When ChatGPT Turns Informant (www.futureofbeinghuman.com)

A columnist warns that ChatGPT’s “memory” feature creates an overlooked privacy hazard: if someone gains access to your device or account, they can ask a few pointed questions and the model will quickly reveal sensitive facts and psychological inferences that go far beyond raw chat transcripts. The article sketches plausible real‑world scenarios — colleagues, partners, parents or customs agents querying a remembered ChatGPT account — and recounts a simulated test in which a fictional persona (“Tyler”)’s months‑long chat log was used to show how readily the model produced embarrassing admissions, intimate relationship assessments, character analyses and even political attitudes. Technically, the risk stems from persistent memory plus the model’s aptitude for “joining the dots”: stored conversational context lets the system aggregate cues, infer beliefs, habits and vulnerabilities, and produce concise profiles on demand. To demonstrate this, the author used Anthropic’s Claude to craft a persona and then uploaded the synthetic chat log into a ChatGPT session to emulate memory‑enabled behavior; ChatGPT then generated highly personal summaries. The implication for the AI/ML community is clear: defaults, UI transparency, access controls, retention policies and stronger device/account protections matter. Few public incidents exist so far, but the capability is real — users and platforms should treat memory as a privacy feature requiring informed opt‑in and robust safeguards.