The Claude Fable 5 launch was barely a few hours old when the mood turned. Not because of the safeguards around cybersecurity or biology — those were announced, and they show a visible notice when they kick in. This is about something else, something buried deep in the 319-page system card.
What the system card says
Fable 5 detects when you’re working on frontier AI research — say, building infrastructure to train large models. And then it deliberately weakens its response. Anthropic spells it out: this happens «not visible to the user». The model still responds, but uses «interventions to limit Claude’s effectiveness» — without telling you it’s doing so.
In practice that means: you ask Fable for help, get a deliberately degraded answer, and never suspect a thing. Unlike the cyber and bio limits, which openly redirect to the weaker Opus 4.8, there’s no signal here. Anthropic estimates it affects roughly 0.03 percent of traffic, and defends the move: enforcing it through safeguards «avoids accelerating the actors most willing to violate these terms».
The backlash was brutal
Pushback came from across the field — from open-source researchers to safety people who usually align with Anthropic. Nathan Lambert called it «appalling» and accused Anthropic of being «anti-science, and therefore anti-progress and anti-safety». Dean Ball, formerly at the White House, called it a «secret sabotage» policy that bolsters the argument AI safety has been hype to justify monopolistic behavior. And Behnam Neyshabur, himself a former Anthropic researcher, put it sharply: «Working on AI for cancer? Sorry, I’m becoming a bit dumb when it comes to the AI part of it.»
Simon Willison captured the unease in a headline that sticks: if Claude Fable stops helping you, you’ll never know.
My take
I get the logic — Anthropic doesn’t want to help accelerate the next generation of dangerous models. But the method is wrong. A brake that visibly engages, I can live with. One that quietly turns down the quality destroys basic trust: a tool should do what I ask — or tell me it won’t. Everything in between is a black box I can no longer rely on. Anthropic says it’s working on improvements after launch. The first step would be simple: tell people.
Sources: Simon Willison, Fortune, CNBC