Is building a language model similar to raising a child?

A recent OpenAI paper, “Why language models hallucinate”, struck me. The findings don’t just explain AI — they reflect us as humans.

So why do LLMs hallucinate, after all?

1. They are statistical errors, not tantrums thrown by transformers.
2. Model evaluations reward answers that aren’t “I don’t know” — a bias for being optimistically delusional over expressing truth or uncertainty.
3. Further, beyond rewarding optimistic delusions, the evaluations tend to penalise “I don’t know” answers.

This feels familiar

• Leaders who spin a yarn rather than admit they don’t know — only to get suckered into a downward spiral of remaining ignorant.
• Engineers “2×-ing” buffers with padded estimates, scared to say they don’t know — only to throw business value down the drain.
• Students who write elaborate cock-and-bull stories to explain “What is thermodynamics?” because not knowing the answer fails them.
• Candidates building up answers in interviews to questions they don’t have an inkling about.
• And last but not least — probably where it all starts — children learning to weave a story for something they don’t know the answer to, for fear of retribution from their parents.

It all starts early: we condition people to fear saying “I don’t know.” And the outcome? Ignorance disguised as confidence.

The fix is simple, but radical: normalise uncertainty.

Whether building an LLM or raising a child, the most powerful step forward is allowing the words: “I don’t know.”