AI firms sharpen privacy pitch as enterprises demand control over data

Here is how OpenAI, Anthropic and Google are approaching data privacy as businesses demand greater control over sensitive information shared with increasingly capable AI models

As businesses use artificial intelligence (AI) models to process increasingly sensitive information, questions over data retention, privacy and safety are becoming harder to ignore. AI firms are now trying to strike a balance between monitoring models for misuse and giving enterprise customers greater control over their data, with providers adopting different approaches to storage, training and safety checks.

 

Among the latest is from OpenAI, which has introduced Private Safety Processing (PSP) alongside its Zero Data Retention (ZDR) policy for eligible application programming interface (API) customers. According to the US-based AI entity, the system is designed to detect potential safety risks across multiple interactions while keeping customer prompts and responses inaccessible to OpenAI employees.

 

Anthropic, meanwhile, requires 30-day retention for certain covered models for safety monitoring, while Google has outlined specific conditions under which its Gemini services can offer zero data retention.

 

To understand why the AI firms’ approach differ when it comes to privacy, it is important to first know what happens to data after an AI request is made.

 

When a company uses an AI model through an API, its software typically sends an input to the model. This could be a question, document, piece of code or structured business information. The model processes that input and generates an output.

 

There can be several stages at which data may exist: while a request is being processed, in temporary logs or caches, for abuse monitoring and debugging and, depending on the provider and customer's agreement, potentially for longer-term storage.

 

These purposes should not be confused with model training. Data being temporarily retained to provide or secure a service does not automatically mean it is being added to the training data used to build a future model.

 

Similarly, a company saying it does not use customer data for training does not necessarily mean the data disappears immediately after a response is generated. That distinction is increasingly important as AI providers publish more detailed policies. 

 

OpenAI puts Zero Data Retention in focus

 

OpenAI's latest announcement is built around this distinction between privacy and safety.

 

The company says eligible API customers using ZDR do not have their prompts or model responses retained after processing, and customer content is not available to OpenAI personnel for review. It also says enterprise customer data is not used to train its models unless the customer explicitly opts in.

 

OpenAI is nevertheless introducing Private Safety Processing because some risks cannot be identified by looking at a single interaction.

 

For example, a series of apparently harmless requests could collectively reveal an attempt to bypass safeguards or conduct an attack. OpenAI says its new system is designed to identify such patterns across related interactions without giving its personnel access to the underlying customer content.

 

Under the proposed system, automated safety systems can identify potential misuse and send OpenAI a limited safety signal indicating the type of activity involved. The underlying prompts and responses are not made available to OpenAI personnel.

 

Customers can also investigate alerts using information in their own systems and choose to share information with OpenAI if they want to appeal an enforcement decision or assist with an investigation.

 

This is significant because AI providers face a difficult balance. The more capable models become, the more important monitoring can become. But greater monitoring can also require access to more information.

 

OpenAI says Private Safety Processing is currently being tested with early customers, with a rollout and technical white paper planned for September. 

 

Paying for AI means paying for usage

 

For enterprise customers, using an AI model through an API is different from subscribing to a consumer chatbot. API services are typically priced according to the number of tokens processed.

 

A token is a unit of text that an AI model reads or generates and can represent a whole word, part of a word or, in some cases, multiple characters. The amount and type of tokens processed therefore determine how much a customer pays for using the model.

 

When a company sends a request, the input consumes tokens. The model's response also consumes tokens. The provider then calculates the cost based on the number and type of tokens processed, with pricing varying by model and sometimes between input and output.

 

Put simply, a company using an API is paying for the amount of AI processing it consumes.

 

OpenAI's recent pricing announcement around GPT-5.6 illustrates how this model is evolving. The company said improvements in model efficiency, inference systems and context management allow customers to get more useful work from the same amount of computing resources while lowering the cost of some models.

 

This is important to the broader data debate because customer data and payment are separate parts of the relationship. A customer paying for API usage does not automatically give the provider the right to use that customer's data for training.

 

Data used for training should be governed by the provider's terms, privacy commitments and the customer's agreement. It should not be described as a form of payment unless a company explicitly structures its service that way.

 

For enterprise customers, the more relevant question is not just what they are paying for, but also what the AI provider is permitted to do with the data they share.

 

Anthropic takes a different approach

 

Anthropic's recent policy update shows why there is no single industry-wide definition of zero data retention.

 

In July, Anthropic said it would require limited data retention and review for certain covered models, including Mythos-class models and future models with similar capabilities. Prompts and outputs for these models are retained for 30 days to support safety work.

 

The reason is similar to OpenAI's explanation: some forms of misuse only become visible when multiple requests are examined together. Anthropic gives the example of repeated attempts to bypass safeguards, where hundreds of slightly different prompts may be used to find one that succeeds.

 

Anthropic says that, by default, its personnel cannot read retained conversations. Human review is possible through a controlled process when content is flagged by automated safety systems, and such access is recorded. After 30 days, the data is automatically deleted except in cases involving flagged material or legal requirements.

 

The policy applies specifically to organisations that had configured ZDR for the covered models. Anthropic notes that customers using these models through its API, Amazon Bedrock, Google Cloud Agent Platform or Microsoft Foundry may need to enable retention, depending on the setup.

 

This approach highlights an important point: zero data retention can come with exceptions when providers believe additional retention is necessary for safety.

 

Google's conditions around zero retention

 

Google Cloud's Gemini Enterprise Agent Platform documentation takes another detailed approach. Google says it will not use customer data to train or fine-tune AI or machine-learning models without prior permission or instruction.

 

At the same time, the company's documentation explains that achieving zero data retention can require customers to configure specific features carefully.

 

For example, Google may log prompts for abuse monitoring in certain circumstances. Some grounding features also involve limited storage. Grounding with Google Search can involve storing certain derived queries for up to three days, while Grounding with Google Maps can retain prompts, contextual information and outputs for 30 days for reliability engineering.

 

Google also says its published Gemini models may use in-memory caching by default to improve performance. This data is isolated at the project level, has a 24-hour time-to-live and customers can disable the caching feature.

 

Once again, zero data retention needs to be understood in the context of the particular model, feature and configuration being used.

 

Why companies are spelling it out now

 

The growing attention to data retention is partly a consequence of AI moving deeper into enterprise systems.

 

A chatbot answering general questions may not handle particularly sensitive information. But an AI coding assistant could receive proprietary source code. A financial application could process internal forecasts. A healthcare system could handle sensitive records. An agent could potentially access multiple internal systems while completing a task.

 

The stakes therefore become much higher. OpenAI itself points to the sensitivity of enterprise data, citing financial records, health data, confidential business plans and proprietary research among the information handled by its customers.

 

At the same time, AI companies are becoming more concerned about the capabilities of their models. OpenAI recently said it had temporarily slowed some model development while strengthening monitoring, alignment and security safeguards. It said increasingly capable models require stronger controls across training and deployment.

 

This creates the central tension: companies want AI providers to protect their data, while providers want enough visibility to detect serious misuse.

 

The question goes beyond 'is my data stored?'

 

For businesses adopting AI, the data question is becoming more detailed than simply asking whether information is retained. They need to know how long it is retained, where it is stored, who can access it, whether it is used for training, whether safety monitoring creates exceptions, whether caching is enabled and whether those settings change depending on the model or feature.

 

OpenAI's ZDR policy, Anthropic's 30-day retention requirement for covered models and Google's detailed zero-retention documentation show that providers are moving towards more explicit answers on how enterprise data is handled.

 

For customers, that clarity may become as important as model accuracy or price. The AI industry is moving towards a system where intelligence is increasingly bought and consumed as a metered service. But alongside the price per token, another consideration is becoming important: the cost of giving up control over data.

 

For enterprise customers, the choice of an AI model may therefore increasingly depend not only on how well it answers a question, but also on what happens to the question after the answer is delivered.