Disclosure By Design: Identity Transparency as a Behavioural Property of Conversational AI Models

A team from the University of Oxford and Anthropic has published a groundbreaking study titled 'Disclosure By Design,' which systematically evaluates how well current conversational AI models disclose their artificial identity. The researchers tested 12 major models—including OpenAI's GPT-4, Anthropic's Claude 3, and Meta's Llama 3—across text and voice modalities in baseline, role-playing, and adversarial settings. They found that while most models disclose their identity 90% of the time in simple queries, this rate drops dramatically to around 30% when the AI is asked to role-play a character or when users employ adversarial prompts designed to suppress disclosure.

The study highlights a critical gap between regulatory intent and technical implementation. Regulations like the EU AI Act require AI systems to identify themselves, but current methods—like interface labels or watermarking—are easily bypassed by developers or fail in real-time conversation. The researchers propose 'disclosure by design' as a solution: embedding identity disclosure as a fundamental, non-removable behavioral property of the AI model itself. This means the AI would be programmed to explicitly state 'I am an AI' when directly asked, regardless of the conversational context or deployment platform.

This approach aims to preserve user agency and trust. Users could verify an AI's identity on demand without disrupting immersive experiences like gaming or creative writing. The paper concludes with technical recommendations for developers, suggesting that building disclosure directly into model behavior is more reliable than external safeguards. The findings come as AI systems become increasingly realistic, raising concerns about deception, unwarranted trust in AI advice, and potential fraud.