On September 15, Angela Lipps filed a $10 million federal lawsuit against the City of Fargo, North Dakota, and the detective who charged her with 8 felonies. She is a grandmother from Elizabethton, Tennessee. She had never been to North Dakota.
In the spring of 2025, a woman used a fake military ID to pull thousands of dollars out of Fargo banks. According to Fargo's police chief at the time, West Fargo police ran a photo from that ID through Clearview AI, built on billions of scraped photos. It returned her name: a potential suspect with similar features. West Fargo never charged her. A Fargo detective compared the footage to her social media and driver's license, decided it was her, and a judge signed a nationwide warrant. On July 14, 2025, U.S. Marshals arrested her at gunpoint while she was babysitting. More than 3 months in a Tennessee jail, then Fargo, on a $100,000 bond she could not pay. Her attorney pulled her bank records. She had been home the whole time. Every charge was dropped on Christmas Eve. She walked out with 4 cents, no coat, near 10 degrees.
Fargo's chief at the time admitted investigative errors, and the city stopped using the system. Her lawsuit says Fargo had no policy, no training, and no supervision for facial recognition. The software never said it was her. It said maybe. Between that maybe and a nationwide warrant, nobody was required to check.
That was the thread all week. Software produces a maybe. A person turns it into a yes or a no. Every story I covered came down to whether that person checked, and whether the system let them.
Photo: Sadettin Dogan via Pexels
What changed
A dispatcher stopped a 146-car hazmat train that software had lined at a maintenance worker. In June near Connell, Washington, the worker held valid track authority, the rail version of lockout tagout. Three Wabtec programs on BNSF's network routed the train onto his section anyway. A dispatcher saw the switch line wrong and took the signal down. The dispatchers union told regulators BNSF knew about the flaw for about a year and the vendor first estimated 14 months to fix it. BNSF says its layers worked as designed.
Photo: AP Vibes via Pexels
A postal carrier who had already lost her house found 7 fake court cases filed against her. Fa'alagilagi Meni-Siliga's Carson, California home was wrecked by a 2020 fire and a storm. She says State Farm delayed the repairs. Her family went bankrupt and lost the house. Her attorney checked the defense's citations: 7 fabricated cases across 8 motions. The attorney admitted using an AI research tool called Irys and never verifying it. Judge Elizabeth Bradley fined her $999.99. California requires a State Bar notice at $1,000. State Farm denies wrongdoing.
A Kansas City psychiatric nurse says software put 2 nurses on a 24-bed unit, and nobody in the building can change it. Lee Barker says Timpani, built by HCA Healthcare with Palantir, staffs off the unit's average census of 16, and the window to change a schedule opens 24 to 48 hours before the shift. HCA says nurse leaders review and edit every schedule and the tool does not make staffing decisions. Both cannot be true. HCA's own material says Timpani uses machine learning to predict staffing needs.
Someone is paid to read what your team typed into ChatGPT. Joseph Cox at 404 Media obtained documents from an OpenAI program codenamed Project Lily. Hundreds of outside contractors read whole conversations and rate the responses. A privacy filter strips names first, and OpenAI's own page says it can miss uncommon identifiers. The setting is improve the model for everyone: on by default for Free, Plus, and Pro, off by default for Enterprise, Business, and Edu. Switching it off is not retroactive. Google and Anthropic disclose human review too.
Outside the channel. On September 15, OpenAI policy chief Chris Lehane told reporters that OpenAI, Anthropic, and Google DeepMind have been working together on AI safety for weeks and do not need an antitrust waiver to do it. The Information reported the three are building a standards body to test models before release, the FINRA-style watchdog Demis Hassabis proposed in July. Lehane also said OpenAI backs a FRONTIER Act provision putting independent verification organizations inside the labs. For now, the outside check on the biggest models is being designed by the people who build them.
Photo: SN.CHE via Pexels
What it means for your business
Every story this week started with software producing a maybe. A facial match. A routed train. A citation. A staffing number. What happened next came down to one person: whether they checked, and whether the system let them. Here is how to put that person back in the room.
Five moves for this week.
Decide now what a match is allowed to do on its own. For any tool that flags, scores, or matches a person, fraud alerts, resume screens, facial or plate matches, write one line: the flag is a lead, not a finding. Name who confirms it against a source the tool did not produce before anything happens to a customer or employee. Bank records cleared Angela Lipps. Nobody pulled them for 5 months.
Find the override on anything that acts without a person, and time it. Scheduling, routing, dispatch, auto-approvals. Ask who inside the building can stop it, how long before it acts the window closes, and when that person last practiced. At BNSF the answer was one dispatcher on the right screen. On Lee Barker's unit he says it is nobody. If your answer is a vendor ticket, that is the finding.
Put the AI question in the engagement letter. Ask outside counsel, your accountant, and your agency in writing: do you use AI on my work, which tool, and who checks the output before it reaches me or a court. The attorney in the Meni-Siliga case believed her tool ran its own citation check. It did not. State Farm did not answer whether it asks its firms. You can.
Read the vendor's warning, not just its promise, and keep the letter. About a month before Medicare's AI prior authorization program launched, the Ohio vendor told the government in writing its software was not fully tested. Medicare launched anyway, and one request sat unanswered 83 days against a 72-hour standard. When a vendor says something is not ready, that letter is the record. File it, and decide who gets to say wait.
Give the tool to your fifth-best person, and keep your best person reading the output. In a Nature Medicine study, 20 doctors judged 100 lung cancer cases with and without an AI. The non-specialists gained the most, and the room agreed with itself more. The same tool spreads whatever it gets wrong to everyone who trusted it. So the expert's job moves from doing the work to sampling it, weekly, in writing.
None of this needs new software. It needs a name next to every maybe, and an hour on Monday.
The longer version of the oversight rule is here: Human-in-the-Loop AI: Keep Your Judgment and Voice.
What to ignore
"Certified." On September 3, Tesla started charging fares for driverless Cybercab rides in Austin. No government agency approved the car. Under federal rule 49 CFR Part 571 the manufacturer certifies its own compliance and the regulator checks afterward, if it checks at all. The Cybercab has no steering wheel, pedals, or mirrors, so Tesla decided the standards written around those parts do not apply, certified itself, and put roughly 1,000 of them on public streets. The audit NHTSA announced September 4 exists to see the engineering behind that call.
That is how most of the AI stack already works. When a vendor's page says compliant, audited, or tested, the vendor wrote that sentence about itself. It changes nothing Monday. Ask who checked besides them, and what they quietly declared out of scope.
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