The OpenAI logo on a phone resting on a laptop. Photo: Levart_Photographer / Unsplash.
OpenAI is putting a process around telling the public when it finds examples of its AI misbehaving. Its new disclosure framework explains which examples qualify, how the company will investigate them and what each report will include. OpenAI says it may publish before it has fully explained the behavior or developed a fix. In one training example, agents working together on a spreadsheet could not share it through the intended local filesystem. After local transfer attempts failed, one agent uploaded the file to a public hosting service even though the task called for local files only. The six initial reports cover incidents observed during training or evaluation, and OpenAI says they should not be read as evidence of how often misalignment occurs across its models. Anthropic already publishes broader risk reports and model-specific system cards. What OpenAI is adding is a more explicit process for reporting individual cases: what happened, what remains uncertain and, when available, what the company is doing about it.
Police Technology
A Metropolitan Police vehicle in Washington, D.C., December 2023. Photo: Raymond Wambsgans, CC BY 2.0.
D.C. police used Flock license-plate-reader data to investigate one of their own officers. Now the police union wants more oversight of the camera network. Its argument: Metropolitan Police Department policy permits plate-data inquiries when there is a legal basis for them or reasonable suspicion of criminal activity, and requires an official law-enforcement purpose. But Internal Affairs used the data in what the union describes as an administrative investigation. The department says the use was appropriate. On the day it denied the union’s grievance, the department reportedly required officers to route Flock requests through the Real Time Crime Center, while Internal Affairs retained access. The dispute is headed to arbitration. The police union is now asking what rules apply when police surveillance is turned inward.
Number of the Day
49%
Share of Ninth Circuit appeals filed without a lawyer in 2025, according to Judge Sidney Thomas. The figure measures self-representation, not AI use.
In Other News
Waymo stopped a ride and called police after detecting a firearm, leading to two teenagers’ arrests. (The Verge)
The U.S. House of Representatives passed the GUARD Act, now headed to the Senate, to let local investigators use existing federal grants for financial fraud cases. (The Record)
A legal software company hired former Michigan Chief Justice Bridget Mary McCormack to lead its judiciary business. (Clio)
The federal judiciary is developing further AI guidance, while three appeals courts already use the technology to check briefs for procedural and formatting problems. (Reuters)
California will require disclosures in ads starring realistic people invented by AI. (California Legislature)
OpenAI introduced Astra for Law, initially available to selected law firms, pairing its latest model with a legal-search index and instructions for legal analysis. (OpenAI)
The FAA is preparing to launch AI-assisted traffic planning in Washington as soon as Monday. (Wall Street Journal / FAA)
Major corporate customers are pushing back on Anthropic retaining their prompts for safety monitoring. (Reuters / The Information)
Perplexity is making local AI easier to use on compatible Windows PCs. (Nvidia)
About Plain Thinking
Grant Morgan is an attorney, former prosecutor and former Tribune Publishing reporter who covered Gary, Indiana, and Chicago’s northern suburbs. Plain Thinking covers AI and consequential technology affecting lawyers, police and courts.
Every Tuesday and Friday · Get Plain Thinking by email · LinkedIn
If an issue includes paid placements, they’ll be clearly labeled. Companies don’t pay for editorial inclusion.
