Enterprise customers demand ironclad data guarantees from OpenAI and Anthropic
Major corporations are restricting how their employees can use advanced AI models, imposing zero data retention requirements and air-gapped deployments as distrust over proprietary data handling spreads across the industry.
Anthropic and OpenAI are fielding mounting pressure from large enterprise clients who fear their confidential business information could end up in model training pipelines. Worried organizations are now imposing strict controls on which AI systems their staff can access and for what purposes, according to reporting by The Information. At the heart of the tension lies anxiety that proprietary intellectual property might be absorbed into the models themselves.
The friction intensified following a June policy shift at Anthropic involving its flagship Claude model. Under the revised terms, Anthropic gained the ability to retain customer data, citing the need to monitor for misuse. This change alarmed several enterprises concerned that sensitive operational information would be swept into the company's data collection apparatus.
Both OpenAI and Anthropic maintain that they do not train models on data provided by enterprise contract holders by default. However, the companies do gather metadata from these same customers. The specifics of what metadata entails remain murky, though OpenAI states it uses such information "to better understand how our services are used." Anthropic contends that any usage data it collects undergoes aggregation and anonymization before being excluded from model training.
Telecommunications provider C Spire has negotiated agreements with both firms explicitly barring them from using its data for training purposes. Yet the contracts still permit both companies to harvest C Spire's technical usage information, encompassing details about which applications connect to the AI models and general usage patterns. C Spire also worries that the AI providers might be capturing information about model behavior between response generations.
OpenAI denies using this "chain-of-thought" data for training. Nevertheless, C Spire maintains that the AI companies lack sufficient transparency about their data collection scope, according to The Information's reporting.
Isolation as insurance
Some organizations are sidestepping the issue entirely by deploying their own infrastructure. Aerospace manufacturer Northrop Grumman has opted to run open-source AI models on isolated, air-gapped servers rather than relying on external providers like OpenAI or Anthropic.
Microsoft is capitalizing on these privacy anxieties by marketing its own secure AI offerings. The company's isolated cloud environments execute AI models on private infrastructure with no data transmission to outside AI vendors. This solution carries substantial costs, though at least one customer is reportedly weighing Microsoft's option.
Pharmaceutical firm Novo Nordisk has chosen a middle path, continuing to use Anthropic's Claude for certain applications while imposing a blanket prohibition on feeding any proprietary information into the model.
The erosion of trust is translating into tangible business consequences. A major U.S. utility company scrapped plans to pilot Anthropic's Claude after the company declined to commit to a nonrevocable zero data retention policy. The utility had wanted to assess whether Claude could manage its core power grid operations.
Nvidia has similarly constrained its use of Claude to non-sensitive tasks, citing the absence of zero data retention guarantees. The chipmaker reserves its proprietary in-house AI systems for work it considers too confidential for Anthropic's platform. Nvidia CEO Jensen Huang has previously stated that company employees should consume AI tokens equivalent to half their yearly compensation annually.