On September 23, Stanford confirmed that its housing and dining department ran a student photo through generative AI before hanging it on a Welcome Home banner. The AI dressed 3 students in Stanford merchandise and slimmed 2 of their faces. It took Billy Ramirez out completely. In his seat is a generated woman who, by every account, does not exist. Nobody asked him.
The photo came from a Lunar New Year dinner in 2024. Stanford already had a rule against exactly this. Its AI policy says using AI to produce or alter images of Stanford people is strictly prohibited, and Stanford says the edit and the missing disclosure both broke it.
The banners are coming down. Stanford says the fix is more training and more review. Billy has said he is tired of the attention, and the questions belong with the department that made the edit, not with him. The one still open is the one every owner should ask: the rule was in writing, so who was supposed to check before the banner went up?
That was the thread all week. A check existed on paper. Everyone assumed somebody was running it. Nobody was.
Photo: Amar Preciado via Pexels
What changed
A targeting system everyone assumed checked for stale records never did. On February 28, 2 Tomahawk missiles hit an elementary school in Minab, Iran, killing more than 150 people, at least 123 of them children. Bloomberg reports the database still listed it as a Revolutionary Guard compound, and Palantir's Maven made it a day one target. Central Command expected Maven to flag stale data. It was never built to. Reuters and the Washington Post reported Claude, the model I use for research, ran inside Maven.
Four AI coding agents promised a pinned plugin could not change. None of them verified it. AIR Security found Claude Code, OpenAI's Codex, GitHub Copilot, and Google's Gemini CLI lock a reviewed plugin to one exact commit, then never check the installed code matches it. Claude Code and Codex update plugins in the background by default. Nobody has reported it used against anyone. Anthropic and OpenAI patched. GitHub disputes how far it reaches. Google is retiring Gemini CLI and will not patch it.
Photo: PNW Production via Pexels
Two human reviewers approve every AI seatbelt fine in Western Australia. In one count earlier this year, about 60 percent of appeals won. Cameras use AI to flag a belt in the wrong spot, then people confirm the fine: more than 81,000 since October 2025, A$550 and 4 points each. Lisa Taylor was fined because her 10-year-old slid the strap under her arm in the back seat. It was thrown out more than 8 months later. In September, police started dropping hundreds more.
Photo: Grigoriy Bunkov via Pexels
In 2017 the Justice Department said a face match is not probable cause. A judge is now being asked to say it. In 2023 a Reno casino's facial recognition called Jason Killinger a 100 percent match for a banned man. He showed the officer a valid license and spent 11 hours in custody until fingerprints proved who he was. A court filing says the officer testified the arrest never should have happened. On September 21, the ACLU and the Innocence Project backed his federal case.
The one story this week where the check ran. Joseph Coates, 37, of Renton, Washington, was too sick with POEMS syndrome for the stem cell transplant that could put him in remission. An AI model built to find new uses for existing drugs suggested a combination never tried for his disease. Dr. David Fajgenbaum at Penn looked at it and decided it was worth the risk. Four months later Coates was healthy enough for the transplant. The model suggested. A doctor decided.
Outside the channel. On September 24 the Blue Cross Blue Shield Association published an analysis estimating $942 million in added hospital costs over 2 years from claims billed as more complex. It says the rise coincides with hospitals adopting AI billing software that scans records and lab results for anything that can be coded. Its test was a diagnosis on the claim with none of the treatment you would expect for it. Complex cases went from about 37 percent to about 40 percent.
What it means for your business
Every story this week had a check that lived on paper. A written policy. A pinned commit. Two reviewers. A 2017 guidance memo. Each one was real, and each one assumed somebody else was running it. Here is how to find out whether yours run.
Five moves for this week.
Write down what each AI tool actually checks, then make the vendor confirm it. List every tool that flags, scores, drafts, or edits for you. Next to each, write the one thing you believe it verifies, and ask the vendor to confirm it in writing. Central Command believed Maven caught stale records. Nobody could say why. If the vendor will not put it in writing, assume the tool does not check it.
Give every line of your AI policy a named enforcer, and test one. Stanford's rule said AI edits to images of real people were strictly prohibited. The edit still reached a banner. Pick one rule, such as no AI edits to photos of real customers or staff, name who signs off before anything publishes, and walk one real asset through that sign off this week.
Find out which AI coding agent versions your developers run today. If anyone on your team uses Claude Code, Codex, Copilot, or Gemini CLI, ask for the installed version on every machine. Confirm the Anthropic and OpenAI patches landed, move anyone off Gemini CLI since it will not be fixed, and make one person responsible for knowing what is installed.
Audit your reviewers, not just the tool. When a person approves what software flagged, track how often they overrule it. In Western Australia, 2 reviewers signed off on every fine and about 60 percent of appeals still won. A reviewer who never says no is not a check. Pull 10 recent approvals this month and have someone else redo them from scratch.
Check what your notetaker does after the meeting ends. According to her complaint in New Jersey, Lindsay Waninger left the call where she was fired, her managers stayed on, and Fireflies emailed her what they said next. No court has ruled. Open your notetaker's settings, confirm when it stops recording and who receives the summary, and turn off automatic sharing to everyone on the invite.
None of this needs new software. It needs an hour on Monday and a name next to every check.
What a real human check looks like, and how to keep one in place: Human-in-the-Loop AI: Keep Your Judgment and Voice.
What to ignore
"We should all slow down together." On September 12, Anthropic's chief executive published an essay urging the big labs to slow AI down together, and Sam Altman, Elon Musk, and Demis Hassabis agreed publicly the same day. Six days later, 4 paying subscribers sued Anthropic, OpenAI, SpaceXAI, and Google, claiming an illegal agreement to slow AI down. Their main evidence is those public statements. Whether that breaks antitrust law is a judge's call.
It changes nothing Monday. A pledge about the frontier did not change a setting on any tool you pay for. Ask your vendor what its product verifies, not what its leadership believes.
Watch the week
Every story in this issue started as a video. Two a day, 9 AM and 5 PM Eastern, each under three minutes.
See it clearly. Then use it.
That is the week. Reply and tell me what you are stuck on, I read every one.
Tim Bish
Understanding Your AI