Thursday, September 17 September 17, 2026

Good morning. It's Thursday, September 17th, 2026 — and today AI safety isn't just a tech industry talking point. It's the topic of a royal audience.

King Charles the Third is meeting this morning at Dumfries House in Scotland with the top executives from OpenAI, Anthropic, Google DeepMind, and Nvidia. Sam Altman, Dario Amodei, and their counterparts sat down with the British monarch to discuss one urgent question: who makes sure this technology stays under human control? Reuters is reporting that King Charles personally urged the group to ensure AI remains, in his words, "firmly in the service of humanity." It's a striking moment — a head of state convening the four most powerful AI labs in one room to talk about guardrails. And it's happening the same week that multiple other safety signals are flashing.

OpenAI, Anthropic, and Google DeepMind have confirmed they are in active talks to form a shared AI safety standards body. The initiative is called CAISI, and five labs are already participating in early testing. This is significant. These are companies that compete fiercely on model capability, and now they're sitting down together to agree on shared safety benchmarks. Whether it produces real teeth or stays a PR exercise is the right question — but the convergence itself is newsworthy.

Meanwhile, Microsoft quietly dropped a 37-page document last Sunday called the Humanist AI Code of Conduct. It's a formal rulebook for Microsoft's own AI models — covering everything from how systems should handle sensitive topics to how they should behave when pushed toward harmful outputs. Thirty-seven pages of self-imposed rules from the company shipping AI into enterprise software used by hundreds of millions. That's a big surface area.

On the capability side, Google DeepMind leaders are floating something that sounds like science fiction: self-improving AI within a year. That's tied to Gemini 4, which is currently in training. To be clear, this is speculative language from internal Google discussions — not a confirmed product announcement. But when the people building the technology start using phrases like "self-improving" on the record, that's worth noting. The gap between capability and governance has never felt more live.

There's also a sharp public debate breaking open between tech founders. Mark Zuckerberg took a pointed shot on social media at Anthropic, suggesting that major AI labs should be focused on safety rather than building faster. That's notable because Meta's Llama models have been some of the most aggressively open and fast-iterated releases in the industry. Zuckerberg calling for a slowdown — even implicitly — signals how much the public narrative has shifted in the last six months.

And here's a governance gap that deserves more attention: no U.S. federal law currently requires AI developers to publicly disclose dangerous model behavior. Not OpenAI, not Anthropic, not Google. If a model exhibits deceptive conduct or dangerous outputs, labs can choose whether to report it. Given today's summit and the CAISI talks, this gap will almost certainly come up as one of the first items any serious standards body would need to address.

Finally, a funding note: Liquid Compute emerged from stealth yesterday with fifteen million dollars in seed funding. They're building a regulated marketplace for trading AI compute — essentially a financial exchange for GPU capacity. As inference costs dominate AI budgets at scale, a liquid market for compute infrastructure could matter a lot to enterprise AI buyers.

The theme tying all of this together is accountability. Safety summits, shared standards bodies, conduct codes, compute markets with regulatory structure — the industry is building scaffolding around a technology it acknowledges it can't fully control yet. Whether that scaffolding is strong enough is the multi-trillion-dollar question.

One business idea worth acting on today: an AI safety audit firm that helps mid-market enterprises comply with emerging standards like CAISI before they become mandatory requirements. As the big labs self-regulate and governments watch closely, the companies buying AI products will face downstream liability questions. A boutique firm that audits AI deployments against emerging standards — and produces certification-style reports — is positioned exactly where regulatory demand will land next. Think SOC 2 for AI, but built before the auditors.

That's your MorningAI briefing for September 17th. Stay sharp out there.