OpenAI’s Latest Bid to Fight Anthropic: A Promise Not to Keep Customer Data

OpenAI promised not to retain data from businesses using its artificial-intelligence models while increasing the safety of its products, a bid to snatch customers from rival Anthropic who are upset the maker of Claude is keeping their information.

The maker of ChatGPT said Wednesday that it is previewing new technology to some customers that lets it detect patterns and safety risks across multiple interactions with its models. The goal is to move beyond previous AI safety systems that monitored interactions individually, which made it difficult to ensure tools were being used properly without retaining some customer data.

The pledge stands in contrast to Anthropic’s policy of keeping customer data for 30 days to ensure the safety of its latest models. The move has drawn criticism from tech leaders including White House AI adviser and venture capitalist David Sacks, Microsoft CEO Satya Nadella and Palantir’s Alex Karp, who have argued that businesses need certainty that AI companies aren’t keeping their data, particularly in regulated industries such as healthcare and finance.

“We’ve heard very loud and clear from businesses that this is important. They often have their own commitments that they have made to their customers if you think about somebody who is serving other enterprises, who are trusting them with incredibly sensitive data,” said Aleah Houze, OpenAI’s head of product policy.

The move highlights OpenAI’s push to catch up to Anthropic in the AI race while mollifying critics who fear the company’s tools don’t have adequate safeguards to prevent cyberattacks or other harms caused by its models going rogue. The company recently told investors that its second-quarter revenue growth lagged behind Anthropic’s, disappointing some shareholders, The Wall Street Journal reported.

OpenAI said customer data would stay on infrastructure controlled by the customer. It is also working on a system in which OpenAI could store the information with encrypted keys owned by the customer, meaning its employees wouldn’t see model prompts or responses or other sensitive information.

The new system sends OpenAI a signal when it identifies risks like a cyberattack or biological weapon based on model activity. The signal helps determine whether an action by the company is needed. Customers can appeal those decisions or share more information with OpenAI if desired. OpenAI only keeps the high-level signal detailing the severity and category of the risk.

Anthropic’s popularity with businesses that pay to use its coding tools and advanced models has vaulted the company ahead in the AI race, but some of the company’s policies have fueled a backlash in Silicon Valley and the Trump administration.

Sacks said in June that retaining customer data is “creating a new level of AI haves and have-nots,” and fueling outrage among software developers using the tools.

The issue is related to a debate about open-weight models that allow users to download the numerical parameters that determine their behavior. Many tech companies have endorsed such models, even those made in China, and asked the Trump administration not to place restrictions on them. Anthropic has warned that the tools pose safety risks because they often don’t have the same safeguards as the most powerful closed models.

OpenAI recently signed an industry letter in favor of open-weight models and makes some of its own, though its most powerful tools are closed. The company has had talks with administration officials about potential restrictions on Chinese open-weight model providers that are accused of unfairly using advances made by U.S. companies in a process called distillation.

News Corp, owner of The Wall Street Journal, has a content-licensing partnership with OpenAI.