California’s new partnership with Anthropic sparks questions about security and state data
SACRAMENTO, Calif. —
California’s new government partnership with artificial intelligence company Anthropic is raising questions about data security and the cost to taxpayers as state agencies begin using the company’s tools.
Gov. Gavin Newsom announced the partnership June 29, which allows state agencies, cities and counties to purchase and use Anthropic’s chatbot Claude at a 50% discount, along with free workforce training and technical support from the AI company. According to the governor’s office, the tools are meant to improve government efficiency by assisting with tasks such as drafting and summarizing documents and analyzing information.
More than two months after the announcement and amid several new warnings around the evolving technology, KCRA 3 struggled to get California’s government leaders to do an on the record interview about the partnership. The questions come as Gov. Gavin Newsom is slated to announce executive action Friday related to AI safety.
This week, the California Department of Technology would only respond to questions on the record if they were sent via email. KCRA 3 also tried to interview the two lawmakers who are in charge of the privacy and data committees at the state capitol, Assemblymember Rebecca Bauer-Kahan and State Sen. Chris Cabaldon. Neither were willing to talk about it.
Anthropic did not respond to a request for comment.
Which state agencies are using Claude?
In an email, a spokesperson for the California Department of Technology said the partnership is still in its early stages, and several state agencies have used it. That includes the California Department of General Services on behalf of Government Operations Agency, the state’s Fair Political Practices Commission, the Public Entity Risk Management Authority, and the statewide financial system known as FI$Cal.
At the local level, the city of Gilroy, Southern California Regional Rail Authority, city of Corona, city of Agoura Hills, city of Union City, County of Santa Cruz and Kaweah Health have also used it.
The Department of Technology said 60 other state departments or local agencies have expressed interest in using Anthropic’s services but would not say which ones exactly, citing contract negotiations that may be underway.
How much will this cost?
The agreement with Anthropic is considered a “zero-dollar contract”, meaning the state government as a whole isn’t paying for it. Each state agency, city, or county negotiates and pays individually for the tools and use, so it’s not clear how much the partnership could end up costing taxpayers.
According to a spokesman for the California Government Operations Agency, the partnership allows for a rate of $10 per individual user per month. There are also usage charges based on how much text Claude processes, billed per “tokens” or chunks of texts. Staff has the choice between Claude Sonnet, which charges $3 per million input tokens and $15 per output tokens, or Claude Haiku, which charges $1 per million input tokens and $5 per million output tokens.
The agency spokesman said six staff members in GovOps started using Claude in August of 2026. The agency is monitoring usage costs to determine the ongoing business value, he said.
Will Claude keep state data safe?
The type of information that state and local workers can put into Claude will depend on the agency. The Department of Technology said some examples include annual reports, large volumes of scientific research or documents related to policy and compliance. Data that includes personally identifying information is generally prohibited from being used, the spokesperson said.
The department said statewide policies, individual department policies and the contract security terms add extra layers of security to keep state information safe. The Department of Technology said the agreement between the state and Anthropic explicitly prohibits Anthropic and Claude from using state data to train or improve their models.
“It is like going to a hotel and the hotel promising you, ‘hey, we won’t look through your luggage,’ instead of storing your luggage in a vault that the hotel doesn’t have a key to access,” said Ansgar Grunseid, CEO of Confidential AI, a San Francisco-based startup specializing in secure AI use.
He noted that similar agreements or promises have failed in the past, citing OpenAI’s breach of a data promise to some of its customers due to a lawsuit with The New York Times. He said technically and legally, Anthropic cannot guarantee the information put into Claude will be secure or not processed in other ways.
“This is not a technical means, this is a promise, and we’ve seen over and over these promises fail,” Grunseid said. “All these systems are controlled by software, software has attacks, bugs, vulnerabilities, human error, misconfiguration.”
“What’s interesting about this partnership with California and Anthropic is California is saying we will allow state employees to put a lot of state data in Anthropic, and that’s putting a lot of trust in Anthropic,” he said.
KCRA 3 also asked the Department of Technology who will be legally responsible if a state employee does something based on bad information from Claude. A spokesperson said the state government will be responsible for oversight and verification.
Stefano Bellasio, CEO of a workplace AI startup, Anthropos, said the partnership could be beneficial despite concerns around AI, including those flagged by Anthropic’s own leader.
“I think it’s actually smart,” Bellasio said of California’s agreement with Anthropic. “We want these people to be a lot more productive, we want them to use AI, upskill themselves with AI. That’s going to be the future. And I ultimately think it’s going to be very good for the taxpayers.”
The Department of Technology said Anthropic will not be the only company California works with, stating that the state has other pending requests for Generative AI services.
What does Anthropic get out of this?
When KCRA 3 asked the Department of Technology this question, a spokesperson said, “Anthropic is recouping costs for Claude access as outlined in their cost sheets.”
“If you’re Anthropic, what you want is you want to extend your tendrils and see as much data as possible and see as much use as possible,” Grunseid said. “Obviously, there’s the very simple, shortsighted goal of revenue. More dollars equals better. And so capturing an entire state at a time in terms of their usage, even with that 50% discount, it matters a lot.”
“There’s tremendous momentum in getting your tendrils into the state of California and getting all the employees used to using a certain set of tools and processes,” he said.
See more coverage of top California stories here | Download our app | Subscribe to our morning newsletter | Find us on YouTube here and subscribe to our channel.