The robots must now confess

Yesterday, parts of the European AI act started being officially enforced. You may have seen some headlines about it being delayed; ignore that, it’s only for high-risk stuff like hiring algorithms, credit scoring, or biometric ID.

As of now, the following are law:

• If you run a chatbot, it must signal to people that they’re talking to a machine and not a human. This has to be clearly indicated from the first hello, not hidden in some 8 pixel grey font in a footer.
• AI-generated or manipulated images, audio, or video that look realistic have to be labelled as such.
• If emotion detection or biometric categorisation AIs are pointed at people, you have to tell them they’re being read.

There’s one more, but with a caveat:

• AI-generated content has to carry machine-readable markings that detection tools can recognise. This only applies to new content. Anything generated before yesterday gets a stay of execution until December 2nd. At that point, it too gets the mark of Cain.

The commission has helpfully published some AI badges for your labelling needs. They’re not mandatory and carry no legal force (The labelling is mandatory, though). Just a little freebie to save you time…

There’s also something of a trap in the fine print: if you publish AI-written text on matters of public interest, you have to disclose the AI use unless a flesh and blood human reviews it and holds editorial responsibility (clicking “approve” and closing the window doesn’t count – you have to mean it).

For most people, the job today is to check their chatbot and make sure it announces loudly and without shame that it’s a clanker.

Colin