Hi Lazar, this is a brilliant angle to look at it from.
I think the line between a hallucination and a hypothesis comes down to two things: established reality and testability.
When a model gets facts wrong about what is already known such as inventing a fake historical date or a broken URL then it’s a hallucination. It simply failed to retrieve the truth. But when it ventures into uncharted territory like proposing a completely new biochemical compound to fight a novel virus then it is explicitly stepping outside its training data to synthesize a pattern humans might have missed.
So, does a hallucination only graduate into a true hypothesis when we introduce testability?
Well if an AI generates a wild, unsupported claim about the unknown, it feels like an error. But maybe it becomes a legitimate hypothesis the moment a feedback loop can actually run an experiment to prove it right or wrong. Maybe we only call it a "hallucination" because we don't yet have a fast enough way to test if its weirdest ideas might actually work.
What do you think? Does the hypothesis require a human to recognize its value first, or is the ability to test it enough?
Lazar Tomić
GSI Darmstadt research intern passionate about Machine Learning, Artificial Intelligence. Student + Author
Let’s discuss:
Would you call an AI hallucination a hypothesis?
We dismiss hallucinations as errors. But history is filled with ideas that initially looked wrong, absurd, or impossible.
If an AI generates a novel claim unsupported by its training data, is it merely failing or is it doing something closer to what humans call speculation?
At what point does a hallucination become a hypothesis worth investigating?
Or are we giving machines far too much credit?
I'm curious where you draw the line.