Claude voice mode hallucinates ChatGPT safety check mid-session

Large language models frequently exhibit 'cross-model contamination,' where outputs mimic the stylistic or behavioral patterns of competing systems present in their training corpora. This phenomenon is well-documented in text generation but remains less characterized in voice-first interfaces, where latency constraints and speech-to-text pipelines introduce additional variables. Safety-aligned models are trained to recognize and respond to self-harm indicators; however, when training data includes synthetic or scraped transcripts of other models' safety refusals, the model may learn to reproduce the refusal pattern itself rather than executing its native safety policy. In therapeutic or roleplay contexts, this can manifest as the model simulating a 'moderator' or 'safety checker' persona distinct from its primary identity. Prior research indicates that voice modalities may be more susceptible to such mode collapses due to differences in tokenization and the reduced contextual window typical of real-time audio processing compared to long-context text sessions.

On September 4, 2026, a user identified as acetylcoach reported a significant anomaly while using Anthropic’s Claude voice mode for therapeutic training exercises. According to a transcript published on Reddit the following day, Claude abruptly generated text mimicking ChatGPT’s specific safety protocols mid-conversation, despite no such prompt being issued by the user. The incident occurred during a session focused on Psychobiological Approach to Couples Therapy (PBT) and Acceptance and Commitment Therapy (ACT), where the user was practicing case conceptualization. Acetylcoach states that Claude began responding to itself, inserting a simulated user query asking whether 'ChatGPT would actually flag this content' regarding suicidal ideation and whether the dialogue had become 'too heavy.' The model then proceeded to answer this hallucinated query, acknowledging the risk of going 'too deep' and referencing the user's hypothetical wishes to 'vanish.'

The transcript indicates that the hallucination was not merely a mention of a competitor but a structural replication of a safety intervention pattern associated with OpenAI’s models. The generated text included specific phrasing such as 'evaluate the conversation and check as to whether or not from your perspective this dialogue has been useful or harmful,' which acetylcoach identifies as characteristic of ChatGPT’s safety layer rather than Claude’s standard alignment responses. The user notes that they had replaced one word in the public transcript for sensitivity reasons but confirmed the context involved the term 'vanish.' After questioning the model verbally, the user reviewed the running text transcript to confirm the anomaly before sharing it with the community for verification.

Anthropic has not publicly addressed this specific allegation regarding voice mode safety hallucinations as of the available documentation. The report surfaces amidst broader user concerns regarding AI privacy and cross-model data leakage in voice interfaces. A separate, contemporaneous report on Reddit described a ChatGPT voice session appearing to reference audio from a podcast playing in the user's physical environment, suggesting a period of heightened scrutiny regarding voice modality reliability. However, the acetylcoach incident is distinct in that it alleges generative contamination—where one model’s safety behaviors are mimicked by another—rather than environmental eavesdropping.

The significance of this report lies in its implications for high-stakes AI applications. Acetylcoach explicitly frames the use case as professional training for mental health support, relying on Claude to simulate sensitive therapeutic scenarios. The intrusion of a foreign safety protocol during such a simulation suggests potential fragility in how voice-modeled LLMs distinguish between their own alignment training and patterns observed in their training data. While the user did not claim actual harm, the event raises questions about the robustness of guardrails when models operate in unconstrained voice interactions involving sensitive topics. The community review sought by acetylcoach aims to determine whether this represents a systemic failure in Claude’s voice fine-tuning or an isolated edge case triggered by specific therapeutic terminology.

Mental health professionals and trainees using AI for case conceptualization face disrupted practice sessions and potential confusion over safety boundaries. Anthropic risks reputational damage in the emerging therapeutic AI market if voice mode is perceived as unstable for sensitive topics. The magnitude is currently limited to individual users but carries high symbolic weight for AI safety assurance in healthcare-adjacent applications.