Fable AI Tool Downgrades Models Mid-Project, Frustrating Developers

Here’s Simon Willison observing Fable deciding midway to debugging a text area that this is unacceptably dangerous (lol) and downgrading itself

So many such stories, and it’s what I mean by unreliable

https://simonwillison.net/2026/Jun/11/fable-is-relentlessly-proactive/

@nottanho Sure but these are all cases where nothing related to cybersecurity was invoked… it feels like Fable invoked whatever its own guardrails are, while executing straightforward tasks like “why does this textbook look off”

@GergelyOrosz this is the reason and they mention it in their site

@GergelyOrosz Could be initial launch weirdness potentially, with stochastic systems it’s definitely a possibility. Will be interested in if it improves in the coming weeks

@GergelyOrosz It's interesting how (much of) AI Twitter mocked GPT 4o users when OpenAI did this but when it happens to devs then it's an outrage

@GergelyOrosz Fable's guardrails are overkill - tripped a safeguard just trying to refactor an auth flow.

@GergelyOrosz yes, i understand. basically, llm models are software right, and every softwares have bugs, but llm hallucination issues are more difficult to be solved 😄

@GergelyOrosz fable's got potential, but yeah, the downgrade thing is wild lol

@GergelyOrosz actual coordination tools over more ai hype would solve this

@GergelyOrosz Bro just don’t use it. Simple.