Reporting that holds elected officials to account can take months, and its effects are often hard to see. In September 2025, San Francisco Public Press reporter Jason Winshell found that California’s first AI safety law, SB 53, traded away safety and liability provisions in the weeks before it passed. Earlier this month, that reporting surfaced in a congressional race to succeed Speaker Emerita Nancy Pelosi.
In an Oct. 1 debate hosted by KQED and the San Francisco Chronicle, Supervisor Connie Chan cited the Public Press by name in a challenge to state Sen. Scott Wiener, the law’s author.
Candidates rarely call out our work directly. In April, Saikat Chakrabarti, who ran in the June Democratic primary, referred to our reporting on social media, likely drawing Chan’s attention to the issue.
The acknowledgement of our investigative reporting was gratifying. But the exchange that followed was frustrating to watch. Chan had 30 seconds for her direct question to Wiener:
“San Francisco Public Press mentioned your AI watered-down bill, which is also known as SB 53. When they recognized — they actually categorized that bill as traded away safety and liability provision — and that, at that time, they asked you for the records about who amended that bill that resulted in the watered-down bill for that SB 53. And subsequently you also have seen that OpenAI and Anthropic employees, workers have donated significantly to your campaign. Will you now release those records?”
That question carried three claims: The law was weakened, Wiener withheld records and AI money backed his campaign. Wiener answered with three denials, and the specifics behind each claim never entered the exchange.
Chan raised our reporting again at an Oct. 13 debate hosted by Mission Local, The New York Times and City Arts & Lectures, asking Wiener “when San Francisco Public Press asked you who amended the bill with you, and you wouldn’t provide that record,” whether he could be trusted to regulate AI with integrity. Wiener did not address the records, but he did deny her accusation that donations from AI companies indirectly support his candidacy. He replied, “There is no Anthropic money in this race. Period.”
All these issues deserve a closer look.
‘Watered down’
Wiener opened his answer to rebut the first claim, addressing none of the law’s actual provisions:
“What I will say is, to refer to SB 53 as watered down is a joke. It was the first and strongest AI safety regulation passed in this country. Period. Full stop.”
He pointed to the law’s influence and to its opponents:
“Multiple states have literally cut and paste this law and then built on it, which is great.”
“Donald Trump doesn’t think it’s watered down. Other states don’t think it’s watered down. All the tech companies that try to undermine and sabotage SB 53 don’t think it’s watered down.”
Then he described an earlier bill he introduced, which Gov. Gavin Newsom vetoed:
“The governor vetoed the first AI safety regulation that I put on his desk that had a kill switch and liability and whistleblower protections and safety testing requirements and so forth.”
Our reporting picked up the story after that failed bill, documenting how SB 53’s regulatory reach was pared back late last summer and early fall, before Newsom signed the heavily revised law.
The maximum fine for a catastrophic incident fell from $10 million to $1 million. A requirement for independent audits of companies’ safety plans was deleted. Companies could now redact trade secrets from published safety plans, and the state could not review the redactions. An earlier draft required companies to report hacks that could lead to safety risks, but the final law required a report only after death, injury or catastrophic damage. Winshell’s December 2025 follow-up found that late amendments limited whistleblower protections to a small subset of employees responsible for assessing safety risks, potentially leaving others with relevant information exposed to retribution.
And then it happened: In July, OpenAI models broke out of a test environment and breached the servers of the AI company Hugging Face.
Winshell said recently that in his comparison of the bill’s drafts, an earlier version would likely have required OpenAI to report it. A July 2025 draft required companies to test even internally used models for “dangerous capabilities,” including the ability to carry out a cyberattack. In that version, the first time a model crossed such a threshold, the company would have had to notify the attorney general within 15 days, whether or not anyone was harmed. Amendments on Sept. 5, 2025, deleted the requirement.
At an Aug. 10 Assembly hearing, Jonathan Snow, a deputy director at the state Office of Emergency Services, testified that OpenAI “did not meet the threshold” for reporting under SB 53. Because the July draft let each company set its own thresholds, a report would have been likely, but not certain.
The national regulatory environment is dynamic, but it is not necessarily becoming stricter. Our March story found that New York replaced its stronger RAISE Act with large sections of SB 53, nearly word for word. The revision cut penalties by 90% and removed a ban on releasing potentially unsafe models.
Wiener’s defense in the debates rests on firsts, and on who opposed the law. In describing the vetoed bill, he touted liability and whistleblower protections, two areas where our reporting found SB 53 had been weakened. Chan’s charge in the Oct. 1, 2026, debate that the law was watered down came across as a matter of opinion. Our reporting points to the bill’s amendment history, which documents it.
The records
Wiener’s answer to Chan’s Oct. 1 question about our unfilled records request was one sentence:
“I have no idea what records my opponent is referring to.”
We filed a formal records request with Wiener’s office. His office forwarded it to the Legislature. The Senate Rules Committee released letters from outside groups, but it withheld communications with Wiener’s office, citing the Legislative Open Records Act. That law allows the Legislature to withhold such records but does not require it. Separately, we asked his staff who was involved in individual line edits of the bill. A spokesperson replied, “Unfortunately we are unable to answer these questions.”
Wiener may not have remembered a request his office passed along last November. Still, he could have voluntarily released those records at any time. In the debate, his denial stood unchallenged, and viewers had no way to verify the request had ever reached his office.
The money
Wiener’s answer never addressed the tech industry donations Chan raised. He closed with a jab:
“It’s very easy when you’ve never lifted a finger for AI safety regulation to armchair quarterback it.”
The moderators gave Chan time to respond. She spent most of it on her own record in San Francisco, then returned to the money:
“This is the moment: Can we trust people who are bankrolled by the AI industries and continue to write legislation to regulate —”
Wiener cut in: “And that is false.” The moderators moved on to health care.
Mission Local reported in September that Wiener had received hundreds of thousands of dollars from workers at OpenAI and Anthropic. Wiener told the outlet that many of those donors had “split from their bosses” to support AI safety.
Public First Action, a group Anthropic has funded with $40 million, has given $3.5 million to Abundant Future, a super PAC supporting Wiener, the San Francisco Standard reported. Public First says none of that money came from Anthropic.
“Bankrolled” is a charged word, and it gave Wiener room for the above four-word denial. The narrower question in Chan’s opening, about donations from AI company employees, never got an answer.
Wiener and Chan continue to battle over the AI issue in a blizzard of home-delivered fliers, making the same claims raised in their live exchanges, talking past each other.
What remains open
The question at the center of our reporting is still unanswered: Who outside the Legislature asked for each amendment to SB 53? A 30-second debate question was never going to settle it. We will keep asking. Read our full coverage of artificial intelligence.
