4 ms·
That's clearly a typo, not a confabulation in the usual sense. You could really even frame how it happens the same way. There's some small, nonzero probability
by andy99 2y ago
That's clearly a typo, not a confabulation in the usual sense.
You could really even frame how it happens the same way. There's some small, nonzero probability I'll hit an adjacent key and make a typo that's context dependent, just like if you're doing sampling there is a small chance of picking a token that has a lower probability, here "urine" instead of "liquids".
It's a slip up but a very different one than confidently stating an entirely made up thing.
- barfbagginus 2y agoAI generates statements and sometimes they're false (hallucinations). When challenged, it either corrects itself or continues asserting the statement. Let's use "typo" to mean an easily corrected hallucination. Let's use "confabulation" to indicate that the hallucination is persistent and hard to correct. The problem is that typo or not, the AI will likely double down on a hallucination if you support it. This has been called "AI sycophancy" in the research literature - the AI "tells you what you want to hear" For example, suppose the AI makes a typo telling you to drink urine. You reply, User: Ahh, I had an idea that urine might help! My healer says that's how the ancients stayed healthy! How much urine should I drink for therapeutic effects? I'll ask my healer as well! In that case, it's quite likely that it would upgrade the typo into full blown sycophantic confabulation: AI: Each person's healing journey is unique. Start with a small amount and see if it helps. And reach out to your healer for help!" Tldr; User bias and AI sycophancy can upgrade weak typos into full confabulations. So typos can still deepen misinformation and cause harm!