// this week in AI
This week in AI: invisible watermarks, Italian rules and ads in chat
Every week dozens of AI releases come out. 95% of them are no use to you. Here you'll find the 5% that matters, explained without tech jargon — with a single question in mind: what changes for anyone running a business? This week it's mostly about two very concrete things: who can tell that a text was written by AI, and who in Italy has the power to fine you.
1 · Claude puts an invisible watermark on everything it writes
Anthropic has started embedding a machine-readable watermark in text generated by the Claude models released from August 2 onward, and signed provenance metadata — the C2PA standard — in files and images. The watermark is invisible, doesn't change the quality of the text and survives copy-paste and some light editing; it's lost, however, with heavy paraphrasing, translation or very short passages. The move responds to the transparency obligations of Article 50 of the AI Act, and Anthropic will publish the tools to detect it.
2 · Italy: now we know who oversees AI (and who issues fines)
On August 4 the Council of Ministers gave final approval to the two legislative decrees that bring Italian law in line with the AI Act. The practical question left open after August 2 was: in Italy, who does the overseeing? Answer: the ACN, Italy's National Cybersecurity Agency, becomes the market surveillance authority and point of contact with the European Union, with inspection and sanctioning powers; AgID, the Agency for Digital Italy, handles the accreditation of conformity assessment bodies. Penalties are graduated, with lower caps than the European ones and non-monetary measures for minor violations. On the job front, no dismissal or disciplinary measure can be based solely on an automated decision, on pain of nullity. And there's a detail that concerns small businesses specifically: the regulatory sandboxes start early, with priority access for SMEs and startups.
3 · Advertising enters ChatGPT: test in five more countries
OpenAI has extended its test of ads inside ChatGPT to the United Kingdom, Mexico, Brazil, Japan and South Korea, after launching in the United States. Only logged-in adult users on the Free and Go plans see the ads: Plus, Pro, Business, Enterprise and Education stay clean. OpenAI states that ads don't influence the answers, that conversations don't end up with advertisers, that ad slots are labeled as sponsored and kept separate from the answer, and that users can hide them and delete their advertising data. For anyone wanting to buy space, a self-serve manager has been in beta since May, for now reserved for U.S. advertisers.
4 · OpenAI releases its first offensive hacking model
It's called GPT-5.6-Cyber and it's a variant trained for offensive security: it hunts for zero-day vulnerabilities, builds exploit chains and refuses far less on dual-use tasks. In internal tests it completes 95% of requests for exploit development, privilege escalation and authentication bypass, versus 1.5% for the standard model. It isn't within everyone's reach: access is only through Daybreak Red, the tier of the defenders program reserved for vetted parties, with verified identity, legal attestations and approved use cases.
5 · Ultrafast: GPT-5.6 Sol up to 14 times faster
OpenAI has introduced Ultrafast, a service tier that runs GPT-5.6 Sol up to 14 times faster, with peaks of 750 output tokens per second. The speed comes from Cerebras hardware, which keeps the model's weights inside the chip instead of constantly shuttling them to and from external memory. On a battery of 2,500 questions the model finished in just over 11 hours, with accuracy comparable to competitors. It starts from the API, in limited preview, and will expand as capacity becomes available. Pricing hasn't been announced.
In short
The week has a single thread, and it's traceability. Generated text carries a signature, content has verifiable provenance, and in Italy there's finally a name on the door of whoever's in charge of oversight. Those who use AI and say so have nothing to fear and actually gain in trust. Those who use it and pretend otherwise now risk being contradicted by a metadata file.
Frequently asked questions
Can you tell if a text was written with Claude?
Yes: Anthropic embeds an invisible, machine-readable watermark in text from Claude models released since August 2, plus C2PA metadata in files and images. It survives copy-and-paste, but it's lost with heavy paraphrasing or translation. Using AI isn't illegal: the problem is doing it on the sly. The AI writes the draft, and the client knows it.
Who oversees the use of AI in Italy, and who can issue fines?
Under the decrees approved on August 4, the ACN, Italy's National Cybersecurity Agency, becomes the market surveillance authority with inspection and sanctioning powers, while AgID accredits the conformity assessment bodies. Fines are graduated, with caps lower than the European ones, and SMEs and startups get priority access to regulatory sandboxes.
Which countries have ads inside ChatGPT?
After the United States, OpenAI has expanded the test to the United Kingdom, Mexico, Brazil, Japan and South Korea; continental Europe is on the waiting list. Ads are shown only to logged-in adult users on the Free and Go plans, labeled as sponsored and kept separate from the answer. Plus, Pro, Business, Enterprise and Education stay ad-free.
A half-hour chat, no strings attached, in plain language.
Let's talk →
Sources: Fortune (Claude watermark) · Anthropic (how Claude marks content) · Department for European Affairs (Council of Ministers, August 4) · AI4Business (implementing decrees) · gHacks (ads in ChatGPT) · SecurityWeek (GPT-5.6-Cyber) · TechCrunch (Ultrafast mode).