Human Approved

The Luxembourg government has published guidelines for using generative AI in its communications[1]. The rules, released Tuesday, require human verification of all AI-generated content before publication and mandate transparency for deepfakes and AI-assisted text.

Prime Minister Luc Frieden signed the response to a parliamentary question from LSAP MPs concerned about AI-generated images in government campaigns. The MPs worried about photographers, graphic designers, and visual artists losing public commissions to machines.

The government acknowledges the risks to creative professionals: income loss, recognition, cultural diversity, copyright. But it also highlights productivity gains and new creative opportunities. It has no data yet on whether AI image generation has actually hurt employment or income in the creative sector.

Since March 20, guidelines have been in place for central government. Every piece of AI-generated content must be proofread and checked before publication. The staff member who approves it bears ultimate responsibility. The human does not always have to be flagged, though.

Transparency obligations under Article 50 of the EU AI Act took effect August 2. Texts informing the public on matters of general interest, plus images, audio, or video constituting deepfakes, must be identified as AI-generated or AI-modified. Some human-verified texts are exempt from disclosure.

The backdrop: on March 4, Frieden announced that all civil servants would get access to two sovereign AI tools based on Mistral AI technology, as part of a three-year, 40 million euro partnership. Sixteen days later, the government council adopted the guidelines. The state is accelerating AI adoption while trying to set its limits.

So the government is giving civil servants AI tools and telling them to verify everything themselves. What could go wrong. At least the transparency rules are real, and the EU AI Act provisions are now in force. The creative industries concern is legitimate but unanswered. The government essentially says "we know there might be a problem but we have no data yet." That is not a policy, that is a shrug.