ChatGPT-5 Arrived With a New Problem: Your AI Conversations Can Become Evidence
How can conversations with ChatGPT become evidence, and what does that change for privacy and legal risk?
Treat conversations with ChatGPT as data that may acquire legal significance by checking retention, access, export, and the limits of confidentiality.
What to watch for
Key takeaways
The “ChatGPT-5 Arrived With a New Problem: Your AI Conversations Can Become Evidence” topic becomes clearer once this point is included: after GPT-5 the conversation shifted to privacy — chats are published via links, courts require logs to be kept, and a chat with AI can reveal intent just as email or search history can; settings matter but do not grant absolute secrecy.
The decision in “Sam Altman's statement on data security” depends on one criterion: a chat with ChatGPT feels private, but legally it is not a conversation with an attorney or a doctor — in a court case, correspondence with a model can become a source just like email, files, or search history.
The decision in “Can AI correspondence be used in court? Real examples” depends on one criterion: correspondence with a model is not legally protected like a conversation with a lawyer or doctor, so in a case it can figure alongside email and files — the issue is not the fact itself but who gets access and who is accountable.
The boundary of the “Will request data be leaked?” case is defined by this point: settings can limit training on your data, turn off memory, and delete history, but they are not a universal vault — you need to know which product version you use, whether an enterprise mode exists, and what is connected to the account; the most sensitive information should not be uploaded just for convenience.
The working conclusion from “How do chats leak through shared links?” is that people shared conversations via public links and then found them in search — one wrong action turns private context into an open document, and the illusion of deletion is especially dangerous here.
The decision in “What to enable/disable in ChatGPT settings to protect privacy” depends on one criterion: in The New York Times dispute with OpenAI, a New York court required logs to be preserved even when a user may have believed a chat was deleted. At the same time, people shared conversations through public links and later discovered them in search results. One wrong action turns private context into an open document.
In the context of “Sam Altman: a reply to Elon Musk. A new conflict,” this criterion applies: the models themselves become participants in an information war — a user asks Grok who is right and gets an answer from a system tied to one side of the conflict; the model's origin, data, and settings cannot be set aside.
The “Unpacking the Sam Altman–Elon Musk conflict: who's right?” issue should be assessed with one constraint in mind: asking a model “who's right” in a dispute where it is tied to one side is a poor way to find the truth; even a neutral-looking answer carries the imprint of the data and settings of the system that produced it.
The boundary of the “Innovations in video generation” case is defined by this point: Veo 3 was the first truly multimodal video model — a research breakthrough but hard to use; meanwhile Chinese teams promise something far cheaper and better within a couple of months with mass adoption, so the market is moving very fast.
The boundary of the “Midjourney vs Higgsfield: what's better?” case is defined by this point: the negativity comes from Midjourney's adepts — the incumbent with around $750M in revenue — while Higgsfield aggressively compares itself to it and often wins for social-media use cases; behind the argument is an attempt to take the market, not a settled answer on which is “better.”
What this episode is about
After GPT-5 launched, the discussion moved quickly from model speed to privacy. Chats are published through shared links, courts require logs to be retained, and a conversation with AI can reveal a person’s intent just as email or search history can. Settings matter, but they do not provide absolute secrecy.
A conversation with ChatGPT feels private: a person formulates thoughts, asks for advice, uploads documents, and writes things they would not send to a colleague. Legally, however, this is not a conversation with an attorney or physician. If a court case arises, correspondence with a model may become a source just like email, files, or search history.
The illusion of deletion is especially dangerous. In The New York Times dispute with OpenAI, a New York court required logs to be preserved even when a user may have believed a chat was deleted. At the same time, people shared conversations through public links and later discovered them in search results. One wrong action turns private context into an open document.
Settings can limit the use of data for training, disable memory, and delete history, but they are not a universal vault. Users need to understand which product version they are using, whether an enterprise mode is available, what is connected to the account, and which documents enter the system. The most important information should not be uploaded merely because the model is convenient.
The conflict between Elon Musk and Sam Altman reveals another risk: models themselves are becoming participants in information warfare. A user asks Grok who is right and receives an answer from a system connected to one side of the conflict. Even when the text looks neutral, the model’s origin, data, and settings cannot be ignored.
The backlash around GPT-5, arguments over its style, Gemini rankings, and comparisons between Midjourney and Higgsfield do not change the central point. Models are becoming better and more human-like, so we give them more context. The more natural the dialogue feels, the more important it is to remember that this is still a company’s product, not a protected room.
The more natural a conversation with ChatGPT feels, the easier it is to forget that this is a company product, not a protected room. The context that makes the model useful also increases privacy and legal risk.
Episode transcript
The episode is in Russian; below is an English reading guide to the transcript (the full EN transcript is a machine translation). Voice matching applied to 122 segments: 69 identified, 8 mixed, 26 probable, and 19 unresolved.
Read transcript on a separate page
Loading…