An AI-generated likeness of Christopher Pelkey in the video played at sentencing. Video still: Court TV.

An Arizona appeals court threw out a 10½-year manslaughter sentence after a judge relied on a deepfake of the victim. The court ruled Wednesday that the AI-generated video of Christopher Pelkey made Gabriel Horcasitas’s sentencing fundamentally unfair. His conviction stands, but he must be resentenced. Plain Thinking noted the family’s video in May.

Judge Todd Lang called the video “genuine” and treated its message of forgiveness as coming from Pelkey. The appeals court found that it presented the sister’s imagined sentiments as the victim’s own. Actual footage of Pelkey embedded in the presentation was permissible; the invented statements were not. The court ordered resentencing even though defense counsel had not objected when the video played.

Legal Practice

California’s bear flag. Håkan Dahlström / CC BY 2.0.

California is putting rules for lawyers’ use of AI into state law. Starting January 1, SB 574 allows AI assistance but bars lawyers from delegating legal practice to it. Lawyers must disclose generative-AI use in court documents, take reasonable steps to check its output and correct errors, and personally verify every citation in filings. The law also restricts access to confidential information entered into AI systems. That could give firms more reason to run AI on their own computers, where they can keep tighter control over client data.

Newsom also signed a separate workplace-surveillance law restricting employers, including government agencies, from using AI to infer workers’ emotions or collect data from their nervous systems, with exceptions including safety uses.

State Government

Vermont’s State House in Montpelier. Ted Dawson / CC BY-SA 4.0.

Vermont is using AI to find weaknesses in state computer systems, but lawmakers overseeing those systems say they were kept in the dark. A limited group of state employees is using Anthropic’s Claude Code to examine state software for security and operational problems. One member of the legislative IT oversight committee said they learned about the project at a conference, while another said she could not get basic answers about which systems the tool could access. Vermont’s technology department says it restricts the tool’s access and requires human authorization.

In Other News

Maryland’s new criminal laws targeting AI impersonation and fabricated records took effect Thursday. (Maryland General Assembly)

An Ohio appeals court weighed in on a facial-recognition search, siding with prosecutors while leaving the AI questions unresolved. (Ohio Eighth District Court of Appeals)

A lawyer suspended over fake AI citations and dishonesty was also barred from immigration practice. (Board of Immigration Appeals)

Federal civil-rules staff say Rule 11 already covers fake AI citations and recommend no change. (U.S. Courts)

Prosecutors in a Utah county are drawing boundaries for AI’s role in criminal cases. (Cache County Attorney’s Office)

Deputies in Hays County, Texas, tested AI-written police reports, but prosecutors raised concerns and the sheriff’s office declined to seek funding. (KUT)

An Ohio county is spending nearly $200,000 to have AI summarize medical-dispatch calls and flag missed steps. (Cleveland.com)

Seattle is struggling to sort out its own rules for police use of traffic cameras. (KOMO)

New reporting questions how NYPD used Flock without a contract with the company. (New York Daily News)

San Francisco will cut retention of ALPR data on Flock’s platform from a year to 30 days. (CBS News)

Tempe, Ariz., may enact an ordinance banning police use of ALPRs. (Axios)

California is setting up a new fire unit to see whether AI can help fight wildfires. (StateScoop)

Columbus lost its immunity argument against a lawsuit over its 2024 ransomware breach. (Ohio Tenth District Court of Appeals)

Microsoft says attackers wiped out Azure storage accounts in a seven-minute deletion spree. (Microsoft)

Google says hackers have already exploited more security flaws this year than in all of last year. (Google Threat Intelligence Group)

AI experts and law firms were targeted with fake job offers designed to steal access to their accounts. (Proofpoint)

California subpoenaed OpenAI over an unintentional cyberattack on a popular online AI platform earlier this year. (California Department of Justice)

Major AI companies promised stronger safety controls at the White House this week. (News4Jax)

OpenAI is coming out with a new feature aimed at giving customers more control over sensitive data. (OpenAI)

Google is pushing back into the AI race with Gemini 4 Argon, initially available only to selected testers. (Google)

A legal-tech vendor bought an AI company named after the judge you’re still mad at from law school. (Clio)