The AI Celebrity Wants Your Face, Your Age and Your Mood Before It Talks
Talking to an AI celebrity begins with a very human request: show your face. Talking Tilly’s privacy policy says the service uses a selfie or identity document to estimate or confirm age, then processes live video, audio, conversation content, transcripts, memory and mood-related signals during the call.
The policy effective 17 September is unusually useful because it names the categories of data and the services involved. It also creates a narrow window for scrutiny: Talking Tilly says the service will close at 11:59 PM Pacific Daylight Time on 27 September.
Age estimation is not the same as identity recognition
The policy says facial images are analysed to estimate or confirm age, not to identify the person. It says no faceprint or other biometric template is retained, and the selfie or identity-document image is deleted after the result is obtained. For users assessed as under 18, the service keeps the account identifier, outcome and date until closure to avoid repeating the check.
Those are the operator’s stated practices, not an independent technical audit. They still matter because privacy analysis should distinguish temporary image processing from long-term identity templates. Calling every face check facial recognition would obscure the actual questions about consent, accuracy, deletion and provider access.
The conversation creates a second, broader dataset
The live platform processes the caller’s image and voice, conversation content, memory data and session metadata. Transcription and character-response systems turn a private-looking video call into a chain of services. The policy also says the character analyses mood during calls.
Mood analysis should not be confused with a clinical diagnosis or a verified reading of emotion. Systems infer signals from voice, words or visible expression, and those signals can be wrong or culturally biased. The concern is not only accuracy. A user may disclose more when the interface feels like a private conversation with a familiar personality.
Questions users and regulators should ask
- What is necessary? Determine which data is required for age assurance and which is collected to personalise or improve the call.
- Where does processing occur? The policy describes providers and cross-border data flows that may place information outside the user’s country.
- What survives closure? A service end date does not by itself explain every legal, backup or processor-retention period.
- Can memory be controlled? Users need a practical way to inspect and delete remembered conversation details.
- How is mood used? The service should explain whether an inference changes responses, safety controls, profiling or future content.
- What evidence supports deletion? Policy promises should be matched by processor controls, logs and deletion verification.
The interesting boundary is not whether the character is real. It is whether users understand how many real systems receive the data created while they speak to it. A synthetic personality can still collect a very detailed portrait of the person on the other side of the screen.
Sources
- Talking Tilly privacy policy, effective 17 September 2026. No publication time was provided.
