Every AI Fix You've Already Made Has the Same Expiration Date

If you turned off an AI training setting, wrote a one-page AI policy, or signed off on what an agent is allowed to touch at some point this year, stop assuming that decision is still in effect. None of those three fixes hold on their own, and 2026 has already produced a separate, unrelated incident proving each one expired quietly.

Three fixes, three expiration dates

In September, users found that ChatGPT's opt-out from model training reverts to "on" after app updates, with no dialog and no email. A setting someone configured once, correctly, stopped being true the next time the app updated — and nothing told them.

The same month, Sam Altman and Dario Amodei briefed the UN Security Council on AI risk and proposed treating AI failures as reportable incidents, the way a hospital logs an adverse event. That only works for a team that already has a policy defining what counts as a minor slip versus a serious one — and a policy written in January about the tools you had then says nothing about the agent you approved in August.

And when Meta's Muse agent started buying directly instead of clicking Amazon's sponsored listings, it wasn't just a fight between two platforms. It was proof that the permission you granted an AI tool on day one — what it can see, buy, or send — describes what that tool did on day one. It doesn't describe what a platform update, a new agent mode, or a feature nobody announced lets it do now.

The habit, not the checklist

Each of those three posts ends with a specific fix: recheck the toggle, write the policy, audit the permission. All three are right, and all three share the same failure mode if you stop after doing them once. A setting, a policy, and a permission grant are not facts about the world. They're claims that were true on the day you checked, made by a product that has no obligation to tell you when it stops being true.

The fix isn't a better one-time audit. It's moving the check from "something I did" to a recurring line on a calendar — tied to events that actually invalidate the thing you configured: an app update, a new tool request, a quarter passing. Nobody downstream is going to flag the drift for you. The ChatGPT toggle reset with zero notice. Amazon didn't email Meta's users to say their shopping agent's behavior had changed underneath them.

Priya: the recheck that isn't optional anymore

Priya runs operations at a 60-person logistics brokerage that connected an AI assistant to its dispatch and invoicing systems last spring. She approved the integration once, read the permission screen once, and moved on. After the Muse story broke, she went back and found the vendor had shipped two feature updates since her approval — one of them letting the assistant auto-send invoice corrections instead of just drafting them. Nobody had asked her again. She now puts a 15-minute recheck on her calendar every time the vendor posts a changelog, not just once a quarter, because the changelog is the thing that actually tells her when day-one permissions stopped being accurate.

Dale: the counterpoint

Dale is a solo grant writer who never wrote a policy at all, reasoning that a one-person shop doesn't need one. He still does something equivalent without calling it that: before each new client, he spends five minutes confirming which AI tool and tier he's using for that client's draft, because a policy was never really the artifact that mattered. The artifact is the recurring question underneath it — am I still operating the way I decided to, or the way a login screen happened to default to this week. A written policy makes that question easier to repeat consistently across a team. It doesn't replace asking it.

The takeaway

A privacy toggle, a written policy, and a permission grant are not three different problems. They're the same problem wearing three outfits: something you configured once that a vendor, a platform, or an agent update is free to quietly change out from under you. The fix that survives is the one with a recheck date attached, not the one that felt thorough the day you did it.