OpenAI previewed a new system called Private Safety Processing on August 19, 2026. It’s an automated abuse-detection layer that scans patterns across related conversations without giving OpenAI staff access to the underlying content, while staying compatible with the company’s Zero Data Retention policy. TechCrunch framed the move bluntly: OpenAI is trying to one-up Anthropic on customer privacy.
I pay for both ChatGPT and Claude every month and route real client work through each. So a ChatGPT vs Claude privacy comparison isn’t an abstract policy question for me — it’s the difference between which tool gets which kind of draft. It’s also a question worth a durable answer, rather than a one-week reaction. The underlying trade-off between the two tools will still matter long after this specific announcement is old news, which is why this post is built as a reference to revisit, not just a recap of one day’s headline.
In this article
- What OpenAI actually announced on August 19
- ChatGPT vs Claude: the privacy default behind each chat
- How Claude’s retention rules actually work
- How ChatGPT’s retention rules actually work
- Where Private Safety Processing changes the ChatGPT vs Claude gap
- What this means for client data in practice
- Where each tool still falls short in this ChatGPT vs Claude matchup
- FAQ
What OpenAI Actually Announced on August 19
Private Safety Processing is built to identify patterns of misuse across a user’s related interactions. That’s the kind of signal that normally requires a human reviewer to read message content, but it’s designed to work without exposing that content to OpenAI personnel. The system is designed to stay compatible with Zero Data Retention deployments, meaning enterprise and API customers who’ve opted into ZDR don’t lose that guarantee just because a new safety layer got added on top of it. OpenAI has said a broader rollout and a technical white paper are coming in September 2026. That means the version previewed this week is explicitly a preview, not a finished feature.
The framing matters as much as the mechanism, for anyone tracking the ChatGPT vs Claude race generally. Anthropic has spent much of the past year building its public identity partly around safety and interpretability research. A system that lets OpenAI claim comparable privacy-preserving abuse detection is a direct answer to that positioning, not a response to a specific incident or regulatory demand that’s been reported.
I wrote about ChatGPT’s privacy refresh back in May, and settled on four small workflow changes for handling client drafts. Private Safety Processing is a different kind of update. It’s not a settings toggle I flip. It’s an infrastructure change to how OpenAI’s abuse detection pipeline works underneath every conversation, whether or not I ever notice it running. That’s a meaningful category difference from most “privacy update” stories, which usually just mean a new checkbox in account settings. It’s exactly the kind of structural shift a running ChatGPT vs Claude comparison needs to account for, since it changes how the tools work, not just what they claim.
ChatGPT vs Claude Privacy: The Default Behind Each Chat
The single biggest structural difference between the two products is what happens by default, the moment you start typing. On ChatGPT, chats are saved indefinitely to your account until you delete them yourself. OpenAI is currently the only major provider that lets you keep that history without being automatically enrolled in model training. On Claude, the opposite trade-off is built in. As of the September 2025 update to Anthropic’s consumer terms, Claude trains on your conversations by default, and keeping that default off requires manually opting out.
That’s a genuinely different philosophy, not just a different toggle position. ChatGPT separates “do I keep my history” from “do I consent to training.” Claude and Google’s consumer tools generally bundle those two questions into a single choice: keep your history and accept training, or opt out of both.
Gemini follows the same bundled pattern as Claude, rather than ChatGPT’s split model, for what it’s worth. Across the three tools I use weekly, OpenAI is the outlier on this specific design decision, not the norm. That’s worth knowing before assuming any one provider’s approach represents an industry standard. A ChatGPT vs Claude privacy comparison is really a comparison between one company’s unusual choice and the more common bundled-consent approach everyone else has converged on.
Neither model is objectively “more private” once you actually sit with the trade-off. ChatGPT’s approach means a user who never deletes anything builds an ever-growing, permanently retained archive — one that’s exempt from training, but not from existing indefinitely on OpenAI’s servers. Claude’s approach means a user who doesn’t actively opt out gets a shorter retention window, in exchange for having their conversations feed model training for up to five years in de-identified form. Which one you’d rather live with depends entirely on which risk actually worries you: a permanent archive, or training exposure. Reasonable people land in different places on that question.
How Claude’s Retention Rules Actually Work
Anthropic introduced an opt-in toggle — “You can help improve Claude” — for Free, Pro, and Max users in August 2025, giving people until October 8, 2025 to make an active choice. Opt in, and Anthropic can retain de-identified conversations for up to five years and use them for model training. Opt out, and the older 30-day retention window applies, with no training use at all.
Privacy researchers have flagged the interface itself as a concern. The September 2025 rollout used a large, prominent “Accept” button with the training toggle pre-set to “On” — a pattern critics have compared to a manipulative dark pattern that risks running afoul of GDPR-style consent standards. That criticism is the piece of the Claude side of this ChatGPT vs Claude comparison that Private Safety Processing does nothing to address. It’s a consent-interface problem, not a data-processing one. No amount of OpenAI catching up on abuse detection changes how Anthropic’s own toggle is presented to a new user.
On the API side, the picture is cleaner. Anthropic explicitly excludes API data from training, and the five-year consumer retention rule doesn’t apply there at all. Notably, Claude Opus 5’s own August 2026 launch dropped the 30-day retention requirement that still applies to Anthropic’s Fable and Mythos-tier models. That means retention policy now varies even within Claude’s own model lineup, not just between opted-in and opted-out consumer accounts. Anyone doing a serious ChatGPT vs Claude comparison for their own work needs to check which specific model tier they’re actually using, not just which company — the answer isn’t uniform even within one product family anymore.
How ChatGPT’s Retention Rules Actually Work
OpenAI’s consumer default keeps chat history indefinitely, with deletion left entirely to the user. Turn history off, and a conversation is retained for roughly 30 days for abuse review before deletion — though legal holds and active abuse investigations can extend that window. Business and Enterprise customers get a stronger baseline. OpenAI states it does not use ChatGPT Enterprise or Business content for model training by default, and API traffic follows the same rule, with 30-day retention limited to abuse monitoring unless a customer is on a Zero Data Retention plan.
That tiered structure — consumer, Business, Enterprise, API, each with a different retention and training posture — mirrors the same complexity showing up on the Claude side, with Opus 5’s dropped retention requirement. Anyone treating “OpenAI” or “Anthropic” as a single privacy policy is already oversimplifying. The real unit of comparison is the specific plan and specific model, not the company name on the login screen.
OpenAI’s structural advantage has always been letting you keep your history without consenting to training — Private Safety Processing is OpenAI trying to close the gap on the piece Anthropic has arguably done better: catching abuse without a human reading your content.
That’s the trade each company has been optimizing on a different axis. This announcement is the first real sign OpenAI is trying to compete on Anthropic’s axis, instead of just defending its own — one more reason the ChatGPT vs Claude comparison keeps shifting instead of settling.
Where Private Safety Processing Changes the Comparison
Before this announcement, “does a human read my flagged conversation” was a real point of friction on both platforms. Automated safety systems still routinely escalate content to human reviewers when a pattern looks concerning. That’s standard practice across the industry, and one of the more uncomfortable realities of using any hosted AI tool for sensitive work. Private Safety Processing is OpenAI’s attempt to narrow that gap: pattern detection across related interactions, without a person in the loop reading the actual content, at least for the abuse-detection use case OpenAI has described so far.
It’s worth being precise about scope, because the preview announcement doesn’t cover everything. This is specifically about abuse-pattern detection, not a blanket promise that no OpenAI employee ever reviews any flagged content under any circumstance. Legal holds, active investigations, and edge cases still exist on both platforms. The September white paper OpenAI has promised should clarify exactly where those lines sit.
If Private Safety Processing ships as described, it moves one specific point of the ChatGPT vs Claude comparison in OpenAI’s favor, without touching any of the others. It doesn’t change the indefinite-retention default. It doesn’t add a training opt-out that doesn’t already exist. And it doesn’t resolve the dark-pattern criticism aimed at Claude’s own consent flow. It narrows exactly one gap — automated versus human-reviewed abuse detection — and leaves the rest of the comparison exactly where it stood before August 19.
What This Means for Client Data in Practice
My own routing hasn’t changed based on a single preview announcement, and I wouldn’t recommend anyone else’s change either. A feature that ships in September as promised is a different thing to evaluate than a feature previewed in August. What I have adjusted is which questions I’ll actually ask before the next client engagement that touches either tool: whether Private Safety Processing’s September rollout extends to consumer ChatGPT plans or stays enterprise-only, and whether Anthropic responds with anything that narrows its own training-opt-out dark-pattern criticism.
For the client work I already do, the practical split hasn’t moved. Anything sensitive enough to worry about goes through Claude with training opted out and the 30-day window active, or through Nous Hermes 4 running locally through Ollama when it shouldn’t touch a hosted API at all. ChatGPT still handles the lighter, less sensitive work — quick reframes, headline drafts — where the indefinite-history-without-training default is a genuine convenience rather than a risk.
That routing decision predates this announcement, and doesn’t depend on which company currently has the stronger abuse-detection story. The checklist I actually run before opening either tool for a new engagement stays the same, regardless of whichever company shipped the more flattering announcement that month:
- Does this content include a client name, financials, or anything covered by an NDA? If yes, it doesn’t go through a hosted API at all — Nous Hermes 4 through Ollama, or nothing.
- Is training opted out on the account I’m using? On Claude, that’s a setting to actively confirm, not assume.
- Would a 30-day or five-year retention window matter if this specific conversation leaked? If the answer is “not really,” ChatGPT’s indefinite-but-training-excluded default is a reasonable trade.
- Am I making this decision based on this month’s headline, or on the tool’s actual, current settings? A privacy comparison is only as good as the date it was checked.
That last question is the one this whole comparison keeps coming back to. A useful ChatGPT vs Claude privacy comparison for actual client work has to start from the sensitivity of the material, not from which company published the more flattering blog post that week. The safest default is still routing anything genuinely sensitive away from any hosted API, Private Safety Processing or not — and saving the head-to-head comparison for the material that was never going to be catastrophic if it leaked anyway.
Where Each Tool Still Falls Short
Claude’s opt-out-by-default training setting, however defensible Anthropic’s five-year de-identification claims are, remains a design choice that puts the burden on the user to notice and act. The dark-pattern criticism of that interface hasn’t gone away just because Private Safety Processing shipped. ChatGPT’s indefinite default retention, meanwhile, means anyone who doesn’t actively delete conversations is accumulating a growing archive of everything they’ve ever typed into it, training exclusion or not. That’s a different kind of exposure than Claude’s, but not obviously a smaller one over a long enough timeline.
Neither company has published enough independently verifiable detail about how Private Safety Processing or Claude’s de-identification process actually works under the hood, for an outside user to fully audit either claim. Both companies are, at the end of the day, asking to be trusted on the specifics — worth stating plainly, rather than letting either company’s press coverage stand in for independent verification.
Framed as a straight ChatGPT vs Claude scorecard, neither company wins outright. OpenAI wins on the training-consent split, and — if Private Safety Processing ships as promised — on automated abuse detection. Anthropic wins on API-level training exclusion being unconditional, and on Opus 5 already shipping a reduced retention window without waiting for a September announcement. A reader looking for a single winner in this ChatGPT vs Claude matchup, or any ChatGPT vs Claude comparison written this year, is asking the wrong question. The right one is which specific risk — permanent retention or training exposure — actually applies to the work in front of them.
I’d also flag that this entire comparison is a snapshot, not a settled state. Both companies have changed their consumer terms at least once in the past year. A ChatGPT vs Claude privacy comparison written in December could easily look different from this one, which is exactly why I’m treating September’s white paper as a scheduled recheck, rather than assuming today’s split holds indefinitely.
For me, I’m not switching either tool off the back of a single preview announcement. I’d treat anyone who tells you this settles the ChatGPT vs Claude privacy question as overselling a feature that hasn’t fully shipped yet. What I am doing is treating September 2026 as a real deadline to revisit this post, once OpenAI’s white paper and broader rollout actually land. A preview and a shipped feature are different enough that the honest answer today is “watch this,” not “here’s the winner.” I’d rather publish an honest mid-story update in September than a confident verdict today that a single white paper could make wrong within weeks.
Sources
- OpenAI — Offering Zero Data Retention for frontier models
- TechCrunch — OpenAI seeks to one-up Anthropic with new customer privacy protections
- Anthropic — Updates to Consumer Terms and Privacy Policy
- Privacy Watchdog — OpenAI (ChatGPT) Privacy Policy Review 2026
FAQ
ChatGPT vs Claude: does OpenAI read my ChatGPT conversations?
It depends on the situation. Routine conversations aren’t read by staff as a matter of course, but flagged content can still be escalated to human review under OpenAI’s existing abuse-monitoring process. Private Safety Processing is designed to reduce how often that escalation is necessary, not eliminate human review entirely.
Does Claude train on my conversations by default?
Yes, for most consumer Free, Pro, and Max accounts, unless you’ve manually opted out through the “You can help improve Claude” setting. API traffic is excluded from training regardless of this setting.
Is Private Safety Processing available to consumer ChatGPT users yet?
Not yet. As previewed on August 19, 2026, it’s positioned as a preview ahead of a broader rollout and technical white paper OpenAI has said will arrive in September 2026.
Should I stop using ChatGPT or Claude for client work over these differences?
No. The ChatGPT vs Claude privacy gap isn’t wide enough to justify abandoning either tool for routine work. Both remain usable for the lighter end of client work. The honest adjustment is routing genuinely sensitive material — client financials, proprietary strategy, anything under an NDA — away from either hosted tool, and toward a local model or an explicit zero-retention configuration instead.
In the ChatGPT vs Claude race, will Anthropic match OpenAI’s Private Safety Processing approach?
Not yet confirmed. Nothing in Anthropic’s public communications as of this post addresses Private Safety Processing directly. Any response would likely wait until OpenAI’s September white paper clarifies exactly what the feature does.
AI-assisted research and drafting. Reviewed and published by ToolMint.