Unpopular opinion: most "AI bugs" are really input-trust bugs wearing an AI costume.

The model didn't hallucinate. Someone fed it text they never sanitized, from a source they didn't control, and acted on the output without a check. That's the oldest problem in software, trusting your input, dressed up in new clothes.

I build every AI feature from the input inward. A scraped job posting is adversarial. A webhook claiming a payment settled is adversarial until I've verified it against the ledger. The model's output gets validated against a schema before anything downstream touches it, the same way you'd validate a form.

The teams shipping reliable AI usually don't have the fanciest prompts. They have the boring habit of treating every boundary as hostile, even after a language model joined the stack.

If your AI features get weird in production but work in the demo, this is usually why. The demo used clean input. Production never does.

Agree? Or do you think I'm underselling the model's own failure modes? Let's argue in the comments, I'll defend it.