Three reports from the past few days share the same lesson: AI at work calls for agreements. De Volkskrant found AI text in letters from ministers, a settings file from Claude Code put passwords on GitHub, and Binnenlands Bestuur explains when a civil servant may refuse AI.
76 parliamentary documents with AI text, nowhere disclosed
De Volkskrant analysed 3,597 letters to parliament and answers to parliamentary questions from the Jetten cabinet, sent between 23 February and 24 October. According to the AI detector Pangram, 76 of the documents contained AI-generated text, either wholly or in part; two letters consisted of it entirely. None of the documents stated that AI had been used. In some cases civil servants had worked with the free version of ChatGPT, which goes against the government's own rules. Because the prompts had not been saved, ministries were often unable to trace how the AI passages ended up in the documents.
What does this mean for you? If you work for or on behalf of a government body, use only the approved business environment and record what you used AI for. The difference between a free consumer account and a business variant is set out in the GDPR check for ChatGPT in government, alongside those for Copilot and Claude. You will find all the options for your field at AI for government.
Claude Code: "don't ask again" saves your password
If you choose "Yes, and don't ask again" in Claude Code, the tool saves the approved command verbatim in `.claude/settings.local.json`, including any passwords or API keys it contains. According to The Hacker News, security firm Apiiro found real secrets in 1,122 of 104,185 public copies of that file on GitHub, including 379 database passwords, 43 SSH logins and 23 AWS keys. The median: they had been exposed for 217 days. Claude Code does add the file name to git's global ignore list on your own computer, but that list does not apply to colleagues who clone the project and does not remove a version that was committed earlier, DEV Community writes. Earlier reports to Anthropic were closed automatically, according to The Hacker News.
What does this mean for you? Replace every key that has ever appeared in such a file, remove the file from your git history, add it to your `.gitignore`, and refer to secrets via a variable such as `$DB_URL`. Better still, approve a pattern rather than a complete command. If you work in software development or cybersecurity, also read the GDPR check for Claude in software development and compare it with GitHub Copilot.
Can a civil servant refuse to use AI?
Binnenlands Bestuur set out the position. Under the employer's instruction right, your employer determines which tools you work with, so you cannot simply refuse an approved AI system on principle. You can refuse if a task would breach the law, such as the GDPR, or if there are concrete risks of errors, discrimination or privacy violations. The AI Regulation also requires AI literacy, and you do not have to accept outputs blindly.
What does this mean for you? If AI becomes mandatory, ask about the agreements on privacy and training. For those working in HR, this is the moment to put those agreements on paper.
In brief
- In 76 parliamentary documents there was AI text without any disclosure, sometimes from the free ChatGPT.
- Claude Code can store passwords in `.claude/settings.local.json`; keep that file out of git and replace any leaked keys.
- A civil servant may refuse AI where it breaches the law or poses concrete risks, not on principle.
- All models and privacy scores are at models and safety.
