Part 2 · How AI Actually Works
AI Has Limitations (Hallucinations, Bias, Errors)
You ask an AI assistant to recommend a book on a subject you care about. Back comes a title, an author, a publisher, even a page count, confident and specific. You go to buy it, and the book does not exist. Not out of print. It was never written. The AI made it up.
Why This Matters
Every AI tool you have met in this pillar shares one flaw: it can be confidently wrong, and it will not warn you when it is. If you cannot recognise the failure, you will trust the wrong answer at the wrong moment. By the end of this lesson you will be able to:
- Recognise an AI hallucination and give two examples
- Recognise how bias enters AI outputs and give two more examples
- Recognise a confident error for what it is
This is the Ethics and Risk Awareness that keeps you in control.
Where AI Fails, and Why
The ordinary documents, emails, and spreadsheets around you are the raw material an AI learns from. That is the key to this final lesson: an AI is only as good as the patterns in that data, and patterns can mislead, in three ways you need to recognise.
- Hallucination: invents plausible facts that were never real.
- Bias: repeats the unfairness baked into its training data.
- Confident error: states a wrong answer in the same certain tone.
First, hallucination. An AI predicts the next likely words; it does not look facts up. When it has no real answer, it does not stop; it produces one that fits the pattern and states it as fact. That is how you get a citation for a court case that never happened, or a quote no one ever said.
Second, bias. An AI learns from human data, and human data carries human bias. Train a hiring tool on records that favoured one group, and it quietly favours that group too. Ask an image generator for a nurse, and it may return only women: the pattern in its training, not the truth.
Third, confident errors. An AI states everything in the same fluent, certain tone, right or wrong. It gets a simple sum wrong, or misreads the date on your invoice, and reports it with the confidence of a correct answer. The tone never signals doubt.
One cause sits under all three: an AI predicts patterns. It does not know what is true.
Key Takeaway
One cause sits under all three failure modes: an AI predicts patterns, and it does not know what is true.
Check your understanding
Self-check only. Nothing is scored and nothing is recorded; open an answer to compare it with your own.
Raj is a procurement officer. He asks an AI tool to summarise the safety record of a supplier he is evaluating. The AI returns a detailed paragraph citing two industry safety audits with specific dates and scores. The summary looks thorough and well-sourced. What should Raj do next?
- Use the summary as-is in his procurement report, the AI cited specific audits, so the information is reliable.
- Search for the two cited safety audits independently to confirm they exist and that the scores match what the AI stated.
- Reject the AI output entirely because AI cannot be trusted with factual claims about suppliers.
- Ask the AI to confirm whether its own citations are accurate.
Show the answer
Search for the two cited safety audits independently to confirm they exist and that the scores match what the AI stated.
Mei works in HR. She uses an AI tool to help shortlist CVs for a data analyst role. She notices the AI consistently ranks candidates from certain universities higher, even when candidates from other institutions have stronger relevant experience. Which AI limitation best explains this pattern?
- Hallucination, the AI is inventing qualifications for candidates from those universities.
- Confident error, the AI is wrong about which candidates are stronger but states its ranking with certainty.
- Bias, the AI learned from historical hiring data that favoured graduates of certain institutions and is reproducing that pattern.
- This is not an AI limitation, the AI is correctly identifying that those universities produce better candidates.
Show the answer
Bias, the AI learned from historical hiring data that favoured graduates of certain institutions and is reproducing that pattern.
Saved in this browser only. Carry progress between devices with the resume link on the course page.