Updated September 19, 2026
AI + Business
Free · Independent · Grant-funded
AICV News
News · Views · Field Notes

The Week in AI

the week in ai: september 14 to 18, 2026

September 14 to 18, 2026 · curated by · 8 pieces

The way people find a business is turning into a list somebody has to let you onto.

Meta opened its Muse agent to outside companies this week, and the process is an application. You describe what your connector does, Meta reviews it for functional, security and legal requirements and runs end-to-end testing, and if it passes you land in a directory where, in Meta’s own words, editors review connectors for featured placement. A person decides what gets seen. OpenAI, the same week, started testing ads in ChatGPT you can hold a conversation with — you click, and you get a separate chat with an agent the business is paying for. Neither company will say what it costs or who qualifies.

Set that beside the other half of the week. Reach, which publishes the Mirror, the Express and dozens of local sites, cut about 220 editorial jobs and closed three local brands outright. Google referrals were down 55% in the first half of the year, its website audience down 40%, and the company is dropping page views as its headline number. The owner of our own Desert Sun has said it may turn Google off entirely rather than keep feeding it. Pew watched people use Google with an AI summary at the top and found they clicked through 8% of the time, against 15% without one.

So the old door is closing while the new one is built with a review process in front of it. That is the shift worth understanding: from a search result you could influence to a directory you have to apply to.

Two different answers to “who checks any of this” also arrived. California now keeps a register of who is allowed to audit an AI system, and has asked experts to report back within two months on whether independent verifiers should sit inside the labs themselves. Anthropic and Accenture each committed at least a billion dollars over five years to evaluators working inside Anthropic with access comparable to an employee’s — while Anthropic said plainly that no standards yet exist for what those evaluators should see, or how they report what they find. In court, Universal and Sony argued that Suno’s newest model is “the fruit of the same poisoned tree,” that a training problem cannot be laundered by training on your own model’s output.

The encouraging part was local, and it was small. On one Saturday in Palm Desert, five teams formed and submitted five agents, one of forty-eight cities doing the same thing. Whatever these new lists end up asking for, the people who can build to them are already here.

reported1 piece

the click stops paying2 pieces

agents that sell and spend2 pieces

who checks the work2 pieces

what the models were trained on1 piece