OpenAI chases Anthropic’s biz customers with zero data retention pledge
OpenAI has unveiled Private Safety Processing, a system designed to let it scan customer interactions with its AI models for safety risks without breaching its Zero Data Retention (ZDR) commitments. The move appears aimed at winning over business customers who prioritise data privacy, positioning OpenAI as offering stronger data protection guarantees than rival Anthropic, which has acknowledged that its own ZDR policy still involves retaining some data for safety purposes.
Anthropic has confirmed that, since 9 June 2026, prompts and outputs from its top-tier models (Mythos 5 and Fable 5) are retained for 30 days under ZDR agreements to support safety reviews, with human staff able to access flagged content via a controlled process. OpenAI's approach differs in that customers using ZDR deployments can control the infrastructure themselves or store data on OpenAI's systems, where encryption keys will soon be customer-controlled; automated systems flag potential misuse without exposing content to OpenAI staff, with human review reserved narrowly for suspected child exploitation material. Full technical details of Private Safety Processing are expected to be published next month.
- OpenAI launches Private Safety Processing to enable safety scanning under zero data retention.
- Move targets business customers wary of Anthropic's 30-day data retention policy.
- Full details of the new system are due next month.
New here? Start with this
Business customers can use generative AI to analyse information, draft content and support other work. Their prompts and the models’ replies may contain confidential material, so companies can set rules about whether a provider stores that data and who can access it.
AI providers also check for harmful or prohibited uses, which can involve reviewing prompts and replies. Zero Data Retention agreements are intended to limit how long a provider keeps customer data, but the details of safety checks and human access can differ between providers.
OpenAI and Anthropic are competing for business customers, including organisations with strict privacy requirements. Their data-handling arrangements matter because companies must weigh privacy protections against the providers’ ability to identify misuse.
Both sides, in good faith
The strongest fair case each way — we don't pick a winner.
The case for
Data minimisation is a foundational privacy principle, and customers choosing ZDR deployments have made an explicit commitment to zero human access. OpenAI's technical approach honours this rigorously: automated systems handle safety screening without exposing content to staff, whilst customer-controlled encryption ensures data remains under customer control. For organisations handling genuinely sensitive information—healthcare records, legal communications, financial data—this meaningful privacy protection builds appropriate trust and eliminates real risks of human access, breach, or policy creep that brief retention models cannot prevent.
The case against
Responsible AI safety requires human judgment that automated systems cannot fully replace. Anthropic's limited 30-day retention with audited human review allows safety experts to catch subtle harms, dangerous patterns, and serious issues like potential child exploitation that automated filtering misses. OpenAI's claim of zero data retention whilst using data for safety purposes is misleading marketing that presents a false choice. Enterprise customers deploying AI should value qualified human expertise reviewing flagged content, not just algorithmic signals, and should understand that meaningful safety sometimes requires limited, transparent human oversight rather than perfect privacy.