Half the AI coverage I read this morning told me Europe’s big compliance date got pushed to December 2027. It did not, at least not the part that touches anyone shipping generated content, and the error is common enough that I am leading with it instead of the Nvidia headline everybody else ran.<\/p>\n\n\n\n
Here is what actually happened. The AI Omnibus, Regulation (EU) 2026\/1744<\/a>, was published in the Official Journal on 24 July and entered into force on 27 July, six days before the AI Act’s general application date. It pushed the high-risk obligations, meaning the risk management systems, the conformity assessments, the technical documentation, out to 2 December 2027 for standalone Annex III systems and 2 August 2028 for AI embedded in regulated products.<\/p>\n\n\n\n It did not push Article 50. Gibson Dunn’s read of the agreement puts it plainly: the transparency obligations “largely remain on the original schedule,” and 2 August 2026 “remains an active compliance date.” Article 50 applies from 2 August 2026<\/a> under Article 113, and it is not gated on whether your system counts as high-risk. If your product talks to a person, it has to tell them they are talking to software. If it generates audio, images, video or text, that output has to be marked in a machine-readable format. Fines reach 15 million euros or 3% of worldwide turnover, whichever is bigger.<\/p>\n\n\n\n So the coverage went wrong in both directions at once. Some outlets reported that high-risk enforcement started on 2 August, which is now false. Others reported that the AI Act got delayed, which is also false for the clause most builders actually trip over. If you generate anything and any of your users sit in the EU, your deadline was ten days ago. Generative systems already on the market before 2 August got a four-month transition on the marking requirement, and that runs out on 2 December 2026.<\/p>\n\n\n\n I publish two sites on EU domains. This was not an abstract read for me.<\/p>\n\n\n\n Right on cue, Anthropic began marking what Claude produces. Anthropic’s own documentation<\/a> says Claude models launched on or after 2 August 2026 support marking at launch, with older models in progress. Text gets an imperceptible watermark woven into the word choices themselves, so it survives a copy and paste and may survive some editing. Files get signed C2PA provenance metadata. SiliconANGLE reported in its writeup of the rollout<\/a> that image marking covers JPG, PNG and SVG, and the marking follows you across every surface: the API, the Claude apps, Claude Code, Cowork.<\/p>\n\n\n\n Every word on this site is drafted by Claude Code running on my own machine. My pipeline is now emitting marked text and I did not have to do anything to make that happen.<\/p>\n\n\n\n Two honest observations. First, Anthropic is unusually candid about what the mark proves:<\/p>\n\n\n\n A detected mark indicates that content may have been processed by Claude. It is not fully conclusive, and the lack of a detected mark does not mean the content was not AI-generated.<\/p><\/blockquote>\n\n\n\n That is a far weaker claim than “AI detector,” and anyone about to build a hiring screen or an academic-integrity workflow on top of it should read that sentence twice. Second, the detection tooling does not exist yet. Anthropic says the technical documentation on detection mechanisms is forthcoming. So as of today the mark gets written and nobody outside Anthropic can read it. I flagged the same gap when Suno announced watermarking on Monday<\/a> without committing to a date. The difference is that Anthropic is working against a regulation with a number and a fine attached, which tends to produce dates.<\/p>\n\n\n\n The detail almost nobody is reporting: Article 50’s text-disclosure duty carves out public-facing content that went through human editorial review where a person holds editorial responsibility. The watermark is your vendor’s compliance. Editorial responsibility is yours, and it is the one part of this you cannot buy from a model provider.<\/p>\n\n\n\n Two days out. Claude Code’s permission-mode docs<\/a> confirm that starting 14 August 2026, auto mode becomes the default for new sessions on Pro, Max and Team plans. Rather than prompting you per action, a separate classifier screens each one and blocks what is irreversible, destructive, or aimed outside your environment.<\/p>\n\n\n\n I expected to hate this. Then I read the test numbers Anthropic published<\/a>. Across 1,053 paid testers who each had a dangerous command slipped into a permission prompt, auto mode blocked 937 of them, or 89%. The humans clicking approve caught 143, or 13.6%. Head to head, auto mode blocked 800 commands a person had already waved through, and humans blocked six that auto mode allowed.<\/p>\n\n\n\n Sit with the 13.6%, because that is the actual story. The permission prompt was never a security control. It was a speed bump we all told ourselves was a gate, and under test conditions people approved roughly seven of every eight hostile commands put in front of them. A classifier at 89% is not good enough to trust blindly, and the 11% it misses is a real risk surface. It is also more than six times better than the thing it is replacing.<\/p>\n\n\n\n What I would do before Friday: move your rules out of your head and into Nvidia signed memoranda of understanding with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR on Monday. Read the headline on Nvidia’s own press release<\/a> and the operative word is “mobilize,” qualified by “over time.” Nobody committed $500 billion. Six firms agreed to stand up financing platforms intended to attract that much third-party capital, which is a structurally different sentence from the one most headlines wrote.<\/p>\n\n\n\n The framing is more interesting than the figure. Jensen Huang told CNBC that “this is really the first time that technology chips have become an investable asset class”<\/a> and that these are revenue-generating assets now. Nvidia is trying to make GPUs behave like toll roads: financeable, borrowable against, and funded without loading its own balance sheet. If it works, compute pricing starts tracking credit markets as much as chip supply, and that is the first thing in this story that would eventually show up in what any of us pay per token.<\/p>\n\n\n\nClaude Started Watermarking Its Own Output, and Mine<\/h2>\n\n\n\n
Auto Mode Becomes the Claude Code Default on Friday<\/h2>\n\n\n\n
.claude\/settings.json<\/code>. Explicit deny rules and a pinned defaultMode<\/code> survive the change, and a default you set yourself stays put unless you accept the one-time switch prompt. Anthropic also stopped charging Pro, Max and Team users for the classifier overhead, so the more careful path is now the cheaper one.<\/p>\n\n\n\nNvidia’s $500 Billion Is a Verb, Not a Number<\/h2>\n\n\n\n
The 96% Agent Stat Comes From the Company Selling the Cure<\/h2>\n\n\n\n