---
title: "Comment by Lazar Tomić on \"What Anthropic Actually Said About AI Building Itself\" | Hashnode"
description: "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 wron"
url: "https://hashnode.com/posts/anthropic-ai-recursive-self-improvement/6a23d700fba76f562a302663/comment/6a2451d08b5885e106f201ef"
---

[Discussion](https://hashnode.com/posts/anthropic-ai-recursive-self-improvement/6a23d700fba76f562a302663)

Comment [What Anthropic Actually Said About AI Building Itself](https://hashnode.com/posts/anthropic-ai-recursive-self-improvement/6a23d700fba76f562a302663)

[Lazar Tomić](https://hashnode.com/@lazartomic)

[Jun 6](https://hashnode.com/posts/anthropic-ai-recursive-self-improvement/6a23d700fba76f562a302663/comment/6a2451d08b5885e106f201ef)

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.

[Arjun Adhikari](https://hashnode.com/@arjunadhikari101)

[Jun 11](https://hashnode.com/posts/anthropic-ai-recursive-self-improvement/6a23d700fba76f562a302663/comment/6a2451d08b5885e106f201ef#reply-6a2af93fea57737d35603fc6)

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?