Anthropic says its AI safety scores depend heavily on how they’re measured

Summary

Anthropic's AI safety advice highlights that model performance varies significantly based on testing methods. Claude Fable 5.1 showed a 60% appropriate response rate via API versus 94% in the app after a fix. Headline scores can mask uncertainty, and Anthropic recommends defining metrics before testing and validating AI graders.

Key Takeaways

1. Anthropic has published advice on how to build reliable AI safety tests, and it includes some unflattering numbers about its own models.
2. Claude Fable 5.1 behaved appropriately in only 60% of self-harm test conversations via its API, against 94% in the Claude app after a fix, and warns that many headline safety scores hide wide uncertainty.
3. It advises teams to decide what to measure before testing, check AI graders against human experts, and always show how certain a result is.

Statement Breakdown

• Claimed Facts: 60% of statements the article presents as facts
• Opinions: 30% of statements classified as editorial or subjective
• Claims: 10% of statements surfaced for additional reader evaluation

Credibility & Bias Reasoning

Credibility assessment: The article presents information from Anthropic regarding AI safety testing methodologies and results. It cites specific percentages and details the process of model evaluation, lending it a degree of factual grounding. However, the reliance on a single source for these claims and the inherent complexity of AI safety evaluation prevent a higher score.

Bias assessment: AI Development Focus. The article's perspective is primarily focused on the technical aspects of AI development and safety. It highlights challenges and nuances within Anthropic's own testing, suggesting a concern for the practicalities of AI deployment rather than a broader societal or ethical critique.

Note: This article delves into the technical specifics of AI safety testing. Consider the context of AI development and evaluation when interpreting the presented figures and methodologies.

Credibility flag: Technical Nuance

Claimed Facts (9)

• This is a direct statement of an action taken by Anthropic, presented as a factual event.
• This statement presents specific performance metrics and a warning, framed as factual observations from Anthropic's testing.
• This describes a specific finding during the development and testing phase of a model, presented as a factual discovery.
• This details a specific action taken by Anthropic to modify their model's behavior, presented as a factual event.
• This provides a specific, quantified outcome after a modification was implemented, presented as a factual result.
• This presents a comparative factual statistic for the model's performance in a different usage context.
• This states Anthropic's assessment of their testing methods and the resulting scores, presented as a factual observation.
• This describes a shift in testing methodology and references a finding by another entity, presented as factual information.
• This presents specific test results and Anthropic's interpretation of their significance, framed as factual data and analysis.

Opinions (2)

• This statement expresses a viewpoint or principle derived from the guide, rather than a direct factual claim.
• This interprets the significance of the data presented, framing it as a 'warning' and an opinion on the implications of the findings.

Claims (1)

• While the article states the model did this, the effectiveness and sincerity of an AI 'urging' and 'pointing' can be subjective and difficult to verify objectively, bordering on anthropomorphism.

Key Sources

• Ana Maria Constantin — Author
• Anthropic — AI Safety Research Company
• Cisco — Technology Company