Blog

What Happens to Your Project Data When Someone Pastes It into ChatGPT

Jared Tesche · October 29, 2026 · Updated October 30, 2026

Long read?

If someone on your team pastes a contract, a bid or a sub’s pricing into ChatGPT, what happens next depends on which account they used. On a personal account, free or paid, the text can be kept and used to train the next version of the tool unless someone turned that setting off. On a business plan, the software companies say they don’t train on your data by default. On every plan the chat is stored for a while, and a court can order it kept. The fix is a one-page rule about what your people can paste and where, and the rest of this post shows you how to write one.

I was walking past a director a while back and they said to me, “Look, use ChatGPT to make your email clearer. You can just paste it in.” It was their own paid account. You can tell people to take the client names and numbers out first, but at some point, if there’s no information left in the email, what’s ChatGPT supposed to rewrite?

Every policy in this post was checked on October 29, 2026. These pages change often, so treat the date as part of the answer and check the company’s own page before you rely on any detail.

What are your people already pasting on their own accounts?

This is the bigger risk, because you can’t see it. A 2024 Microsoft and LinkedIn survey of 31,000 office workers found that 78% of people using these tools bring their own to work, and 80% at small and mid-sized companies (Microsoft). That survey is about office workers in general, not contractors, and it’s more than two years old, so don’t read it as a number for your crew.

A lot of people assume that paying for a personal plan makes it private. It doesn’t. ChatGPT’s paid personal plans are called Plus and Pro, and on both of them OpenAI, the company behind ChatGPT, can use your chats to train its tools unless you go into Settings and turn it off (OpenAI). The company that owns the email has no say in any of it.

When your team uses a business plan, the person running the account can see what’s going on. On ChatGPT Business, the admin can view, export and delete members’ conversations, and on Enterprise they get an audit log (OpenAI). Claude’s audit logs come with its Enterprise plan (Anthropic). Microsoft logs Copilot chats in the company’s own Microsoft 365 account, where they can be searched and kept like email (Microsoft).

When your team uses personal accounts, you can’t see any of it. You don’t know what was pasted, what was kept, or whether anyone turned training off.

What actually happens when you paste something in?

Every chat tool runs on what’s called a model. It’s the engine behind the chat window, built by feeding it a huge amount of text until it learns which words tend to follow which. OpenAI says the finished model is a set of numbers, not a stored copy of what it read (OpenAI). Anthropic says the same about its Claude models: they don’t store text like a database, and they don’t look things up in the original training text once they’re built (Anthropic).

What you paste into a chat is separate from the model. The company stores it as part of your conversation, and depending on your plan and settings, it may use that conversation later to help train the next model. So there are really two questions: is it used for training, and how long is it kept?

Is this just a ChatGPT problem?

No. ChatGPT is the example because it’s the one everyone has heard of, but Claude, Gemini and Microsoft Copilot all work the same way where it matters here. Each one has a personal version and a business version, and the rules are different for each. If your team uses one of the others, everything below still applies, so go read that company’s page.

Personal plans versus business plans: who sees what?

Here’s the short version for the four tools a small contractor is most likely to run into.

Tool Personal plans Business plans
ChatGPT (OpenAI) May be used for training unless you opt out, and the opt-out only covers new chats Not used for training by default
Claude (Anthropic) You choose whether chats are used for training Not used for training by default
Gemini (Google) Human reviewers read some chats Not human reviewed or used for training outside your company without permission
Copilot (Microsoft) You can opt out of training in your account Work accounts aren’t used to train the underlying models

Checked October 29, 2026. Sources: OpenAI, Anthropic, Google personal, Google Workspace, Microsoft personal, Microsoft work.

A few details change how much you can lean on that table.

On ChatGPT’s Free, Plus and Pro plans, OpenAI says it may use your content to train its models, and you can turn that off in Settings or in its privacy portal. After you opt out, OpenAI says it won’t use your new conversations. It doesn’t say it will pull back the old ones. On ChatGPT Business, Enterprise and Edu, OpenAI says it doesn’t use your inputs or outputs by default (OpenAI).

There’s a catch with the thumbs up and thumbs down buttons. On personal ChatGPT plans, if someone rates a response, OpenAI may use the whole conversation for training even after an opt-out. On ChatGPT’s business plans, OpenAI says rating a response doesn’t let it train on the conversation (OpenAI). Anthropic’s business plans have a similar exception of their own: if someone explicitly sends Anthropic feedback or a bug report, it can use that (Anthropic).

On Google’s personal Gemini app, human reviewers read some chats, and Google tells you straight out not to enter confidential information you wouldn’t want a reviewer to see (Google).

Temporary or incognito chats help, but only partly. OpenAI says temporary chats aren’t used for training, and it may still keep a copy for up to 30 days to check for misuse (OpenAI).

How long is the data kept, and who can demand it?

The number you’ll hear is 30 days, and it’s true for one thing: a chat you delete. OpenAI says a deleted chat is scheduled for permanent removal within 30 days, unless it has to keep it for security or legal reasons (OpenAI). Anthropic says the same for a deleted chat. After that, the numbers get longer:

On business plans, the company running the account has more say. OpenAI says admins on Business and Enterprise control how long data is kept (OpenAI), and Microsoft lets admins set how long Copilot chats are kept, the same way they do for email (Microsoft).

Then there’s the legal side. Turning off training doesn’t mean the data is gone, because a court can order a company to keep it. In the New York Times lawsuit against OpenAI, a court ordered OpenAI to keep ChatGPT conversations on the Free, Plus, Pro and Team plans that would normally have been deleted, whatever setting the user had picked. OpenAI says that obligation ended on September 26, 2025, but it still holds a limited set of April to September 2025 data for the case, and it says ChatGPT Enterprise and Edu were never part of the order (OpenAI).

Can a competitor get my bid out of it?

This is the fear that comes up most. Company X can’t log in and ask for what Company Y pasted. The model doesn’t work like a filing cabinet, so there’s no search box that returns your bid by name, and the companies say the finished model can’t look up its original training text.

That said, researchers have shown that models can sometimes repeat text they were trained on. In 2020, a team pulled hundreds of exact passages out of GPT-2, an early OpenAI model, including names, phone numbers and email addresses, even when a passage showed up in only one document. Bigger models were more vulnerable (Carlini et al.). In 2023, a follow-up team got ChatGPT itself to spit out training text at about 150 times its normal rate, and concluded that the training meant to make these tools behave doesn’t get rid of it (Nasr et al.).

What those papers pulled out was text that also exists on the public internet. I’m not aware of a study showing someone pulling a specific company’s private chat out of a tool like ChatGPT. The risk still isn’t zero, and it points at one kind of content: rare, one-of-a-kind text, like your sub pricing or a client’s personal details. That’s exactly the kind of thing you shouldn’t paste into a tool that trains on it, and it’s why the table in the next section has a red row.

What’s safe to paste, and what isn’t?

After 15 years in construction, here’s the sorting I’d hand to a project team. It’s a starting point, so adjust it for your own work.

Tier What goes here Where it can go
Green Public tender documents, published specs, public permit records, your own marketing copy Any account, though a business account is still the better habit
Yellow Your own schedules, meeting notes, scopes, day-to-day emails A business account only
Red Sub pricing and bid breakdowns, signed contracts, client financials, anything under an NDA, personal information about employees or clients Not in a chat tool, unless your company has a business plan and has decided it’s allowed

Public tender documents are the easy example of something safe, because anyone can already download them. The point of the green tier is to let your team use these tools freely on work that carries no risk, so people don’t feel like they’re being told no to everything.

Stripping names out before you paste helps, but only so far. Go back to that director’s email: take out the client, the job, the numbers and the dates, and there isn’t much left for the tool to work with. People stop scrubbing once scrubbing makes the tool useless, so you’re better off counting on the account than on the scrubbing.

Personal information is its own category. Alberta’s private-sector privacy law, PIPA, says an organization has to take reasonable steps to protect the personal information it holds (Government of Alberta). The federal privacy commissioner’s guidance on these tools says to use de-identified information in prompts where possible, and to enter sensitive or confidential information only where it’s authorized (Office of the Privacy Commissioner of Canada). That’s guidance, not legal advice, but it’s a good reason to keep employee and client details out of the prompt.

What does a simple rule look like?

If one of my coordinators asked me tomorrow whether they could use ChatGPT to clean up an email to the owner, I’d tell them no, use the Copilot that’s built into Outlook. If your company runs on Microsoft 365, Copilot Chat comes with the business plan at no extra cost, and when you’re signed in with your work account it can already see the email you’re writing. Microsoft says those chats stay inside the Microsoft 365 setup your company already pays for and aren’t used to train its models. Look for the green shield at the top of the chat window, which is how Microsoft shows you’re on the protected work version (Microsoft).

Microsoft doesn’t run every model behind Copilot itself. Some are run by OpenAI and by Anthropic, the company behind Claude, and for most companies in Canada both are switched on by default (Microsoft on Anthropic, Microsoft on OpenAI). Microsoft holds them to the same contract that covers the rest of your Microsoft 365, so your email is still protected, but it can leave Microsoft’s servers to get answered, and Microsoft’s own pages don’t say which model handles a chat in Outlook. If you need it to stay with Microsoft, your admin can switch those outside companies off, which I cover further down. Web searches are separate: if Copilot looks something up online to answer you, Bing gets a few search words, not your email or who you are.

Copilot isn’t the best tool out there, though. I use Claude every day, and where I can, I’d get a company onto a business plan with a stronger tool like that. But if Copilot is all your company can get to right now, some AI is better than no AI, and it’s a lot better than your people using their own accounts.

Answering that one person at a time gets old fast. A company that uses these tools should have it written down. The formal name for this is an AI management system, and there’s an international standard for it, ISO/IEC 42001, which applies to organizations of any size. A ten-person contractor doesn’t need a certification, and a one-page rule beats having nothing at all.

Here’s what goes on the page:

  1. Which accounts are approved. A business plan on a stronger tool is the better pick, and if Microsoft 365 is what you’ve got, Copilot is a fine place to start. Either way, say personal accounts aren’t for company work, paid or not.
  2. The green, yellow and red table from above, adjusted for your work.
  3. A line on personal information: keep it out where you can, and don’t treat stripping names as a substitute for using the approved account.
  4. Who to ask when someone isn’t sure.
  5. A review date, because these companies change their policies every few months.

If you want help working out what your team should and shouldn’t paste, and getting it down on one page, get in touch with me.

Are there other ways to keep sensitive data in-house?

There are a few, and they’re a bigger step than a written rule.

You can run a model on a computer in your own office, so the data never leaves the building. The trade-off is quality, and how much quality you get depends on how much hardware you’re willing to buy. Artificial Analysis, an independent group that tests these models, scores them on a single scale. When I checked on October 13, 2026, the best online model scored 57.6 (Artificial Analysis). Gemma 4 E4B, a free model from Google, scored 8.9. It runs through a free app called Ollama (Ollama) on a desktop that costs about $4,000, roughly what you’d pay for a good CAD workstation. That’s about where the best online tools were in spring 2024.

You can spend a lot more on bigger hardware and run better models. The best free model you can download, MiMo-V2.6-Pro from the phone company Xiaomi, scored 46.3, a lot closer to the top. It’s also huge. Xiaomi’s own setup instructions spread it across 8 to 16 server graphics cards (Xiaomi), which puts it well past anything you’d run in an office. Artificial Analysis has since changed its test and hasn’t scored the small model on the new one yet, so treat these numbers as a snapshot.

Line chart of the best online model and the best free model by release date, 2023 to 2026. Gemma 4 E4B, which runs on a desktop that costs about $4,000, scores 8.9, level with the best online model of spring 2024. The best free model, which needs a server, scores 46.3. The best online model scores 57.6.

Take the chart as a rough guide. Artificial Analysis estimated some of these scores, and its test leans on hard coding work, which is where a small model looks its worst.

If your company runs on Copilot, your admin can also keep chats with Microsoft. In the Microsoft 365 admin center, under Copilot settings, there’s a page called “AI providers operating as Microsoft subprocessors.” Set OpenAI and Anthropic to “No users” there, and Copilot only uses models that Microsoft runs on its own servers, which Microsoft says keeps the data inside Microsoft (Microsoft). You still get ChatGPT’s models, because Microsoft runs its own copies of them. What you give up is Claude, and some of the newest models and features, which Microsoft says reach Copilot faster through OpenAI’s own servers (Microsoft).

You can also limit which folders and files a tool can reach. Microsoft says Copilot follows the permissions a person already has, so it can only pull in what that person could already open (Microsoft). That keeps sensitive files from showing up in an answer by accident. If you want to know where these tools actually help a construction team and where they don’t, that’s on my what works page.

Sources

All pages checked October 29, 2026, except the Artificial Analysis scores, which are from October 13, 2026. Microsoft’s pages on which models run Copilot were checked October 30, 2026.

By Jared Tesche.