Data, Training, and Dots: What OpenAI Will and Will Not Use

Every agent that works continuously in the background produces a byproduct nobody asked for: a running log of everything it looked at, considered, and decided not to act on. For a chatbot, that byproduct barely exists — no message, no exhaust. For a dot, which OpenAI describes as monitoring projects and responding to changing information even after you close the chat window, that exhaust is constant — a point TechCrunch's launch coverage also underscored in describing dots as agents built to "handle everything" with minimal human intervention. The question of what happens to it — does it train the next model, does it sit in a log somewhere, does it disappear — is one of the more consequential design decisions baked into the September 29, 2026 launch, and it deserves more scrutiny than a single line in a press release usually gets.

The stated policy, plainly

OpenAI's framing draws three distinct lines, and it's worth keeping them separate because they don't all work the same way.

Business, Enterprise, and Edu workspaces: content is not used to improve OpenAI's models by default. This mirrors the data-handling commitments OpenAI has made for ChatGPT Enterprise generally — it's an extension of an existing policy to a new surface, not a new policy invented for dots.

Personal-plan users (Pro, in this case, since that's the tier dots ships to at launch): individuals retain the choice of whether their conversations and work inform model development, consistent with how ChatGPT's existing training toggles work.

Proactive research and private notes: this is the genuinely new carve-out. OpenAI states it does not train on a dot's background research or the private notes a dot writes to itself — the ambient thinking a dot does while watching your tools, distinct from the workspace content it's watching. That's a specific, narrower promise than "we don't train on your data," and it's the one most directly created by dots' always-on design.

Why the third line is the interesting one

The first two lines are inherited from existing ChatGPT policy — reasonable, expected, not much of a story. The third line exists only because dots generate a category of content that didn't exist before: autonomous background thought. OpenAI apparently decided that content needed its own explicit exclusion, rather than falling under whatever bucket the underlying workspace content lands in.

That's a defensible design choice, and arguably the correct one. A dot's private notes are, definitionally, things a human never asked for and may never read — closer to an agent's scratch memory than to a document someone wrote intending it to be seen. Training on that material would mean OpenAI's models learn from millions of dots' worth of unsupervised inference about what mattered in unrelated companies' Slack channels and CRMs. Excluding it removes an entire category of "wait, you trained on what?" headline before it can happen.

The carve-out for private notes is a bet that the riskiest data a dot touches is the data nobody, including the dot's owner, ever actually reviewed.

But the exclusion also raises a question the launch materials don't directly answer: what happens when a dot's private note becomes the seed of an action a human does review — a draft PR, a flagged customer issue, a revised slide? At what point does "private research" become "workspace content" for training purposes? OpenAI's public description doesn't draw that boundary precisely, and it's the kind of detail that tends to surface only once enterprise security teams start asking pointed questions during procurement — a process already underway given the Agent 365 integration OpenAI is building with Microsoft for exactly that audience.

The Pages complication

There's a second wrinkle specific to ChatGPT Space: content that starts as one person's private work can become shared the moment it's moved into a collaborative Page. OpenAI's own framing acknowledges that individual training settings apply per participant — meaning a Page edited by several people, each with different personal training preferences, creates a genuinely ambiguous case. Whose setting governs shared content? The announcement doesn't resolve this cleanly, and it's the sort of edge case that matters enormously to the first enterprise customer who asks their legal team to sign off on Space adoption.

How this stacks against Muse and Grok Bot

Meta's public commitments for Muse are, on paper, more sweeping: users can opt out of training use entirely, conversations are excluded from ad systems, and users can ask the agent to "forget" specific information outright — a deletion mechanism dots doesn't appear to offer in any form as of launch. Meta is also promising a Muse Confidential VM with end-to-end encryption by year-end 2026, which would put a cryptographic floor under the privacy promise rather than a policy one.

xAI's Grok Bot, by contrast, has published no comparable data-handling framework in its own launch announcement or the secondary coverage available at launch. It signs into your existing tools using your own credentials — a fundamentally different trust model — but no equivalent statement about what it trains on, retains, or lets you delete has surfaced. That's not proof one doesn't exist; it's an absence worth flagging rather than filling in with a guess, and it's the sharper of two honest ways to describe an undisclosed policy.

Set side by side, dots occupies a middle position: more specific than Grok Bot's undocumented approach, less user-controllable than Muse's opt-out-and-forget framing. Whether that's the right tradeoff depends on whether you trust OpenAI's internal boundary between "private note" and "training-eligible content" more than you trust a toggle you control yourself — a question every enterprise buyer evaluating an always-on agent is going to have to answer for their own risk tolerance, not OpenAI's.

For teams used to the Claude Code Agent Skills model — where every capability you install is a file you can read, version, and remove — the contrast is instructive. A skill either does what its source says or it doesn't, and you can check. A dot's private notes are, by design, notes you'll mostly never see either.

Part 22 of 30 in the ChatGPT Dots series. Previously: Proactive Research: The Feature That Makes Dots Feel Like Employees. Next: Dots and the Bubbly Avatar: Why Cute Branding Matters. Browse more builder insights or explore AI skills at aiskill.market.