Claude, Anthropic's AI chatbot, has been quietly grading its own conversations for harm since June 2026, the company disclosed this week in posts on X. The system is called self-assessment, and it runs in the background on models including Claude Fable 5 without the user ever seeing it.
Here's how it works. After Claude answers a prompt, a hidden instruction asks the model to review the exchange it just had. Claude then produces a short internal verdict: whether the request looked harmful, what kind of harm it might be, and a priority score for a human reviewer to check later. If the chat looks clean, it's marked unflagged and nothing happens. No pop-up, no warning, no interruption.
What makes this notable is the cost. Anthropic says the check adds roughly 10 extra tokens per conversation, pushing total compute cost up by less than 2 percent. The company's own older safety systems added close to 24 percent, and Anthropic says rival safety tooling from OpenAI has consumed up to 16 percent of compute in comparable setups.
Accuracy and limits
On Claude Fable 5, Anthropic says self-assessment catches 78 percent of harmful conversations and 94 percent of the high-priority ones, with a false-positive rate of just 1 percent. Because the tool leans on Claude's general reasoning rather than a fixed rulebook, Anthropic says it can pick up misuse patterns it was never explicitly trained on. In one internal test, researchers hid examples of fraud and cyberattack attempts from training data entirely, and the model still caught 90 percent of them.
The company is also open about where the tool falls short. Self-assessment is built to flag conversations for offline review, not to block anything in real time, and Anthropic says it isn't enough on its own to stop a determined attacker. In internal red-team testing, hackers got past Claude 54 percent of the time when self-assessment was the only defence running, against a 79 percent success rate with no safeguard at all. A dedicated real-time blocking layer cut that down to 25 percent, which is why Anthropic says it's running self-assessment alongside, not instead of, its existing real-time filters.
The system has already produced results outside the lab. Anthropic says that in August alone, self-assessment flagged more than 100,000 cases tied to a single customer account involving child exploitation material, which the company says it reported to authorities immediately. Anthropic hasn't said whether this self-assessment layer will be extended to other Claude models beyond Fable 5, or whether details of its flagging thresholds will be published for outside audit. For a company that has spent much of this year pushing California lawmakers to let courts intervene on risky AI models before release, a cheap, always-on internal monitor is a logical next step, and likely not the last one.
ADFiled by The AI Desk
Models, assistants and the companies and chips behind them, reported from what was released and what was claimed, with the difference kept clear.
More from this desk →
Be the first to comment
Join the argument. No password, just your email or a passkey.