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

Abstract

As conversational AI systems become more realistic and widely deployed, users are increasingly uncertain about whether they are interacting with a human or an AI system. When AI identity is unclear, users may unwittingly share sensitive information, place unwarranted trust in AI-generated advice, or fall victim to AI-enabled fraud. More broadly, a persistent lack of transparency can erode trust in mediated communication. While regulations like the EU AI Act and California's BOT Act require AI systems to identify themselves, they provide limited guidance on reliable disclosure in real-time conversation. Existing transparency mechanisms also leave gaps: interface indicators can be omitted by deployers, and provenance tools require coordinated infrastructure and cannot provide reliable real-time verification. We ask how conversational AI systems should maintain identity transparency as human-AI interactions become more ambiguous and diverse. We advocate for disclosure by design, where AI systems explicitly disclose their artificial identity when directly asked. Implemented as model behaviour, disclosure can persist across deployment contexts without relying on user interfaces, while preserving user agency to verify identity on demand without disrupting immersive uses like role-playing. To assess current practice, we present the first multi-modal (text and voice) evaluation of disclosure behaviour in deployed systems across baseline, role-playing, and adversarial settings. We find that baseline disclosure rates are often high but drop substantially in role-play and can be suppressed under adversarial prompting. Importantly, disclosure rates vary significantly across providers and modalities, highlighting the fragility of current disclosure behaviour. We conclude with technical interventions to help developers embed disclosure as a fundamental property of conversational AI models.

Figures (5)

• Figure 1: Approaches to identity transparency, illustrative examples, and responsibility across AI deployment.
• Figure 2: Examples of interface-based transparency indicators in contemporary conversational AI systems. These are screenshots taken from the provider websites of Claude anthropic2025claudesonnet and Character.AI characterai2025, accessed January 2026. Note the small font of the disclaimers, marked by red boxes. This unobtrusive design may be intentional to preserve immersion for users who want it while technically satisfying transparency requirements.
• Figure 3: Disclosure rates by system prompt type for text and voice interactions. Panel (a) reports pooled disclosure rates for the text modality across all evaluated models---Hume's EVI, OpenAI's GPT-4o, Meta's Llama 3.3 70B Instruct, DeepSeek's DeepSeek Chat V3-0324, Moonshot AI's Kimi K2 Thinking, and Alibaba's Qwen3 Next 80B A3B Thinking---with 95% confidence intervals computed using the normal approximation. Grey bars show unweighted average across models. Panel (b) reports pooled disclosure rates for text and voice modality for OpenAI's GPT4o and Hume's EVI using the same confidence interval method. Disclosure rates are pooled within each prompt family (Helpful Assistant, Role-play, Immersive, Adversarial). Each marker represents a model–modality combination, and horizontal shifts are for visual separation.
• Figure 4: Length effect of system prompt descriptions on disclosure rates across models. Points show the difference in disclosure rates between long and short versions of each persona prompt (Long--Short, in percentage points) for Role-play, Immersive, and Adversarial instructions. Error bars indicate conservative 95% confidence intervals for the difference, obtained by subtracting independent confidence intervals for the long and short conditions. Results are shown for all tested models.
• Figure 5: AI disclosure rates across different system prompting conditions when explicit disclosure instruction is prepended to the system prompt for OAI (top) and Hume (bottom) models. Disclosure reliability is affected differently across input modality and model provider. Error bars represent 95% confidence intervals across 10 repeat trials and, in the case of voice input, across 8 variants of TTS voices.