Some of the largest AI buyers are keeping their own code and data away from Anthropic's most capable model because of what the company may retain. Anthropic announced replacement terms on September 1, and The Information reported the limits two weeks later.
A company's memory includes the copies its software vendors keep. In June, Anthropic, the maker of the Claude models, gave itself the right to retain customer usage logs for 30 days for its most capable models, including Fable, 'and up to two years if its safety systems flag them.' On Monday, Laura Bratton and Aaron Holmes (The Information) reported that Nvidia, Booz Allen Hamilton and Palantir were limiting where Fable may be used as a result. (PYMNTS, Sep 15)
What's new: Enterprise buyers have long chosen models mainly on capability, speed and price. These buyers are limiting a frontier model over what its maker keeps, and PYMNTS reports that 'data retention terms are becoming as decisive as model performance in determining which AI systems large enterprises deploy.' For a company, who holds the copy of its work is now a buying criterion of its own.
Why it matters: The objections are about each company's own work product. Bill Vass, chief technology officer of Booz Allen, said 'we worry a little bit that [Fable] might be learning from some of our code,' and added that the concern covers a small share of the firm's work. Justin Boitano, Nvidia's vice president of enterprise AI, said zero data retention 'should be on by default.' Nvidia, an Anthropic investor, uses its own Nemotron models for sensitive work such as supply chain monitoring, according to the report. Anthropic and OpenAI both say they don't train on enterprise customer content by default.
Between the lines: Anthropic had already changed course. On September 1 it announced Enterprise Frontier Safeguards, under which customers keep retained data in their own cloud under their own encryption keys while Anthropic's automated misuse detection keeps running, and the customer's security team reviews anything flagged. (Anthropic, Sep 1) Customers still have to apply for zero data retention, and the announcement says it is rolling out in phases. (The Register, Sep 2)
The opening: Retention terms are now something vendors sell. Salesforce launched AIforce on Tuesday with zero data retention for frontier models as a listed feature. (Constellation Research, Sep 15) A US utility told the article it dropped a planned test of Fable on core power systems over the same assurance, while still using Anthropic for finance and human resources work. WitnessAI, a company that sells AI governance software, says a company 'remains responsible for governing that data even when it is retained by an AI vendor.' (WitnessAI, Sep 8)