Somebody on your team has a free ChatGPT account, a free Gemini tab, and probably a free image or writing tool bookmarked from a Slack link nobody remembers approving. None of that is unusual — it is close to universal. The question that matters is narrower and almost nobody asks it: when they paste a customer's name, a contract clause, or a product spec into the free version, where does that text actually go?
The honest answer is that it depends on the tool, the tier, and a settings toggle most people have never opened — and the default on the free consumer tier is usually the one you would not choose if you read it first.
What the free tiers actually do, in short
- Free ChatGPT trains on your chats by default. OpenAI states plainly that content from the free and Plus tiers may be used to improve its models unless you turn off "Improve the model for everyone" in Data Controls, or use Temporary Chat.
- Free Gemini does too, and a human may read it. With Gemini Apps Activity left on — the default when you are signed in — a subset of your conversations can be reviewed by human raters and used to improve Google's models.
- Free Microsoft Copilot carries no enterprise data protection. Microsoft's own guidance is to avoid putting sensitive or proprietary work information into it, in contrast to the commercial version.
- The paid business tiers of all three flip this by default. ChatGPT Team, Business and Enterprise, Google Workspace with Gemini, and Microsoft 365 Copilot with Enterprise Data Protection all state that customer prompts are not used for training and are not human-reviewed, without you having to find a toggle.
- None of this is illegal or hidden — it is published in each vendor's own privacy documentation. It just is not what most staff assume when they open a free tool to save five minutes.
01The distinction that actually matters
Every AI vendor draws the same line, even though the marketing rarely says so out loud: the free consumer product and the paid business product are different data-handling regimes, not just different feature sets. Free consumer accounts are, from the vendor's side, a source of training data and product-improvement signal. Paid business accounts are, contractually, a service you are buying under terms that promise your content stays out of that pipeline.
This is not a conspiracy — it is the economics. A free product has to earn its keep somehow, and for a consumer chatbot that "somehow" is largely the conversations themselves. A business product is earning its keep from the subscription fee, so it can afford to promise your prompts go nowhere.
What makes this dangerous for a small business specifically is that the free tier is good enough for real work. Nobody signs up for the free tier meaning to use it for anything important, and then the deadline arrives and it is 4pm and the free tab is already open.
02What "used to train" actually means for your data
It does not mean a person is reading your prompt as you type it, in most cases. It typically means your conversation is retained, may be sampled into a batch that a model is trained or fine-tuned on, and — for at least some providers — may be reviewed by a human contractor as part of quality checks on that sampling. Google's own Gemini Apps privacy documentation says exactly this: with Keep Activity on, a subset of chats "are reviewed by human reviewers (including Google's trained service providers) to help improve Google services."
The three-minute version
If your staff are on the free tier of ChatGPT, Gemini, or Copilot: check whether training is switched on (it usually is, by default), and treat anything pasted into that tool as potentially retained, sampled, and in a small number of cases read by a person who is not on your payroll. Then decide, tool by tool, what that means is off-limits.
03Free versus business tier, tool by tool
These are the current published policies for the three most common free consumer AI tools a small business runs into, checked against each vendor's own privacy documentation.
| Tool & tier | Trains on your input by default? | Confidentiality / business terms |
|---|---|---|
| ChatGPT Free / Plus | Yes, unless you turn off "Improve the model for everyone" or use Temporary Chat | Consumer terms only. No admin console, no data processing agreement, no visibility for the business. |
| ChatGPT Team / Business / Enterprise | No, by default, per OpenAI's stated policy | Business terms with an admin console and a data processing agreement available. |
| Google Gemini (free / AI Plus / Pro / Ultra, consumer) | Yes, while Gemini Apps Activity ("Keep Activity") is on — the signed-in default. A subset may also be human-reviewed. | Consumer terms. No workspace admin oversight, no contractual data protection for the business. |
| Google Workspace with Gemini (business) | No, per Google's published Workspace commitment | Data stays within Workspace's business terms; not used for model training or human review without permission. |
| Microsoft Copilot (free, consumer) | No enterprise data protection guarantee; Microsoft advises against entering sensitive or proprietary work information | No enterprise data protection. Content is not covered by the commercial commitments below. |
| Microsoft 365 Copilot (commercial, Enterprise Data Protection) | No — prompts, responses and Microsoft Graph data are not used to train the underlying models | Protected under the same contractual terms Microsoft applies to Exchange email and SharePoint files. |
Policies as published by each vendor and checked 18 September 2026; every one of them reserves the right to change these terms, and settings toggles (Temporary Chat, Keep Activity, Data Controls) can override the tier default in either direction. Verify against the vendor's current policy page before relying on this table for a real decision — this is not a substitute for reading your own account's settings.
04What "no admin visibility" costs you, separately from training
Training is the headline risk, but it is not the only one, and it is arguably not the one that will actually bite a small business first. A free consumer account has no admin console. That means:
- You cannot see what staff pasted in. There is no audit log, no usage report, no way to know after the fact whether a customer's contract terms went into a free tool last Tuesday.
- You cannot remove an employee's access when they leave. A personal free account is not tied to your business identity system at all — it belongs to the person, not the company, and it goes with them.
- You have no contract to point to. If a free-tier tool's practices ever caused a problem, you have consumer terms of service, not a negotiated agreement with confidentiality clauses and liability terms. See our guide to what a real AI vendor contract should contain for what that gap actually looks like in writing.
- You cannot enforce a policy you cannot see being broken. This is the connective tissue to shadow AI: free tools are exactly where shadow AI lives, because nobody has to ask permission or expense anything to open a new tab.
05What's actually safe to paste into a free tool
Not "nothing" — that guidance gets ignored within a week because it is unworkable. The useful line is between content that identifies someone or something specific to your business, and content that does not.
- Generally fine on a free tier: drafting a generic email template, brainstorming blog topics, rewriting a paragraph for clarity, asking a general how-to question, summarising publicly available information.
- Do not paste into a free tier: customer names, emails, phone numbers, or account details; contract language, pricing, or terms specific to a real deal; anything covered by a client confidentiality clause; unreleased product plans or financials; employee personal data; medical, legal, or financial details about any named individual.
- Watch attachments and screenshots, not just typed text. A pasted screenshot of a spreadsheet or a dragged-in PDF carries the same exposure as typed text, and it is easy to forget that when the interface makes uploading feel casual.
- Turn off training where you can, even on a free account. It does not fix the admin-visibility problem, but it is a genuine, immediate reduction in exposure, and it takes under a minute per tool: Data Controls in ChatGPT, Keep Activity in your Google Account's Gemini Apps Activity settings, or simply defaulting to Temporary Chat.
- Write the line down and tell people where it is, rather than assuming judgement will land in the same place for everyone. A one-paragraph policy that says "no customer data, no contract terms, no unreleased pricing, in any free AI tool" resolves most of the actual risk on its own.
06When to upgrade to a business plan
The moment a free tool moves from "helps me draft something" to "regularly touches information about a real customer, a real contract, or a real employee," the free tier has become the wrong tool, not just a slightly risky one. That is the trigger to upgrade — not headcount, not budget cycle, not "we should get around to it."
A real business plan buys you three things a free account structurally cannot provide: a contractual promise that your content is not used for training, an admin console so you can see and manage who uses what, and — if you push for it, because not every plan includes one automatically — a data processing agreement that puts confidentiality in writing rather than in a marketing page. Ask for that agreement explicitly when you sign up; it is usually available on request even where it is not advertised on the pricing page.
This is also where a lot of small businesses discover the free-to-paid upgrade is cheaper than they assumed, because it is often a per-seat monthly fee rather than an enterprise procurement process. The barrier is usually inertia, not cost.
07The honest summary
Free consumer AI tools are not a trap set to catch you — the terms are published, and every major vendor discloses this in its own privacy documentation. But "published" and "read by your staff before they pasted a contract clause in" are two very different things, and the gap between them is where the actual risk sits.
The fix does not require banning free tools. It requires one written line about what never goes into them, a minute spent turning off training where the toggle exists, and a clear trigger — real customer data, real contracts, real confidentiality obligations — for when a team moves to a paid plan with an actual contract behind it.
08Common questions
Does the free version of ChatGPT train on what I type?
Yes, by default. OpenAI's published policy is that content from the Free and Plus tiers may be used to improve its models unless you turn off "Improve the model for everyone" in Settings > Data Controls, or use Temporary Chat, which is not saved to history and not used for training. ChatGPT Team, Business, and Enterprise are excluded from training by default, without needing that toggle.
Does free Google Gemini let a human read my conversations?
It can. Google's Gemini Apps privacy documentation states that with Gemini Apps Activity ("Keep Activity") turned on — which is the default while you are signed in — a subset of chats may be reviewed by human reviewers, including Google's trained service providers, to help improve Google's services and models. Turning off Keep Activity, or using a Temporary Chat, stops this for future conversations, though it does not retroactively remove anything already used.
Is free Microsoft Copilot safe for work tasks?
For nonsensitive tasks, generally yes, but Microsoft's own guidance is explicit: avoid putting sensitive or proprietary work information into the free consumer version, because it does not carry the enterprise data protection that the commercial version provides. Microsoft 365 Copilot, available to organisations with an eligible Microsoft 365 subscription, protects prompts and responses under the same contractual terms that cover Exchange email and SharePoint files, and states this data is never used to train the underlying models.
What's the actual difference between a free AI tool and a paid business tier?
Two things, not one. First, training: paid business tiers from OpenAI, Google, and Microsoft all state that customer content is not used to train models by default, while free consumer tiers typically train unless you opt out. Second, and often overlooked, is admin visibility — business tiers come with an admin console, audit logs, and usually a data processing agreement, none of which exist on a personal free account, so you cannot see what was pasted in or remove access when someone leaves.
Can I just turn off training on the free tier and call it solved?
It solves one problem, not both. Turning off training (Data Controls in ChatGPT, Keep Activity in Gemini) genuinely reduces the risk that your content ends up in a future model. It does nothing for the second problem: a free account has no admin console, so your business still cannot see what staff typed in, cannot audit it, and cannot revoke access when someone leaves. For anything touching real customer or contract data, that gap is what actually pushes you toward a paid business plan.
What should never go into a free AI tool?
As a working rule: customer names, emails, or account details; contract or pricing terms specific to a real deal; anything covered by a confidentiality clause; unreleased product or financial plans; and personal data about any named employee or customer. Pasted screenshots and uploaded files carry the same exposure as typed text, which people forget because uploading feels more casual than typing.
Does turning off training remove data that's already been used?
No, and every major vendor says this plainly. OpenAI states that once information is included in a training run it cannot be retroactively removed, and Google says the same for Gemini — conversations already sampled are not pulled back out. The practical implication is that the opt-out only protects conversations going forward, so turning it off earlier is strictly better than turning it off later.
When should a small business actually pay for a business AI plan instead of using the free tier?
As soon as a tool regularly touches information about a real customer, a real contract, or a real employee — not at some future headcount milestone. A paid business plan typically costs a per-seat monthly fee, not an enterprise procurement process, and buys you a contractual no-training commitment, an admin console, and usually a data processing agreement on request. Read what that agreement should actually contain in our related piece on vendor contracts.
Find out what's actually leaving your business
Tell us which free AI tools your team already uses — the ones nobody formally approved count too — and we will map what each tool's current terms say about training, retention, and admin visibility, and where a paid business tier would close a real gap. We are not a reseller for any of these vendors, so there is nothing here for us to upsell.
Ask for a data-exposure checkSources, read 18 September 2026: OpenAI, "How your data is used to improve model performance" (openai.com/policies) and OpenAI Help Center articles on disabling model training and Temporary Chat; OpenAI's enterprise privacy documentation on ChatGPT Team, Business and Enterprise defaults; Google's Gemini Apps Privacy Hub (support.google.com/gemini) on Gemini Apps Activity, human review, and Temporary Chats; Google Workspace's published generative AI privacy commitments for Gemini for Workspace; Microsoft Learn's "Enterprise data protection in Microsoft Copilot and Microsoft Copilot Chat" and Microsoft Support's comparison of free Copilot versus Microsoft 365 Copilot. Policies change; verify current terms directly with each vendor before relying on this article for a real decision. Related: Shadow AI: Staff Using Tools You Did Not Approve and Six Clauses to Check Before You Sign an AI Vendor Contract.
Hero image from Unsplash, used under the Unsplash License.