The Business Plans Have It Switched Off. Yours Does Not.

There is a difference between a company storing your conversation and a company learning from it. The first one you probably assumed. The second one is switched on right now in nearly every AI tool you use, and in most of them nobody asked you first.

This is not a scandal and it is not hidden. Every company named below publishes the setting, explains it in a help article, and lets you change it in about forty seconds. The problem is that the setting arrives already on, it is not named after what it does, and almost nobody has gone looking for it.

Stored versus learned from

Storing is a filing cabinet. Your chat sits in a drawer with your name on it. You can open the drawer, read it, delete it, and once you delete it, it is gone.

Training is a different arrangement. Your conversation goes into the next version of the model along with a few hundred million others, and what comes out is a set of numbers that shapes how the thing answers everyone. Your words are not in a drawer at that point. They have been dissolved into the product, the way sugar dissolves into a cake.

That is the part that does not reverse. You can delete the chat. You cannot delete what was learned from the chat, because by then there is no chat in there to find. Anthropic says this plainly in its own documentation: turn the setting off and your data is still included in training runs already under way, and in models already built.

Deleting the conversation and removing what the model learned from it are two different actions, and only one of them is available to you.

Who gets it turned off for them

The clearest signal of what these companies believe about the setting is who does not have to touch it.

On ChatGPT, training is on by default for Free, Plus and Pro. It is off by default on the Business plans. Perplexity turns AI data retention on by default for Free, Pro and Max, and never uses Enterprise data for training under any circumstances. Anthropic applies its model improvement setting to Free, Pro and Max, and excludes Claude for Work and the API entirely.

Line those up and the pattern is not subtle. An organization with a signed contract gets the protection written into the agreement. An individual gets a toggle and is expected to find it.

Where the switches are

Five tools, five paths. These are current as of today.

ChatGPT. Profile icon, then Settings, then Data Controls, then turn off Improve the model for everyone. Your history stays where it is. If you use ChatGPT without signing in, the same control sits behind the question mark icon in the bottom right corner, which is worth knowing, because signed-out use is not anonymous use.

Claude. Privacy Settings, at claude.ai/settings/data-privacy-controls. The setting covers Claude Code as well when you are signed in on a Free, Pro or Max account.

Gemini. Go to myactivity.google.com/product/gemini. The control used to be called Gemini Apps Activity and is now called Keep Activity. It is on by default. Human reviewers read a sample of what goes through, and reviewed conversations sit on their own retention schedule, which means deleting your activity does not clear them.

Microsoft Copilot. On copilot.com: profile icon, your name, Privacy, then Training on conversation activity and Training on voice conversations, which are two separate switches. Turning them off does not turn off personalization, so Copilot still remembers your details.

Perplexity. Account settings, then Preferences, then the toggle called AI data retention. Notice the name. It reads like a question about storage and it governs training.

What turning it off does not do

Be clear about what you are buying with those forty seconds, because it is real but it is bounded.

Every opt-out applies forward only. Data already used is already used. Every one of these companies also keeps a separate path for abuse and safety review that your preference does not touch: OpenAI holds chats up to thirty days for abuse monitoring, and Anthropic retains inputs and outputs for up to two years, plus the safety scores attached to them for up to seven, when its automated systems flag a conversation. On ChatGPT, pressing thumbs up or thumbs down can hand over the whole conversation attached to that feedback even when training is off.

So the switch is worth throwing. It is not a wall.

The numbers, since nobody puts them side by side

Anthropic keeps chats for five years in its training pipelines if you allow training, and thirty days if you do not. That is the entire trade, stated by the company: one setting, and the retention window moves by a factor of about sixty. Google runs an eighteen-month auto-delete on Gemini activity by default, with human-reviewed conversations held longer on their own schedule. Perplexity keeps uploaded files roughly thirty days on consumer plans and seven on Enterprise.

Five years is longer than most people keep a car. It is longer than the average small business survives. Whatever you pasted into a chat window in the spring of this year, about a client, a diagnosis, a contract, a child, is inside that window right now.

The one with no door

X is the exception worth naming separately. Its terms of service, effective the fifteenth of January this year, widened the definition of your content to include AI prompts, inputs and outputs, and granted the company a worldwide, royalty-free, perpetual license to use that material for AI training. A toggle for Grok exists. The license sits above it and does not expire.

Perpetual is a real word in a contract. It means there is no version of later in which this ends.

What I would actually do this afternoon

Open the two tools you use most. Not all five, because you will not do all five, and a job you abandon halfway is worth nothing. Turn the setting off in those two. It takes about a minute each.

Then adopt one habit that outperforms every toggle on this page: stop pasting things into a chat window that you would not read aloud in a lift. Client names. Bank details. Anything covered by an agreement you signed. The setting protects you from training. It does not protect you from the fact that the words left your building.

The companies did not hide any of this. They wrote it down, published it, and set the default to the answer that suited them, which is what a default is for. The forty seconds are yours whenever you want them.

Forward → Upward ↑ Onward ↗︎
Mstimaj

Sources and Further Reading

• OpenAI Help Center, Data Controls FAQ. The exact path for “Improve the model for everyone”, the signed-out control, and the thirty-day window on Temporary Chats.
• Anthropic Privacy Center, How long do you store my data? Source of the five-year training retention, the thirty-day figure with the setting off, and the two-year and seven-year windows for flagged conversations.
• Anthropic, Updates to our Consumer Terms and Privacy Policy. Which plans the model improvement setting covers and which commercial products are excluded.
• Perplexity Help Center, Data Collection at Perplexity. The “AI data retention” toggle, its default state on consumer plans, and the Enterprise position.
• Gemini Apps Privacy Hub, Google. The Keep Activity control, its default, and how human review is handled.
• Privacy FAQ for Microsoft Copilot. The two training switches for personal Microsoft accounts.