Anthropic's Claude Fable 5 launch introduced separate safety classifiers that detect potential misuse in cybersecurity, biology, and model-distillation attempts, automatically routing flagged queries to an older Opus model instead of refusing outright. This setup paused chats for AI safety researcher David Manheim when he queried topics such as AI control protocols, stochastic games, evolutionary dynamics, and formal verification of alignment, even though the content involved legitimate arXiv papers rather than harmful applications.
I am curious if the new Fable safegaurds would trigger accidentally on meaningfully important work as a false positive that has a real life consequence.
Just realizing this is eroding trust and please change this to just down right blocking which is a fine position to take. Respect the user’s time.
The classifiers triggered on queries touching high-risk domains even when the actual intent centered on theoretical safety research, producing the kind of false positives Rohan Anil warned could erode user trust over time.
Anil noted that routing to a weaker model wastes researcher effort without adding safety value and suggested outright blocks would at least respect users' schedules; exact false-positive rates for alignment topics remain unstated beyond Anthropic's early aggregate claim of under 5 percent triggers overall.
Users criticized Fable's safety safeguards for eroding trust through false positives and opaque decisions that block alignment research while enabling monopoly control and overreach.
@far__el assuming you even believe the 0.03 figure. it seems to be a hair trigger, gotta be a lot higher
Everyone should be as vocal as possible about this unethical Fable safetyist restrictions on AI research TODAY. Even if it doesn’t impact you and it’s only targeting 0.03% of traffic, this is setting a precedent that will later impact most if not all usage. This goes beyond censorship, this is control and gaslighting like we have never seen before. Super dystopian.
@far__el feels like the old slippery slope argument but this time it might actually hold water
whats the endgame here exactly?
@luminxbt There is no logic to this argument, simply refuse my request if you are safetyist. But to sabotage? That is unacceptable
@_arohan_ Claude refuses to work on my master thesis on model alignment. Either fable 5 straight up refuses or it switches to a dummer model, which then also refuses
@far__el It’s the modern day, “Don’t read this book or it may give you bad ideas.”
@max_paperclips @far__el You know they claimed only a low percent of fable conversations get rerouted but I can't even say hi cause the Claude.md mentions cybersecurity. So what ever bullshit figure they give you is likely OOM off
@_arohan_ I would bot trust them if i they where not my costumer, you truly dont know how fare that goes, corupting tools compaditors use.
@far__el your lab just got deprecated LOL
@_arohan_ Anthropic wants to decide what's good for you and what's safe. They want a monopoly on AI so they can make all the money and all the decisions. They love the idea of locking out anybody they consider dangerous. Imagine if the internet or electricity was handled like this.
@_arohan_
@_arohan_ so ur saying the safeguard is worse than just blocking outright because at least blocking is honest about it
@far__el Vocal is good. Cancelling your subscriptions and crippling them financially is better. Anthropic is already at the stage where they should no longer be an organization. They should be stopped. They’re a malicious actor and supply chain risk
@_arohan_ false positives in moderation always kill trust faster than the problem theyre solving
opaque decisions are the real issue here
@_arohan_ "Trust erodes once and rarely rebuilds"
making the distinction between nsfw censorship vs functional safeguards feels like a pretty reasonable ask
@_arohan_ over-cautious moderation always hits the wrong things first
why is letting people block worse than breaking actual work?
@_arohan_ this is actually a solid point. the vague scoring thing makes it worse no one knows why they got flagged.
@far__el the 0.03% now is the whole door later
they always frame it as tiny until its everyone
@far__el a little traffic now, big crackdown later. thats how every overreach starts
Anthropic's Claude Fable 5 launch introduced separate safety classifiers that detect potential misuse in cybersecurity, biology, and model-distillation attempts, automatically routing flagged queries to an older Opus model instead of refusing outright. This setup paused chats for AI safety researcher David Manheim when he queried topics such as AI control protocols, stochastic games, evolutionary dynamics, and formal verification of alignment, even though the content involved legitimate arXiv papers rather than harmful applications.
I am curious if the new Fable safegaurds would trigger accidentally on meaningfully important work as a false positive that has a real life consequence.
Just realizing this is eroding trust and please change this to just down right blocking which is a fine position to take. Respect the user’s time.
