TL;DR

Transparency duties under the EU AI Act took effect on 2 August, obliging providers of the most advanced generative models to tell users they are dealing with AI and to mark what those models produce. Anthropic and OpenAI have both signed up. UK businesses have no equivalent domestic requirement, and will meet these rules anyway through the products they use and the customers they serve.

What the providers have committed to

Anthropic said on Monday that every Claude model released since 2 August carries a watermark embedded in the text itself. Its blog post describes it as invisible, leaving meaning and readability untouched, and — the significant part — travelling with the words when they are copied somewhere else.

OpenAI made its commitment at the end of July, applying Google DeepMind’s SynthID to give generated images an imperceptible overlay, and extending the same treatment to audio. Both companies are also backing C2PA, the cross-industry provenance standard for carrying metadata about how a file was made.

Ashley Casovan of the IAPP argued the shift is larger than people appreciate: from readers making an educated guess about whether something was machine-made, to a verified marker telling them. DLA Piper’s Francesca Pole confirms transparency is where compliance attention has concentrated.

The gap between marking and seeing

Watermarking only works if something reads the watermark. Walter Pasquarelli, who researches synthetic content at Cambridge and advises the OECD, sets out the missing half: you generate text in Claude, paste it into a social platform, and that platform’s compose box needs a detector before any of this reaches a reader. Providers cannot deliver that alone.

He also notes the commercial consequence, which is why platforms may resist. Engagement falls significantly once audiences know content was machine-generated. LinkedIn has already added a way for users to report a post that “seems like AI slop”.

The contrast at home

Britain has published no comparable duty. The direction here is visible in today’s other regulatory story: the Property Ombudsman declining to require complainants to declare AI use at all. Brussels is building disclosure into the infrastructure; UK bodies are, so far, deciding case by case whether to ask.

Looking forward

Any UK firm selling into the EU should establish now whether its AI vendors mark their output, since the obligation reaches the provider but the commercial exposure lands on whoever ships the product.