AI Safety: What You Can Trust AI with and What You Can't. The 5 Levels of Your Data and Where the Line Runs

Mail was stored in the cloud before, banks saw transactions, phones kept messages, but artificial intelligence adds a new layer — it can understand all of it. A system that can be taught to understand a person can be taught to influence one: not with sneaker ads but with advice, an argument and a phrasing that changes a decision. So this is not about paranoia but about grown-up work with AI — what to trust it with, what not to, and where the lines run.

There are countries where every phone call is recorded — in Russia it is the Yarovaya law — and the recordings sit for several years and cannot be deleted. Analysis used to require one person to listen to the calls or search a transcript for mentions of people and actions. Today a conclusion on whether a person reasoned about certain topics takes under a second, and the point is no longer speed but that the system can use who you are.

ChatGPT trains on the user's data by default, but even with training off all chats continue to be stored, and a deleted chat is deleted over thirty days, not at once. A person usually decides to delete data when a threat appears — which means the threat already exists, and the data stays available for another thirty days, perhaps longer. Neither ChatGPT, Gemini nor Anthropic lets you delete chats on a given topic; the host himself does not delete his chats and knowingly trades confidentiality for the quality of the system's work.

Every request now goes into a model that reasons, and inside it there is a safety block: protection against distillation — China runs tens of thousands of bots to train its own models on the answers of American ones — national laws and the companies' internal rules. OpenAI, Anthropic, xAI and Gemini differ in the style and quality of their answers, and the answer depends on the country, the IP address, where the account is registered, previous chats, the language and the prompt. The American companies agreed not to explain how to make bombs and chemical weapons, but otherwise the chat has to choose what to answer people in countries with different laws.

First a request is checked by automated moderators, then by a human. In the Canadian murder case the killer's ChatGPT requests went for review to OpenAI employees, who refused to hand them to law enforcement. Shared ChatGPT chats ended up on the internet in full, and Google states outright that reviewed chats may be kept for up to three years, although ordinary ones are deleted after eighteen months. A conversation with an attorney or a doctor in America cannot be brought into the open; a conversation with ChatGPT or Gemini can.

Sneaker ads are the weak level; real influence begins where the system knows what you fear, what matters to you and where your money is. Even after the chats are deleted, the system has already learned who you are: it sees weaknesses, dreams, conflicts, who influences you and where your position is weak. Influence used to cost a lot — it took a psychologist, a salesman, a political strategist — now it can be scaled, and by giving data you hand over a map of how to talk to you.

OpenAI and Google with Gemini act within US law: the AGI they are working toward must not be used against US citizens, but may be used to defend the country — and therefore against the citizens of other countries in case of war. In China every system passes through a state control system built over decades: five years ago a politically negative message in WeChat already triggered a notice of a legal violation, and cameras catch even litter dropped in the wrong place. Systems built in Iran, Russia or China are fully under control, and chats can train on their data endlessly.

There are two leaders in models — the US and China — but the US has a phenomenal lead. ChatGPT alone has a billion weekly users, all of China together has six hundred million, Gemini has more than eight hundred million monthly, and Google has a billion in search. The US has tens of times more money and infrastructure: Anthropic, xAI and OpenAI are preparing public listings, xAI through Elon Musk's SpaceX, and Goldman Sachs expects the AI infrastructure market to grow to more than seven trillion dollars within five years.

The systems are getting better at detecting attempts to bypass sanctions and access restrictions — not because of the sanctions, but because OpenAI and Anthropic have no interest in their systems being used on foreign markets and work in the interests of certain people, clans and a state. Europe has strict regulation and user rights, but no magical disappearance of data; Pavel Durov, Apple and Elon Musk say Europe demands data in order to manage people — and the host agrees. Data lives in a legal world, not in a vacuum, and you need to understand the rules of your own country.

Green — general questions like a plant disease or a choice of grain; yellow — personal thoughts, plans and drafts, which the host calls the foundation of real AGI; orange — work, contracts, code and client data. Red — finances, medicine, taxes, passports and legal disputes; black — passwords, private keys and other people's personal data without permission, which should be given to an ordinary chat only with great care. The host finds it strange that the world ranks work data above a person's thoughts.

The host has thousands of hours of his own thoughts recorded over ten-plus years, and they seem more important to him than blood counts, cholesterol or an ultrasound, although the world treats medical data as more sensitive. He calmly uploads his tax returns to ChatGPT — there is no Social Security number in them — but does not urge others to do the same. The new GPT Finance module, in his view, was built for access to the user's finances, and that is far less important information than thoughts: it is through those that a person can be influenced on a really large scale.