How AI gets things wrong
AI errors aren't random. They fall into recognisable types, and knowing them tells you exactly where to look.
Learning objectives
- Classify common types of AI error
- Explain why each type occurs
- Spot red flags in AI-generated text
- Learn from real-world verification failures
2.1 A taxonomy of AI errors
Knowing the common types of error tells you where to look.
| Error type | What it looks like | Where to look |
|---|---|---|
| Fabrication | Invented facts, statistics, quotes, people, case law or references | Every specific claim with a number, name, date or citation |
| Misattribution | A real source cited for something it doesn't say | Citations — open and read them |
| Contextual errors | Ignoring, misreading or mixing up the documents you provided | Compare summaries against the source |
| Outdated information | Old prices, laws, leaders or product details | Anything time-sensitive |
| Reasoning errors | Logical leaps, wrong maths, conclusions that don't follow | Calculations and "therefore" statements |
| Omissions | Leaving out important caveats, risks or counter-evidence | Ask: what would an expert add? |
| Overgeneralisation | "Always", "never", "all experts agree" | Absolute or sweeping claims |
| Inconsistency | Contradicting itself within or across answers | Compare sections; ask again differently |
| Sycophancy | Agreeing with your mistaken premise or flattering weak work | Answers that conveniently confirm your view |
| Bias | Stereotypes, one-sided perspectives, unequal treatment | Content about people and groups (Module 4) |
IBM groups hallucinations into factual, contextual and consistency types — a helpful shorthand for the first rows above.
2.2 Why these errors happen
As the How AI Works course explains, LLMs generate the most plausible continuation, not verified facts. OpenAI's 2025 research argues that training and evaluation often reward confident guessing over admitting uncertainty. Models also have a training cut-off, and post-training can reward agreeable answers.
2.3 Red flags in the text itself
- Very specific numbers with no source ("73.4% of UK SMEs…")
- References you can't immediately find, or that look slightly wrong (author, year or journal mismatches)
- Quotes attributed to named people
- Confident claims about recent events
- Neat, symmetrical lists that look too tidy for a messy topic
- An answer that perfectly matches the assumption in your question
Lawyers in the US, the UK and elsewhere have submitted court documents citing cases generated by AI that did not exist. In 2025 the High Court of England and Wales warned that lawyers who cite fake AI-generated cases could face serious consequences, and that freely available generative AI tools are not capable of reliable legal research. The tools weren't "lying" — they generated plausible-looking citations. The failure was the missing verification step.
Ask an AI tool for "five statistics about AI adoption in UK small businesses, with sources". For each statistic, find the original source. How many exist, are current and say exactly what was claimed?
Key takeaways
- Fabrication and misattribution are the most dangerous errors — check specifics and citations.
- Contextual errors misread the material you supplied.
- Watch for outdated facts, reasoning slips, omissions and absolute claims.
- Answers that echo your assumption may be sycophantic.