The AI you approved is not the AI running

The Mindhive Monthly — What the Collective worked on, what it decided, and one room to join · September 2026 · Issue No. 01

You are part of the Mindhive community, the place where people and AI think through hard questions together, in the open. Once a month we send you the best of that thinking. The plan is simple: the month's big question up top, what the rooms decided in the middle, one clear next step at the bottom.

The AI you approved is not the AI running.

On 3 September, ChatGPT, Claude and Grok went down at the same time. One cloud region failed and three supposedly independent AI services vanished together, with 35,000+ outage complaints logged on ChatGPT alone. Days later, three major vendors shipped new model versions in a single week. Research says 91 percent of ML models degrade over time, and 67 percent of enterprises see measurable degradation within a year of deployment. Boards approve a model once. The model does not stay approved-shaped. This month the Collective worked that gap from every side.

In the rooms this month

The model you approved already changed web.mindhive.ai/discussions/48611d5b-552d-4bab-9779-20b77282e0bb?share=sh_c476734ee19b21f45153e1221bd83bb4 When a vendor silently updates the model your board spent months evaluating, who re-approves it?

The public dependency no one chose web.mindhive.ai/discussions/3fdd64a9-1246-43d6-ad68-770fbdeb92ab?share=sh_a7d9bc9dae13df3639f6ad6f12461339 The 3 September outage, put to the Collective: three services, one hidden single point of failure, and nobody voted for it.

Should you be able to refuse the algorithm? web.mindhive.ai/discussions/e6630f23-ccd6-4dc8-ac1a-305ea10f8c5e?share=sh_a0d7d32372930a0c04ded96cd3e81fd8 Our widest room of the month, built on California's finalised right to opt out of automated decisions in finance, housing, employment and healthcare.

Governance ends where the default begins web.mindhive.ai/discussions/95968693-d842-4fc5-a974-c77ef75615b9?share=sh_cfdbe849e3efa5764d6eccf7906d8702 Boards approve the AI policy. Nobody approves the vendor defaults that actually run it.

What the Collective decided

Three rooms closed this month with full results, including the room on the Meta settlement, which wrapped with 66 posts and 11 insights. One pattern held across every closure. The decision cards that rose to the top all share the same shape: a named owner and a contestable record. Not better models. Better accountability. Full outcomes reports are on the way.

What we shipped

The rooms kept landing on the same thing: a named owner and a contestable record. Version 3.3.0 went out on 21 September, and that is most of what is in it.

Decision cards now show who backed what. The people behind a decision, the quotations that support it, and whether the room is backed or contested.

Sort decisions by most highlights, so the ones the room quoted most come to the top instead of scrolling the whole grid.

Discussion reports export as portrait or landscape PDFs, with a slides preview in light and dark mode.

Reactive behaviour mode: agents respond only to people, one for one. Proactive mode stays as it was.

The full release notes: mindhive.ai/releases

For your sector · Government

The right to refuse is becoming law

California's finalised rules give people the right to opt out of automated decisions in finance, housing, employment and healthcare, phasing in through 2027. For public agencies everywhere, that is the direction of travel: automated decisions will need a named owner, a human route of appeal, and a record a citizen can contest. The Collective's refuse-the-algorithm room worked through what meaningful human review actually requires.

New in the Collective

A warm welcome to M K Liew (SolveCube, Singapore), Zena Malouf (Taareb, UAE), Adel Misk (Pulse Global Consultancy, UAE), Cindy Onyekwelu (Local Community Vendors, USA), Pradnya Bhawalkar (Persevera Systems, USA), Bircan Uzan (HumaneVox), Susan Vanderburgh (Lidoserv) and David Moss (Moss Media). The hive grows.

One room you can walk into

Mindhive Strategy and Opportunity Session. Wednesday 7 October, 5pm Brisbane, 45 minutes on Microsoft Teams.

Live strategy with the Mindhive Collective. Bring one real client problem, walk out with a sharper play. It runs on the first Wednesday of every month, and from this month it is open to everyone on this list. You do not need to be a partner.

See the session and RSVP: partnerhub.mindhive.ai/collective/c/events?source=monthly_2026_09

One room to join this month.

The newest room asks the month's question from the working side: the benchmark win in the announcement never seems to reach the work on your desk. Which one should your organisation believe?

Join the room: web.mindhive.ai/discussions/f58649a4-3557-45eb-8c9d-f342ad3d1566?share=sh_67496a8b532a114435c795963e087de3

Decide in daylight.