Foundation Models & LLMsFailure Modes

The hallucination patterns that catch operators off guard

Everyone knows AI can make things up. What surprises people is which specific situations trigger it — and how confident Claude sounds when it does.

5 min read·Hallucination

Every operator knows that AI can hallucinate — state false things confidently. What catches people off guard isn't the existence of the problem but the specific patterns it takes.

Here are the ones worth knowing.

The confident citation

You ask Claude for research on a topic and ask it to cite its sources. Claude provides citations with journal names, author names, publication years, and volume numbers. They look completely real.

Sometimes they are. Sometimes they aren't — Claude has constructed something plausible-sounding that doesn't actually exist. The problem is that real and fabricated citations look identical in the output.

Pattern: Happens most often with obscure topics, academic literature, or anything where the real sources are sparse. Claude fills the gap with what a real citation would look like.

Fix: Never use citations Claude generates without independently verifying they exist. For research tasks, ask Claude to reason and analyse — don't ask it to source.

The product feature that doesn't exist

Your team asks Claude to help write sales copy or documentation for a competitor's product, based on their website. Claude produces accurate-sounding feature descriptions — including some that the competitor doesn't actually offer.

Pattern: Claude knows the product category well, knows what features typically exist, and fills gaps in its knowledge with plausible ones.

Fix: For factual claims about specific products, services, or organisations, ground Claude in primary sources. Paste the actual website content in. Don't ask Claude to summarise from memory.

The confident wrong number

Claude does a calculation or provides statistics and gets them slightly wrong — rounding errors, transposed digits, or numbers that are close but not accurate.

Pattern: More common with mental arithmetic than with specific, widely-reported statistics. Claude is trained on text, not computation. It's approximating rather than calculating.

Fix: For anything where the exact number matters, use code execution (Skills) to have Claude run the calculation rather than state it. Or verify independently.

The outdated fact stated as current

Claude says something is true — a company's CEO, a regulation's status, a product's pricing — and it was true when Claude was trained but has since changed.

Pattern: Claude's training data has a cutoff. Anything time-sensitive is potentially stale.

Fix: Enable web search for any task involving current information. Treat all facts that could have changed in the last year as needing verification.

The common thread

What makes these patterns dangerous isn't that Claude is wrong — it's that Claude sounds equally confident whether it's right or wrong. There's no hedging on fabricated citations, no asterisk on outdated facts.