New self-assessment system reviews conversations in the background, helping identify harmful activity without interrupting users.
Artificial intelligence company Anthropic has introduced a new safety mechanism for its AI chatbot, Claude, to quietly monitor conversations for potential misuse. The feature, known as “self-assessment,” enables the chatbot to evaluate its own interactions after responding to users, adding an extra layer of security without affecting the user experience.
Unlike traditional safety systems that rely on a fixed set of rules, this new approach leverages Claude’s reasoning abilities. After generating a response, the AI is prompted internally to review the conversation and determine whether it may involve harmful, dangerous, or suspicious activity. It then generates a brief internal report that identifies the nature of the concern and assigns a priority level for human review, if necessary.
For users, the process remains completely invisible. No alerts, warnings, or interruptions occur during the conversation. Safe discussions proceed normally, while potentially problematic exchanges can be flagged for further examination behind the scenes.
According to Anthropic, one of the system’s key advantages is its efficiency. Safety checks often require substantial computing resources, making them expensive to run at scale. However, the company states that Claude’s self-assessment process imposes only a minimal computational burden, increasing overall costs by less than two per cent. This makes it significantly cheaper than many traditional moderation systems while still maintaining strong performance.
The company reports that the technology can identify most harmful interactions, particularly those deemed high-risk. More importantly, because it relies on the model’s broader understanding rather than predefined categories, it can detect emerging forms of misuse that may not have been present in the training data.
Anthropic says the feature has already demonstrated practical value. During its deployment, the system reportedly uncovered a large volume of illegal content linked to a single customer account, enabling the company to notify the appropriate authorities. Researchers also found that Claude could recognise signs of fraud and cyber-related abuse, even when examples of such activities were intentionally excluded from its training.
Despite its promising results, Anthropic acknowledges that self-assessment is not a standalone solution. The company emphasises that the technology is intended primarily for monitoring and review, not for instant intervention. Internal testing showed that determined attackers could still bypass the system in certain circumstances.
As a result, Anthropic continues to combine self-assessment with stronger real-time safety measures. Together, these measures aim to create a more secure AI environment while preserving a smooth, uninterrupted experience for everyday users.
