Last week OpenAI quietly published a page explaining how ChatGPT learns from the public web without memorizing you in particular. That mattered to me because I run client briefs, NDA-flagged drafts, and half-baked positioning ideas through the model every day, and ChatGPT privacy is not a setting I can outsource to default behavior. The May 2026 update doesn’t change my contract terms — it changes which toggles I trust and which conversations I keep routing to Temporary Chat. Below are four ChatGPT privacy adjustments I made this week, what each one actually controls, and where the new OpenAI Privacy Filter fits when the in-app settings don’t go far enough.
In this article
- What ChatGPT privacy looks like after the May 2026 update
- The 4 adjustments I made on my account
- Where OpenAI Privacy Filter fits (and where it doesn’t)
- What this means if your draft pipeline runs through ChatGPT
What ChatGPT Privacy Looks Like After the May 2026 Update
The headline change isn’t a new toggle; it’s documentation. OpenAI’s May 6 explainer walks through the data-filtering pipeline that runs before training and confirms two things I had previously been guessing at.
First, on a personal Free, Plus, or Pro account, your conversations are used to improve future models by default unless you opt out — a default I keep forgetting when I sign in on a new device. Second, even with the toggle off, content stays on OpenAI servers for up to 30 days for abuse monitoring (per OpenAI’s Data Controls FAQ).
For my client work, that 30-day retention window is the part I had to internalize most when I rewrote my ChatGPT privacy posture. It’s also a reminder that nothing about the ad-injection update from earlier this month changes how training data is handled — the two systems live in different lanes.
The 4 Adjustments I Made This Week
I changed four settings in under fifteen minutes, none of them dramatic. The point was matching my ChatGPT privacy posture to what each kind of work actually needs.
- Toggled off “Improve the model for everyone” under Settings → Data Controls. Confirmed via OpenAI’s help docs that this stops new chats from feeding training, though server retention still applies.
- Moved client briefs into Temporary Chat by default. Temporary Chat doesn’t appear in history, doesn’t create memories, and isn’t used to improve models. It’s not encryption — but for one-off rewrites of a positioning paragraph, it’s exactly the right shape.
- Created a separate “personal” project for ideation that I do want stored — book notes, weekly review prompts, voice work. Default chat keeps long-term memory; Temporary Chat handles the client material.
- Stopped clicking thumbs-up/down on client-touching answers. A footnote in OpenAI’s Data Controls FAQ flags that explicit feedback on a specific chat authorizes its use for training even when the model-improvement toggle is off. Small thing, easy miss.
This routing is also why my Notion AI vs ChatGPT decision rule for client briefs still holds — not every brief belongs in ChatGPT in the first place.
None of these adjustments changed what I send. They changed what stays.
Where OpenAI Privacy Filter Fits (And Where It Doesn’t)
For sensitive client material the four toggles aren’t enough on their own, which is why the OpenAI Privacy Filter release on April 22 caught my attention.
Privacy Filter is a small open-weight model that masks names, addresses, emails, phone numbers, URLs, dates, account numbers, and secrets — eight categories total — and it runs locally, including in-browser via WebGPU. OpenAI reported a 96% F1 score on its evaluation set.
The catch is in OpenAI’s own “High-Risk Deployment Caution” note: it’s a redaction aid, not a guarantee. For me, that places it in a sensible middle slot — useful for scrubbing meeting transcripts before paste, not a substitute for a real DLP layer when a contract requires one. As one layer in a broader ChatGPT privacy stack, it earns its keep.
What ChatGPT Privacy Means If Client Drafts Run Through You
For me, the takeaway is that ChatGPT privacy in 2026 stops being a single decision and becomes a per-conversation routing problem. Personal projects flow into the default workspace where memory is an asset; client briefs get Temporary Chat plus a Privacy Filter pass when there’s anything resembling a real name. The May explainer didn’t change my contracts. It just made me admit which conversations I should never have left in default mode in the first place.
Sources
- How ChatGPT learns about the world while protecting privacy — OpenAI
- Introducing OpenAI Privacy Filter — OpenAI
- Data Controls FAQ — OpenAI Help Center
- How your data is used to improve model performance — OpenAI Help Center
AI-assisted research and drafting. Reviewed and published by ToolMint.