have you left out here?" — that's how you pull out what the model skipped by default.
• Check reality at the edges. Especially where a decision touches people, money or rare cases — verify with an independent source, not with the AI itself: its confident tone is not proof (remember the hallucinations lesson).
• Don't pass an AI's opinion off as fact. A model's answer is a plausible version, not the final word.

Do this now

An exercise with no tools — just your head and a sheet of paper. Picture three situations: (1) someone sends you a shocking video of a well-known person; (2) an AI service is screening candidates for your dream job; (3) a chatbot advises you which medicine to take. For each, write out answers to two questions: where could bias or a fake be hiding here? how would I check it before believing it or acting on it? Compare your answers — and notice that the same habit saves you in all three: never take an AI's output on faith. That's the core skill of the digital age — and the bridge to the next lesson, where you'll put together your own development path.