OpenAI watches for misuse of its models through a system that never has to store customer data. When something trips, what reaches the company is a narrow signal rather than the prompts and answers behind it. The timing lands while Anthropic defends retention of up to thirty days on its most sensitive models.
Key Takeaways
- The system reads exchanges across several linked conversations, where ordinary moderation only ever sees one session.
- Data can stay on customer-controlled infrastructure, or sit with OpenAI encrypted under keys the customer holds.
- A wider rollout and a technical white paper are due in September.
Have an AI Sum Up This Article
ChatGPTA narrow signal instead of the whole conversation
OpenAI watches for problem usage through a mechanism it is previewing with a small set of customers. The point is to run detection without the company holding on to whatever gets inspected. The lab set it out in its announcement on zero data retention for frontier models.
The technical shift is one of scope. Standard moderation judges a single conversation, which lets anyone split a workaround across separate sessions and slip through. This system reads inputs and outputs across linked exchanges, aimed squarely at that pattern.
What travels back when something trips is not the conversation. It is a signal the lab describes as narrowly defined, naming a type of activity without carrying the prompts or the responses. OpenAI then decides whether enforcement is warranted, and the customer can volunteer extra context if it wants the case understood.
Agents do the triage rather than human reviewers. On paper that answers the objection legal teams raise first, the vendor employee reading a confidential file. It moves the question onto the accuracy of automated detection, where the lab has published no figures at all.
Storage follows the same reasoning. Data can remain on infrastructure the customer runs, or live with OpenAI encrypted under keys the customer controls. The company summed the approach up on its official account, framing it as extending zero retention to longer and more autonomous work.
Zero retention itself is not new here. It has been on offer for frontier models for a while, and both labs broadly honour it for enterprise customers. What changes is stretching that regime to cover abuse spread across sessions, which the classic version never reached.
The customers who balked at Anthropic’s thirty days
Anthropic announced in July that it may keep sessions and conversations for up to thirty days on what it calls covered models, the Mythos class and future systems of comparable capability. Several large accounts handling sensitive material took the news badly.
Anthropic does fence that retention in. Human review runs through a controlled access path, a small set of approved reviewers, and every review session lands in a log those reviewers can neither suppress nor edit. It is a serious arrangement, though it rests on a governance promise rather than on data that does not exist, and the company has already let a safeguard sit switched off unnoticed when its biological risk filter stayed inactive for eleven months.
For a buyer the trade reads clearly. One side offers data that lives for thirty days behind documented safeguards, the other offers data the vendor never holds plus a detection layer nobody has seen benchmarked. Both positions defend themselves, and they appeal to different buyers.
This is ground OpenAI has been losing for months. Enterprise adoption leans toward its rival, as we noted when a professional usage index put Anthropic ahead of OpenAI. Winning on confidentiality alone is a way around the fight over model quality.
The financial backdrop makes the move plainer. Anthropic runs at a 65 billion dollar annualised revenue pace, its recent quarterly growth outpaces OpenAI’s, and both are heading for public markets. A compliance differentiator is worth as much as a benchmark point in that setting.
The teams involved already buy seats. The lab has built its professional offer in tiers, and we covered the latest one when ChatGPT Business introduced a Premium seat at 125 dollars. Public sector demand runs the same way, as we measured when ChatGPT took 88 percent of US Congress AI spending.
More articles on Horizon
- Claude Designed Proteins That Worked in the Lab
- Cursor Origin Takes On GitHub During a Six-Hour Outage
- OpenAI Freezes Its Biggest AI Training Run
A sales argument ahead of September’s white paper
What exists today is a preview for selected customers. The broader rollout and the detailed technical document are due in September, leaving a month in which the claim circulates with nobody able to inspect how it holds up.
Rivals now have to answer it. One lab has asserted that abuse detection works without keeping the data, which turns any retention policy into a decision to justify rather than a technical necessity. Anthropic, Google and Mistral sell to the same IT departments under the same compliance rules.
A caveat sits on OpenAI’s own governance. The lab dissolved the internal group that owned its most severe risk work, which we covered when the catastrophic risk team was closed and its remit spread elsewhere. Promising fine-grained detection assumes people to keep it running.
The other side invites the same caution, since that dormant biological filter went unnoticed by everyone for the better part of a year. A safety system is worth what its operation over time is worth, not what its announcement claims.
One question decides the rest. September’s white paper will show whether the narrow signal is enough to qualify an abuse without ever reopening the conversation, or whether the company ends up asking customers to hand over context on every serious case. In that second version, zero retention survives on paper while the burden of proof shifts to the buyer.
Follow the story on Horizon.
