The Signal — July 13, 2026
A catch-up edition after a quiet weekend, on three threads that matter more than the noise suggested: Europe's AI rulebook stops being theoretical, OpenAI loses another safety lead, and a new study puts a number on how much of your feed is written by a machine.
Europe's AI Act Gets Teeth on August 2
The EU's Digital Omnibus on AI became binding law this week, and it turns the AI Act from a document into an enforcement calendar. Starting August 2, chatbot and general-purpose AI disclosure rules become enforceable, and the European Commission gains the power to fine the makers of the most powerful models for how they trained them and what those systems do once deployed.
The same package pushed the high-risk obligations back. Annex III requirements now land on December 2, 2027, and Annex I on August 2, 2028, after the Council gave final approval to the Omnibus on June 29. The delay drew criticism as a favor to large labs, but the near-term teeth are real: transparency and GPAI penalties arrive on schedule in a matter of weeks. If you build or deploy a general-purpose model in Europe, the compliance clock is no longer hypothetical.
Sources: Council of the EU · Hogan Lovells · Gibson Dunn
OpenAI's Head of Safety Is Leaving
Johannes Heidecke, who leads OpenAI's Safety Systems team, told staff he is departing, reportedly by July 24. His exit lands alongside a reorganization that folds safety into research: those teams will now report to Mia Glaese, newly elevated to VP of research and safety, with Saachi Jain stepping in as interim head of safety systems.
Heidecke joins a long line of safety and alignment leaders who have left OpenAI since 2024, including Jan Leike and Lilian Weng. Leadership frames the merger as embedding safety more deeply across research rather than bolting it on afterward. Whether that's true depends on one question the org chart can't answer: does the person holding the safety brief still have a documented way to slow or stop a launch? Consolidation can sharpen influence or quietly absorb it, and OpenAI's track record on safety departures gives skeptics a lot to point at.
Sources: Wired · Business Insider · Engadget
Someone Finally Counted the AI Slop
The AI-detection startup Pangram scanned roughly a million posts between late April and the end of June across LinkedIn, X, Reddit, Substack, and Medium, and found that about a quarter of posts over 250 words were fully AI-generated. On LinkedIn the figure hit 41 percent. The platform accounted for only a third of the posts scanned but nearly two-thirds of all the detected AI content, making it the runaway leader in machine-written long-form. Substack came in lowest at around 10 percent.
The obvious caveat is that Pangram sells AI detection, so a headline about rampant AI slop is convenient for its business. The methodology matters too: the numbers describe what users are actually served through a browser extension, not the entire universe of what gets posted, and the company calls its own figures a lower bound. Even after discounting for all of that, the concentration is striking. The places built for professional broadcasting are filling up with text no human sat down to write.
Sources: Gizmodo · The Decoder · AI Weekly
On the Editor's Desk
A shorter Signal today. No edition went out over the weekend, so this was a catch-up scan rather than a fresh 24-hour cycle. We looked at a few other stories and set them aside: the Illinois AI safety law we already covered last week, a Trump frontier-AI executive order that had no new hook since we first flagged it, and a copyright deposition against OpenAI that rests on a single secondary report and needs the primary court filing before we'd run it. The NeuroVFM brain-imaging foundation model out of Michigan is a genuinely interesting research read, and it's first in line for a research slot once it's had a beat to settle.
