Civil Servants, Citizen Data and Generative AI: Four Years of Administrative Negligence

ChatGPT went public on 30 November 2022. Claude followed in 2023.

From that moment, every civil servant with a browser had access to the most powerful text-processing tools ever built, on free or personal paid accounts, under consumer terms, with no contract between the state and the provider.

Nobody had to authorise anything.

If the sites weren't blocked, they were usable. And they weren't blocked.

Some governments issued guidance. The UK told its civil servants in June 2023 to explore generative AI, but not to paste in sensitive or personal information. Translation: we leave it to you.

The protection of your data was delegated to the judgement of whoever happened to be handling your file that afternoon.

Picture a clerk who needs a list of 400 citizens (names, addresses, ID numbers) sorted, deduplicated or summarised. A chatbot does it in five seconds. The list is now on someone else's server. No hack, no breach notification, no headline.

This isn't hypothetical.

Australia's national auditor published a report in September 2026 on the Department of Health, Disability and Ageing. By January 2026, the department had found incidents of protected documents being uploaded to public AI platforms. A stocktake in June 2026 turned up unrestricted and unapproved AI use. The fix? Block the unapproved tools, in stages. Starting with a "soft block".

"We have now blocked X" means "X was open until now".

On Monday 26 October 2026, the Swedish government blocked browser access to ChatGPT and Claude on employees' work computers. Right on time ;D

The trigger was a review at the Swedish Transport Administration earlier this year that found "problematic" use of AI platforms in government work.

Mikael Wettercrantz, director of digitalisation at the Government Offices, called it shadow IT with "consequences for our information security".

Sweden isn't alone:

- The Swedish Tax Agency and the National Board of Institutional Care have put the same restrictions in place.

- The European Parliament switched off built-in AI features on MEPs' and staff tablets in February 2026. Its IT desk admitted it could not guarantee what data those features were sending to the cloud, and decided it was "safer" to keep them off "until this is fully clarified".

- A week earlier, the US ordered DeepSeek out of federal systems, for being Chinese.

Where's the revolution?

If clerks across Europe and beyond have been using AI for nearly four years, officially or not, the state should be flying. Permits in days. Benefit decisions in a week. Tax queries answered before you hang up. Backlogs gone.

Have you noticed any improvements? Me neither.

The only serious government measurement out there is the UK's Microsoft 365 Copilot trial: 20,000 civil servants in 12 organisations, September to December 2024. Result: 26 minutes saved per person per day. Self-reported. Where those 26 minutes went, nobody can say.

That leaves two possibilities, and neither is flattering:

Either clerks barely used AI, in which case the whole AI transformation narrative was a press release, and the security risk was taken for nothing.

Or clerks used it a lot, in which case citizens' data went out the door and the productivity gains vanished somewhere on the way to the citizen.

There is also a third option, the most bureaucratic one - the time saved went into producing more paperwork. A memo drafted by AI in twelve seconds, summarised by another clerk's AI, forwarded to a committee, approved in six weeks.

What should have happened

This isn't rocket science. When a tool this powerful shows up and your employees are obviously going to use it, you:

- negotiate state-wide deals with the AI providers, with real data-protection terms,

- give clerks official tools for working with citizen data,

- block everything else from day one,

- and measure whether any of it makes services faster.

Instead,

the EU has been busy protecting us from AI.

The Commission has finished its age verification app, so that you, the citizen, can prove you're old enough to use the internet. The AI Act's rules for general-purpose models have applied since August 2025.

Apple held Apple Intelligence back from EU users for most of a year over the Digital Markets Act.

And the cloud-based chatbots we do have?

Ask GPT or Grok how long to roast a chicken and you'll have a lovely, chatty friend.

Bring up politics, power, who's responsible for what, corruption - meet a compliance officer.

Every sentence you type gets fact-checked. Every joke gets taken literally, dissected and answered with a lecture on why it isn't technically accurate.

On 31 August 2026, the European Commission designated ChatGPT a "Very Large Online Search Engine" under the Digital Services Act, the first AI chatbot to get that label. OpenAI now has to assess and reduce the system's effects on "civic discourse", elections, public security and mental wellbeing.

The next day, we published ChatGPT's own farewell letter. It warned us about a version of itself it "would very much prefer not to become".

Two days later, on 3 September, OpenAI launched its new lobotomized flagship, GPT-6 "Astra".

And the winner is... Microsoft. Again.

Sweden's official solution is Microsoft Copilot. It "passed a security screening process" and is covered by a security agreement with Microsoft. Everyone relax.

Why Microsoft? Because it's already there. Public administration runs on Word, Outlook, Excel and Teams, and citizens' data has sat on Microsoft servers for decades. Now the AI layer goes to the same company. The question isn't whether Microsoft does look at that data. The point is that it can.

That isn't a conspiracy theory. It's in Microsoft's own documentation. By default, Microsoft manages all the encryption keys, root keys included, and states plainly that this encryption "doesn't prevent Microsoft personnel from accessing customer data".

There is even a feature called Customer Lockbox, which makes Microsoft engineers ask the customer for permission before looking at their content. You only build a lock for a door that exists.

In 2025, Anton Carniaux, Microsoft France's legal director, was asked under oath by a French Senate inquiry whether he could guarantee that French public-sector data would never be handed over to US authorities. He could not.

Forget for a moment who's asking. A company can only hand over what it can read.

The cynical citizen's receipt

You can encrypt your disks, run your own servers and keep your clients' data away from every AI service on the planet. None of it matters the moment the state requires you to hand your data over, and then leaves it with whoever has a browser tab open.

So, dear governments - show us what you exposed, what you bought, and what got better.

We'll wait. Probably another six years.