When the government blocks ChatGPT, the question isn't *if* but *how*

A block is a statement, not a solution

Sweden's government offices have reportedly blocked access to ChatGPT and Claude from staff computers from Monday ^[1]. The decision shows something every organisation with an IT policy already knows: writing the rule is the easy part. Enforcing it is harder.

Blocking a service is like locking the biscuit tin. It works until someone finds the key, brings their own tin or decides they are allowed an exception. The question isn't whether staff should use generative AI—many already do, on their phones if not their desks—but how to make sure they use it in a way that doesn't put sensitive information at risk or create records the organisation cannot account for.

Under Article 50 of the AI Act, which applies from 2 August 2026, people must as a rule be told when they are interacting with an AI system ^[2]. For a government employee using a chatbot to draft a response or summarise a document, that notice is straightforward. The harder question is what happens to the conversation afterwards. If a member of the public asks what the AI said, or a supervisor asks what information was fed into it, the answer had better not be "we don't know".

What the rules actually say

The AI Act does not ban generative AI in public administration. Article 50 requires transparency: staff and citizens must know when they are talking to an AI, and AI-generated content that is published to inform the public must be marked as such, unless a human has reviewed it and takes responsibility for it ^[2].

The GDPR sets a different boundary. Personal data processing must be limited to what is necessary, and it must be protected against unauthorised or accidental alteration ^[3]. Typing a citizen's details into a chatbot run by a third party is processing. Whether that processing is lawful depends on the purpose, the safeguards and whether there is a legal basis. A general block suggests the organisation has decided the risk is too high, at least for now.

Neither rule says you cannot use AI. Both say you must know what you are doing with it, and be able to show it afterwards.

The problem a block does not solve

A technical block stops one route. It does not stop the question. Say a civil servant uses a chatbot on their phone to help draft a decision that affects someone's rights. Six months later, that person appeals and asks what information the AI was given and what it suggested. The organisation has two choices: it either has a record, or it says it does not know.

Not knowing is an awkward position. It makes it harder to defend the decision, harder to show that personal data was handled properly and harder to convince anyone that the process was fair. A block prevents the conversation from happening on the office network, but it does not create a record of the conversations that happen anyway.

That is where the evidence question lives. The rule says staff must be informed when they interact with an AI. Fine. But when the interaction is questioned later, who has the record of what was said?

A record of what the AI said

Here, the thing to seal is each answer, or each conversation when it ends. The sealed record includes the question, the answer and the time, and if required under Article 50, the notice that the user was interacting with an AI.

Sealing takes a fingerprint of the record, not the content, so the conversation itself never leaves the organisation's systems. The fingerprint is sent to WitniumChain, which seals it on a chain where any later change would show. If the conversation is questioned months later, anyone the organisation shares the record with can check that it is exactly what was recorded on the day.

Internal audit, supervisors and, if it comes to it, a court can all run the same check. The record shows what the AI said and when. It does not show that the answer was correct, that the decision was fair or that the organisation is bound by what the chatbot suggested. Those are separate questions. But it does mean the organisation is not relying on someone's memory, or on a screenshot that could have come from anywhere.

Starting is straightforward: the REST API, the SDK or the MCP server sends the fingerprint when the conversation ends and gets a witness ID back. The content stays with the organisation. The first 2,000 witnesses a month are free. Read the docs.

How it would work

1. A staff member asks a question in a chat interface. The system records the question, the answer and the timestamp.
2. When the conversation ends, or when the answer is given, the system sends a fingerprint of that record to WitniumChain.
3. If the conversation is questioned later, the organisation shares the record and the witness ID. Anyone with both can check that the record has not changed.

What this means for you

If your organisation is thinking about blocking generative AI, or has already done it, ask these questions:

• Do you know whether staff are using AI tools anyway, on their own devices or accounts?
• If a decision is appealed, could you show what information was given to an AI and what it suggested?
• When Article 50 applies from 2 August 2026, will you be able to show that users were told they were interacting with an AI?
• If you allow AI use under certain conditions, do you have a way to record what was asked and answered, without storing sensitive data with a third party?

A block is one answer. A record that nobody can quietly rewrite is another, and the two can work together. The block stops the risk at the door; the record deals with the conversations that happen regardless.

What to watch

Article 50 of the AI Act applies from 2 August 2026. Organisations using AI to interact with staff or the public should by then have a way to show that the transparency duty was met, and a record of what the AI said.