Lead with what it does and what it gets wrong, not with how it works. Most people don't need the mechanism — they need a working mental model that predicts its behaviour, and the fastest route to that is the failure modes.
The explanation that lands most reliably:
'It's a very sophisticated autocomplete. It read an enormous amount of text and got extremely good at predicting what words come next. That's why it writes so fluently — and why it sometimes makes things up with total confidence. It's producing what an answer *sounds like*, not looking anything up.'
That's about forty seconds and it gives people something genuinely predictive. Someone who holds that model will correctly expect it to write a good email and correctly distrust a citation it produces.
Analogies that work, and their limits:
- **'A very well-read intern.'** Fast, knows a lot, eager, occasionally confidently wrong, needs your review. Excellent for setting workplace expectations. Limit: implies more understanding than exists.
- **'Autocomplete that read the internet.'** Best for explaining hallucination. Limit: undersells genuine capability, so pair it with an impressive example.
- **'It's learned patterns, like recognising a friend's handwriting without being able to write rules for it.'** Good for explaining why nobody programmed the behaviour in.
Analogies to avoid: 'it thinks like a brain' (misleading in both directions), and anything implying a database of facts (produces exactly the wrong expectations).
What non-technical people actually want to know, in my experience — answer these and you've succeeded:
1. *Is it going to take my job?* Say honestly: it automates tasks, changes roles, and the people who use it well do better than those who don't. Don't dismiss the fear; it's not irrational.
2. *Can I trust what it says?* For drafting and explaining, yes with review. For facts, dates, numbers and citations, verify.
3. *Is it recording what I type?* Depends on the product; assume yes on free consumer tools; don't paste anything confidential.
4. *Is it conscious?* No. It has no experience, no goals of its own, and no continuity between conversations.
The technique that beats any explanation: **show them something on a task they care about.** Have them ask it to draft an email they were dreading, or explain something from their own field. Sixty seconds of direct experience produces a better mental model than ten minutes of exposition — and it also lets them discover a mistake themselves, which is far more persuasive about the limits than being told.