The headline version: OpenAI will start watermarking ChatGPT’s text in the EU.
The fine print: the detector goes to approved researchers only, the API mark is off by default, and by OpenAI’s own count a 10% synonym swap cut detection from about 92% to 66%.
OpenAI’s watermark ran everywhere on Monday as a detection story, and the documents describe something closer to a compliance switch that nobody outside OpenAI can test. The ChatGPT side turns on by itself for EU users over the coming weeks. The API side stays off until a developer opts in, which hands the decision to whoever builds on it. And the detector that would prove a mark is present is not available to you, your client or your auditor. Below are six stories from Sunday and Monday, read for people who ship on this stuff, each with a verdict.
OpenAI’s EU Watermark Arrives Off by Default in the API
Matters. OpenAI said on Monday that API customers worldwide can opt in to text watermarking for select models, and that eligible ChatGPT and Codex text generated in the EU will carry an invisible mark, called textGrain, over the coming weeks. TechCrunch’s writeup carries OpenAI’s own numbers on how easily it breaks: swapping 10% of the words for synonyms dropped detection from about 92% to 66%, and short passages, math answers and translations are harder to catch. Detector access goes first to approved researchers and expert organizations, case by case. OpenAI also says a missing watermark does not prove a person wrote the text.
The rule behind this is Article 50 of the EU AI Act, in force since 2 August, and when I covered it Anthropic’s documentation already said Claude models launched after that date support marking. So OpenAI is two months behind on the consumer side and still opt-in for developers. My read, a builder’s and not a lawyer’s: if you serve EU users from the OpenAI API, whether your output carries a mark is now a setting you own. Turn it on for EU traffic, write down that you did, and accept that you cannot verify it until the detector opens up.
Anthropic Went to the Police, Not the Other Way Around
Matters. On September 26 a woman in Bonita Springs, Florida wrote in a Claude chat that she was going to “shoot up” the Lee County Sheriff’s Office, and early the next morning wrote that she had a new gun. Anthropic’s safety systems flagged the conversation, a human review team escalated it and Anthropic notified law enforcement, according to WINK News, which reported that she told investigators she uses AI like a “diary”. She faces a second-degree felony charge under Florida Statute 836.10, with arraignment set for November 2. Tom’s Hardware counts it as at least the third Claude conversation to reach police since August. Anthropic did not respond to Gizmodo’s request for comment.
The page most people find when they ask what Anthropic shares with police is its policy on government requests, dated March 16, and it is written for the opposite direction:
“We will make an exception to this policy if we believe there is an emergency that may result in imminent physical harm or death, and providing such information without delay may avert that emergency.”
That sentence governs requests that arrive from a government. In Florida nobody asked; Anthropic went first, under a privacy-policy clause The Next Web quotes as allowing disclosure when it is “reasonably necessary” to prevent “serious harm”. I think Anthropic made the right call on a credible mass-shooting threat, and I would want my own vendor to make it. The problem is the word “diary”. If you put a model in front of customers, say in your privacy notice that the provider can review messages and, in an emergency, pass them to police, because users assume a chat box is private until a news story says otherwise.
Google Starts Turning Gems Into Skills, and Old Gems Arrive Switched Off
Breaks your stack. Skills began rolling out in Google Workspace on October 5 for Rapid Release domains, with Scheduled Release domains following from October 19, and they reach the Gemini app on October 13. Google’s admin timeline moves Gems into Settings on November 17. No sooner than March 1, 2027 for business and enterprise accounts (June 1, 2027 for education), users can no longer create, edit or use Gems, and Workspace Studio flows with Gem steps stop working. Whatever Gems remain then migrate to draft skills that their creators must review and activate before anyone can use or share them. Sharing permissions carry over, but the activation step does not happen by itself.
This is the second Gemini access change in a week, after the free tier’s drop to one model. The break is quiet: my read is that a team Gem built by someone who has since left becomes an inactive draft with no obvious owner to switch it on. List every Workspace Studio flow that calls a Gem this month and rebuild the important ones as skills well before March. Skills do stack inline in one chat, which is a real upgrade for layered instructions.
ChatGPT Ads Move Into Image Generation
Marketing, in the literal sense. OpenAI announced on Monday a visual ad format that starts testing later this month in the US, shown during image generation, labeled and kept separate from the picture being made, TechCrunch reported. Free and low-cost tiers see the ads, and OpenAI says they will not influence answers. The bigger change for anyone who buys media is the measurement list: attribution partners including AppsFlyer, Triple Whale and Northbeam, plus brand-suitability pilots with DoubleVerify and Integral Ad Science. That is the kit performance marketers ask for before they move budget, and it tells you OpenAI wants performance money more than brand money.
Last month ChatGPT ads learned to talk back, and now they sit in the one moment a user is waiting with nothing else on screen. Test it with a small budget only if your product looks good in a picture and your buyers live on the free tier. Then judge it on conversions you measure yourself, because the partner dashboard is now part of OpenAI’s pitch.
Reflection Announced Beam Without the Weights
Marketing, for now. Reflection AI introduced Beam on Monday, a sparse mixture-of-experts model with 501 billion total parameters and 23 billion active, under an Apache 2.0 license. Reflection’s own post says the weights, technical report and model card arrive “later this month”, and today you can sign up for early access. It claims Beam is competitive with GLM 5.2, approaches Qwen 3.8-Max on coding and agentic tasks, and uses three to four times less inference compute than GLM-5.2. The comparison already looks dated, since Anthropic published its analysis of GLM-5.3 on September 29, a week before Beam was announced. And 23 billion active parameters keeps each token cheap to compute, but you still have to hold all 501 billion in memory, so for most teams this is a cluster model rather than a laptop model.
The Super Intelligence Force Puts the FTC Chair on Both Sides
Marketing, with one deliverable worth watching. On Sunday Trump named Director of National Intelligence Jay Clayton to lead a new “Super Intelligence Force”, with FTC Chairman Andrew Ferguson, Pentagon CTO Emil Michael and OPM Director Scott Kupor alongside him, the AP reported. It follows the September 29 executive order telling federal agencies to say “super intelligence” instead of “artificial intelligence”. SiliconANGLE’s account gives the group 120 days to report on risks and opportunities, with recommendations on how the government discloses security breaches, and a brief to plan for threats “while preventing overregulation and regulatory capture that would stifle innovation and competition.” Ferguson chairs the same agency that, according to a senior official speaking to the New York Post, is preparing civil investigative demands against OpenAI and Anthropic.
On October 1 I declined to call that probe a story because nobody had put it on the record, and that is still the case: one unnamed official, no filing. Now the investigator also sits on a body told to guard against overregulation, an awkward seat while deciding whether to compel testimony from the industry. The rename is branding. The breach-disclosure recommendation, due around the start of February, is the one item that could change what a vendor must tell you after an incident, and it pulls on the same thread as the self-funded AI regulator idea I covered in September.
Where I Land
Four of today’s six stories turn on who gets to see what: the watermark detector, the chat review queue, the ad measurement stack and the breach report. Ask each AI vendor you depend on what it can see about your usage that you cannot, and file the answers next to your contracts. If one of those answers touches a system you run, book a free intro call and we can look at it together.