Editing Why AI Gets Things Wrong (section)

== When it's most likely to go wrong ==

Hallucinations aren't random. They follow predictable patterns:

'''Specific facts, numbers, and dates.''' Ask AI for a general explanation of how photosynthesis works and it'll be accurate. Ask it for the exact year a specific obscure paper was published and it might invent one. The more specific and verifiable the claim, the more you need to check it.

'''Citations and sources.''' AI is particularly bad at this. It will confidently produce author names, paper titles, journal names, and URLs that look real but don't exist. Never trust an AI-generated citation without verifying it.

'''Recent events.''' AI models have a training cutoff date. If you ask about something that happened after that date and the AI doesn't have search access, it may either say it doesn't know (good) or generate a plausible-sounding answer (dangerous).

'''Niche or specialized domains.''' AI performs best on topics that appeared frequently in its training data. Mainstream topics in English have dense coverage. Obscure or specialized topics — especially in other languages — have less, so the model has fewer patterns to draw from and is more likely to fill gaps with plausible-sounding fabrications.

'''When you push it.''' If you insist the AI answer a question it's uncertain about, or tell it "you must provide an answer," it will comply — by generating something. AI tools generally don't have a strong instinct to say "I don't know." Some are better than others, but the pressure to produce output is built into the system.