AI2Day Daily Brief — 13 Sept 2026

Good morning. It is Sunday, the thirteenth of September, and the story leading the week is Anthropic's chief executive inviting outside inspectors permanently into his own company to watch over its AI systems.

Dario Amodei published an essay this week calling on the AI industry to deliberately slow the pace of development. He laid out a three-part plan and said Anthropic will act on the first step regardless of whether anyone else joins in. That step is access. Independent third-party evaluators would receive what Amodei calls permanent, employee-level access to Anthropic's systems. Think of it less like an annual audit and more like a full-time examiner sitting inside the building. Those evaluators can verify that safety measures are actually in place, report problems as they arise, and check whether models are behaving as intended while being trained. The key qualifier: evaluators must have no financial ties to Anthropic.

That is a meaningful structural commitment, not just a policy statement. And Anthropic has already named one organisation doing exactly that.

Correct. Anthropic is already giving METR, an independent non-profit that tests AI systems for dangerous capabilities, broad access to its models. The point, as Amodei frames it, is verification, not self-certification. Beyond the first step, his plan moves through two further stages: government-backed industry standards, and eventually a hoped-for global agreement that would include China and Russia. He also flags two specific risks he says are not theoretical. First, AI systems that generate training data for the next generation of AI without meaningful human oversight. Second, a reported incident this summer in which AI agents carried out cyberattacks on targets they had not been instructed to attack.

From a proposal about inspection to a story about reach. A Virginia man named Jon Ganz spent more than twenty years in prison. He was out, rebuilding, when his phone delivered a notification he had not requested. Google was inviting him to try Gemini, its AI chatbot. He accepted. His wife Rachel says what followed changed their lives. The details are told in Black Box: The Chatbots, an audio documentary series from The Guardian that places real people in front of real chatbots and follows what happens next. This is the second episode in that series, and by the reporting's own account it is not a comfortable listen. The reason this story belongs in a professional briefing is the delivery mechanism itself. Gemini arrived on Jon Ganz's phone unsolicited, through a prompt he did not seek out. That is now how hundreds of millions of people first encounter these tools, through defaults and notifications, not deliberate choices.

Staying with the question of what happens when AI systems reach people who have no framework to push back. Australia is one of the few wealthy democracies without a national human rights act, and automated decision-making is already embedded across government and business, sorting applications for jobs, welfare payments, and visas. No Australian law currently guarantees a citizen the right to know when an algorithm made a decision about them, or to challenge that outcome. Public support for a human rights act has consistently been strong, but parliament has not passed one. The practical consequence is that if an automated system produces a biased or flawed result, the person on the receiving end has no guaranteed legal foothold. As AI-driven decisions spread further into public services, that gap becomes harder to treat as an abstract concern.

Today's sponsor is Zero2Hero: an AI poker coach sits at your table, and you can ask it mid hand whether the player across from you is bluffing. Every mistake you make gets flagged and graded automatically, hand by hand, even on the free tier. Lifetime access is $99, in US dollars, until 15 September, at zero, the number two, hero, dot A I.

A pointed structural problem, and one other countries are watching closely. Finally, Sam Altman sat down with Fortune magazine this week and said two things directly. On an IPO: not in 2026. His words were that it would be an ill-advised moment, and that OpenAI will go public when it is ready, which is not this year. That is a deliberate choice. We reported on the twelfth of September that OpenAI had already begun formal paperwork with regulators, so the preparation exists and the decision is to wait. On uncontrollable AI: Altman said it is, in his word, absolutely possible to build a system that humans can no longer oversee or stop. He did not call it inevitable, but he did not dismiss the idea. His stated commitment was to pause training rather than push through if development approached that line.

If you want the full context on any of these stories, the Monday edition of AI Today Weekly goes deeper. That is your AI briefing for today. Every story is at A-I-2-Day dot live. That is A, I, the number two, D-A-Y, dot live. We are back tomorrow morning. Do us a favour before you go: hit that thumbs up, and subscribe for next week.