Six questions for the person. None for the system.

What the poster says

The NAIH page itself is two short paragraphs: who made the poster, that NAIH translated it for Hungarian data subjects with the authors' permission and CNIL's cooperation, and where the French and English versions are. There is no Hungarian data, no new regulatory position and no enforcement signal, and NAIH does not claim otherwise.

The poster is a single page under a CC BY-NC-SA 4.0 licence, written entirely in the first person singular. The six questions, paraphrased:

• 1. Which app to choose: compare services and pick the one least likely to reuse your prompts for other purposes; read the privacy notice.
• 2. Whether to create an account: use the service without one if you can; if you must, use a strong password and two-factor authentication.
• 3. How to protect your data: use the settings to limit training on your data; use temporary mode and turn off history.
• 4. What to share: no personal or sensitive information; the examples include your real name, address, school, passwords and bank details.
• 5. Whether to trust the answers: AI can get things wrong; verify anything important against reliable sources.
• 6. What to do afterwards: review and delete your history, delete accounts you no longer use, and log out, especially on shared devices.

Two things about dates and audience. CNIL published the original on 27 May 2026 in Korean, French and English, and its page says the poster is distributed through schools, social media and events. The news is the Hungarian publication on 2 September. The audience shows in the examples: your school is among the things to protect, and NAIH's page addresses individual data subjects. This is awareness material for private individuals, in part young ones. NAIH nowhere presents it as corporate guidance, and neither will we.

Five of the six are behaviour. One is a setting.

Look at what decides the answer to each question. The first is a one-off decision. The second is a habit. The fourth, fifth and sixth have to be done again every single time: think before you paste, check what you got back, delete what you left behind. The third is the only one that is a setting: a switch you flip once, after which it works without you.

For one person with one account, that is fine; if they forget, it is their own data at stake. For a company of 20 to 200 people the same six questions look different. Question one becomes a procurement decision. Question two becomes single sign-on. Question three becomes an admin setting or a contract clause. Question six becomes a retention policy. All of these can be solved at the system level, and the better-run companies do solve them.

Question four stays exactly where the poster left it: with the person, at the moment of pasting. On a public service, no setting can tell whether the text just pasted was a client's contract. Only the person pasting it knows that. On a Friday afternoon, before a deadline, three months in, long after the onboarding slide has been forgotten. Anything that depends on someone remembering is awareness. Control starts where they no longer need to remember.

And question five stays with the person on our system too. Private AI gets things wrong as well. This needs saying, because "private" is easy to hear as "reliable". Private is a statement about where the data sits. It says nothing about the quality of the answer.

What we think, including the part that cuts against us

Of everything we read this week, this poster argues hardest against our own position. That is why we are writing about it. Four things need saying plainly.

First, the poster was not written for companies. If we treated it as a weak piece of corporate guidance from a regulator, we would be holding it to a promise it never made, and any informed reader would see that. For its own audience, it is good material.

Second, and this is the harder one: questions one and three say precisely what we compete against. Public AI services can be compared by how much they reuse your prompts, and their settings can bring the risk down to an acceptable level: training off, temporary mode on. A data protection authority is saying this to individuals, but the logic holds at company level too. The enterprise tiers of the large providers commit contractually not to train on your data and let you set retention periods. That exists, it works, and for many companies it is the correct and cheaper choice. Anyone who tells you otherwise is not being straight with you.

Third, what this source leaves us is narrower than usual. A setting is a switch someone can flip back, on a service you do not operate, verified by a web page you cannot audit. That is a promise. It may be a good promise, backed by a contract, and a good contract is worth a great deal. It remains a promise. That applies to us as well: with our cloud edition you are still trusting an operator, and the operator is us. Physical location, by contrast, is a fact: the machine is either on your own premises or in a named European data centre, and you can go and look. This source will not bear more than that, and we do not intend to load more onto it.

Fourth, the material itself is not new. CNIL released it on 27 May; what is new in September is the Hungarian translation.

What we run instead, and why it is not a trick

Our Private ChatGPT runs open-weight models on private European GPUs behind your company login. Prompts, documents, logs and embeddings stay on the platform, on EU infrastructure, with no US data processor in the path. Depending on the edition, it runs on our European hardware or on yours, in your own data centre. Details on the Private ChatGPT page and the editions page.

Take the same six questions on that system. Question one is answered once, by contract. Question two is the company login you already use for email. Question three disappears: there is no switch to flip back, because neither a third party nor we train on your data. Question six is your own retention policy, on your own storage. Question four still belongs to the person, and we will not pretend otherwise; what changes is the stake. A pasted contract stays inside the company environment, on the same hardware, with the same European operator. The data does not leave the company. Question five is unchanged: answers still have to be checked.

Whether that beats an enterprise subscription in your case is a calculation worth doing honestly, and for some companies the answer will be "not yet".

Where we stand

We are not against public AI. We use public models ourselves where the data allows, and the six questions are a good reminder for our own staff. We have no quarrel with the poster. The problem starts where a company's entire AI policy amounts to "be careful what you paste". That is awareness. It is necessary, and it is not enough.

If there is one question to take home this week, make it this one: if a new colleague pastes your most important client contract into a chat window on their first day tomorrow, where is that text at six in the evening? If the answer is a setting, it is a promise. If the answer is the address of a server room, it is a fact.

Guides: Private AI for business · Private ChatGPT · Private AI hosting · Private AI platform · Your own private AI · GDPR-compliant AI hosting