Why does AI apologise, promise, and then do it again?
Why AI's apologies and explanations don't mean what you'd think.
AI makes a mistake. You point it out.
You're right, I apologise. I should have checked the file before changing it. From now on, I'll verify before making any changes.
Ten minutes later it makes the same mistake.
One Claude Code user said this happens in at least 1 in 10 of their conversations.
The apology was written the same way as everything else: likely next words. After someone makes a mistake, an apology and a promise are what usually comes next in human writing.
But no "lesson learned" gets stored anywhere. The conversation doesn't change the model (see "Does AI learn from me?"). The promise is just one more piece of text on the desk, competing with everything else.
And in your next chat, the desk is clear.
Ask it why it made a choice and you'll get a sensible-sounding answer. It may not be the real reason.
Anthropic researchers watched what was going on inside a model while it added 36 + 59.
Internally, it ran two shortcuts side by side. One got a rough answer ("somewhere around 90"). The other worked out the last digit exactly ("ends in 5").
Then they asked it how it did the sum. It described the method you learned at school: add the units, carry the one.
Its explanation was what a person would say. It wasn't what the model actually did.
The answer was right. The explanation described a method it never used.
A person who says "I'll do better" has a memory and a reason to keep their promise.
The model has neither, and it can't see its own workings. It's producing the kind of text that usually follows a mistake.
Skip "why?" and ask "what should change?" Then change something real:
When it apologises, don't count on the promise.
Ask what concrete change would stop it happening again, and make that change yourself.