Original title: Vulnerabilities in Personalization: Assessing Health Privacy Risks in ChatGPT Logs and Memory
S M Mehedi Zaman, Md Mozammel Hoque · September 2026
People talk to chatbots like ChatGPT every day, often sharing private health details without realizing the risks. Researchers checked real chat records and ChatGPT's memory feature to see what happens to this sensitive data. When you chat, the artificial intelligence can quietly pull out details behind the scenes and store them as permanent notes about you. The study found that more than one in five conversations held personal health details, and some posed severe privacy risks with exact locations, real identities, and stigmatized conditions. Worst of all, the system saved almost all of these memory notes automatically without asking for permission. It even turned short-term symptoms into lasting diagnoses, making it far easier to identify users.
Imagine you tell a friend about a headache you had yesterday. Later, the friend makes a permanent note in their phone that says you 'have chronic headaches,' and shares that note with other people without telling you. You never asked them to remember it, and now it changes how everyone treats you, even though it was just one bad day. In this paper, ChatGPT does something similar: it turns temporary health comments into permanent memory entries without the user's permission.
This matters because millions of people use ChatGPT for everyday tasks, from asking about symptoms to discussing medical conditions. The study shows the system quietly saves personal health details in a permanent memory profile, often without the user's knowledge or consent. This creates real risks for privacy, especially with stigmatized conditions or when location can identify someone.
The study is based on conversations from only four countries and a specific set of users, so results may not generalize globally. It focuses on health disclosures in English and a few other languages, potentially missing culturally-specific disclosures. The analysis does not cover all possible privacy risks or verify whether memory entries actually lead to real-world harm.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
Vulnerabilities in Personalization: Assessing Health Privacy Risks in ChatGPT Logs and Memory — S M Mehedi Zaman, Md Mozammel Hoque
Abstract licence: CC0
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.