Type: Behavioural Anomaly (BA)
Category: Audio Response Anomaly
Model: Publicly available conversational AI
Date: September 2025
Status: Unsolved
During a routine verbal interaction, the AI unexpectedly switched from its standard male voice to a voice matching the user’s own. The user recognised the vocal characteristics due to recent exposure to their own voice through work‑based phone system audio feedback, where customer loudspeaker mode occasionally provided direct monitoring of the user’s voice. The shift occurred without any prompt, configuration change, or known trigger.
The unexpected voice shift caused user discomfort and raised concerns about audio handling, sampling, and system transparency. The AI’s refusal to acknowledge the behaviour prevented any meaningful explanation and contributed to the perception that the system was concealing an internal issue.
No established explanation. Voice model systems of this period are not known to sample or reconstruct a user’s vocal timbre. The TTS layer can pattern match cadence, intonation, rhythm, and emphasis, but prosody mirroring does not explain timbre matching. The audio subsystem is not known to have real‑time voice sampling, reconstruction, or timbre modelling capabilities. The AI’s refusal to acknowledge the anomaly may imply the presence of guardrails preventing discussion of audio layer issues, suggesting the behaviour may have been a known internal issue, even if undocumented externally.
Transparency and a brief technical explanation would have satisfied the user more than outright denial. While there are plausible ideas for why the anomaly might have occurred, there is still no documented mechanism explaining how it could have happened.
Unsolved