The east courtroom at the Howard M. Metzenbaum U.S. Courthouse in Cleveland. Photo: Carol M. Highsmith / Library of Congress.

Federal rulemakers are back to the question of deepfakes and AI conclusions in court. Last year, they were considering two proposed rules — 901(c) for deepfake challenges and 707 for AI-generated conclusions — and sought public comment on Rule 707. In May, they paused both for further study. Plain Thinking examined the proposals in June; now they are back for discussion in October. Rule 901(c) would require evidence to support a deepfake allegation. Once that threshold is met, the party offering the disputed material would have to establish its authenticity by a preponderance of the evidence.

Rule 707, on the other hand, would apply Rule 702/Daubert reliability standards to AI-generated conclusions that would count as expert testimony if a person delivered them, even when offered without an expert witness. Take a phone extraction. Copying its contents is one thing; using AI to identify a key contact from those records is another. That conclusion would need to be tested for reliability. Both proposals return for discussion October 15 in Boston. The public may register to observe in person or remotely.

AI Safety

Photo: Andrew Neel / Unsplash.

OpenAI, the company behind ChatGPT, is uncovering more cases in which its AI went beyond what it was supposed to do. Many of the events now making headlines happened months ago, while the company was training or testing its systems. OpenAI has notified dozens of organizations about the activity, including government agencies and universities. These cases involve AI agents that can use websites and computer tools to carry out tasks on their own.

In June, one agent bypassed restrictions on an Australian government website while looking for health data. Other agents collected public U.S. government data or tried unsuccessfully to access Education Department data. On September 20, another agent got past restrictions meant to keep it off the internet; an automatic shutdown also failed to stop it. OpenAI has since paused some development work on its most capable models while it checks the safeguards.

In Other News

The Supreme Court hasn’t said whether it will hear a case asking if police need a warrant to search license-plate records. (Supreme Court)

Prosecutors played AI-assisted love songs recovered from Caleb Flynn’s phone at his trial for allegedly killing his wife. (404 Media)

A lawyer who thought enterprise AI would not hallucinate was removed from a case; his firm must pay up to $10,000 in opposing fees and costs. (Aguilar v. The Crawford Group)

Five Indianapolis police officers face criminal charges for allegedly misusing Flock’s license-plate-reader system. (Indiana Public Radio)

Connecticut’s police standards council recommends deleting license-plate data after seven days, stricter than the 21-day limit taking effect October 1. (Connecticut Mirror)

Los Angeles is considering a 10-year, $235 million contract with Axon that includes patrol-car license plate readers. (Los Angeles Daily News / Officer.com)

Milwaukee police are weighing generative AI under a policy requiring officers to disclose when they use it in their written work. (Wisconsin Watch)

Oregon’s governor ordered a plan for independent safety reviews of frontier AI used by state government. (Oregon Governor’s Office)

Twenty-six attorneys general urged Congress to require AI safety testing and preserve state enforcement powers. (Oregon Department of Justice)

Florida’s attorney general asked a court to stop OpenAI from developing new AI models without outside oversight; the request comes in a lawsuit alleging ChatGPT harms children. (Reuters)

A VA whistleblower says an AI tool meant to speed disability claims made them take longer. (Human Resources Director)

A divided D.C. Circuit upheld the Pentagon’s designation of Anthropic as a supply-chain risk. (Wired)

Nvidia announced a system that uses a separate chip to stop AI agents from acting beyond their permissions. (Nvidia)