What Happens to the Data You Paste Into an AI Tool
TL;DR: "Used for training" is one of four things that can happen to pasted text, and it is usually not the one that hurts you. Knowing which account tier you are on, and what each tier keeps, takes ten minutes and settles most of the fear.
Key Insight
When someone asks "is it safe to paste this into ChatGPT," they are really asking four questions at once.
- Is my text used to train future models?
- How long does the vendor store it?
- Who at the vendor can read it, and under what conditions?
- Does it show up in a legal hold, a breach, or a shared workspace?
Training gets all the attention, but it is the most visible and the most controllable of the four. It is a toggle. Retention and access are the ones that sit in policy documents nobody opens.
The tier matters more than the brand. As of early October 2026, OpenAI says it may use content from personal ChatGPT plans for training unless you turn off "Improve the model for everyone," while its business plans and API are not used for training by default. Gemini, Copilot, Grok and DeepSeek lean the same way on consumer accounts. Claude's consumer app asks you to allow training, and if you do, retention can stretch to five years. Same company, same model family, very different outcomes depending on which door you walked through.
The pro-AI position here is simple. These tools are worth using on real work. The move is to put the real work behind the right door, not to avoid the building.
Why Teams Miss This
Most teams make one of two mistakes.
The first is treating all AI tools as one risk bucket. A policy that says "no confidential data in AI tools" gets ignored within a week, because people need the tools and the policy gives them no way to tell a business workspace from a personal login. The ban pushes usage to personal accounts, which is the worst tier for retention and training defaults.
The second is trusting the label on the tin. "We don't train on your data" is a statement about question one. It says nothing about question two. A vendor can skip training and still keep your prompts for 30 days for abuse monitoring, or longer if a legal hold applies. Pages from the same vendor can also disagree on small points and change without notice, so a screenshot of a policy from March is not evidence about November.
Here is the pattern to watch for. An analyst pastes a customer contract into a personal account and asks for a summary, because the approved tool is slower to open. Nothing dramatic happens. But now a customer's terms sit in a consumer account with consumer defaults, outside any audit trail your company owns. That is the failure, and it is boring.
How to Actually Do It
Do this once per tool, then write down the answers.
- Find the account tier. Personal, team, enterprise, or raw API. Look at the plan name, not the logo. Everything else follows from this.
- Check the training setting. On a personal ChatGPT account, that is Settings, Data Controls, then "Improve the model for everyone." In Gemini it is the Keep Activity setting. In Claude it is the model improvement toggle in Privacy Settings. Turn it off, then look at it again after the next app update.
- Use the no-history modes for one-off sensitive paste. ChatGPT's Temporary Chat, Claude's incognito chats, and Gemini's temporary chats are all documented as not used for training. They still get stored for a short window (30 days for ChatGPT, 72 hours for Gemini, per the vendors' pages), so they reduce exposure and do not erase it.
- Find the retention number. Ask what happens when you delete a chat. The common answer is "removed within 30 days," with exceptions for safety review and legal obligations.
- For anything customer-related or regulated, use a business or API tier with a data processing agreement, and ask in writing about zero data retention if you need it.
- Strip what the model does not need. Most prompts do not need the account number, the full name, or the live API key. A small pre-paste scrub catches the obvious ones.
import re
PATTERNS = {
"email": r"[\w.+-]+@[\w-]+\.[\w.]+",
"api_key": r"\b(sk|pk|ghp|xox[bp])[-_][A-Za-z0-9_-]{16,}\b",
"ssn": r"\b\d{3}-\d{2}-\d{4}\b",
"card": r"\b(?:\d[ -]?){13,16}\b",
}
def scrub(text: str) -> str:
for label, pattern in PATTERNS.items():
text = re.sub(pattern, f"[{label.upper()}]", text)
return text
if __name__ == "__main__":
import sys
print(scrub(sys.stdin.read()))
This will not catch everything, and it should not be your only control. It is a speed bump that makes the careless paste less likely, and it costs nothing to run.
What We've Learned
The useful artifact is a one-page table, owned by someone, with one row per approved tool: tier, training default, retention window, who can see chats, and the date you last checked. Check it quarterly, because every major vendor rewrote its rules at least once between 2025 and 2026.
This week's experiment: pick the three AI tools your team actually uses, fill in that row for each, and find out how many of them are being used on a personal login. That last number is the one that should drive your next decision.
Sources
- Danube Labs, training defaults, opt-outs, temporary chats and retention by vendor (checked 4 October 2026): Does ChatGPT Use My Data for Training? Privacy Settings Guide
- WitnessAI, retention windows by tier across ChatGPT, Claude and Gemini: AI data retention in 2026
- OpenAI, how personal-plan content is used to improve models: How your data is used to improve model performance
- OpenAI, Temporary Chat and opt-out behavior: Data Controls FAQ
Have a specific workflow in mind?
Bring it to a Quick Scan — a live working session where we'll tell you honestly whether it should be an agent, a workflow, or left alone, before you spend a dollar building it. You get 3 prioritized recommendations on the call, a one-page summary after, and the $500 credited toward any engagement within 30 days.
Get new posts + practical agent-ops notes
One email when something new goes up. No nurture sequence, no spam — unsubscribe whenever you want.