Nvidia, Palantir, and others restrict advanced AI model usage over privacy concerns, report claims — 'paranoia' rising over customer intellectual property

Several major AI companies, including Anthropic and OpenAI, are facing scrutiny from their large customers over concerns regarding the use of proprietary data to train AI models. Some companies are demanding assurances about data handling or limiting which models employees can use for specific tasks due to fears that their intellectual property may be included in model training.

This issue emerged after Anthropic modified its Fable model policies in June, allowing the company to retain customer data to prevent misuse. Although both OpenAI and Anthropic claim not to train models with specific enterprise data by default, both collect metadata, which may include information on how customers use their products.

This data is anonymized, but there remains uncertainty over what is collected and used for training. Some customers, like telecom company C Spire, have contracts that prevent AI companies from using their data to train models, but still allow data collection on technical usage. OpenAI and Anthropic have been criticized for not being clear about what data they collect.

Some companies, like aerospace firm Northrop Grumman and pharmaceutical company Novo Nordisk, have opted for air-gapped servers or banned proprietary data use to address privacy concerns. Microsoft offers a secure AI platform with isolated cloud environments to address these concerns, but it is costly. Novo Nordisk continues to use Anthropic's Claude for certain tasks but has a ban on using proprietary data with the model.

One U.S. utility company canceled plans to test Anthropic's Fable due to Anthropic's refusal to agree to a nonrevocable zero data retention (ZDR) policy, causing Nvidia to switch to its in-house AI solution for sensitive tasks.