Is It Safe to Paste Your Code or Data Into ChatGPT? (2026 Privacy Guide)

Pasting a generic function, error message or SQL query into ChatGPT is low risk. Pasting an API key, a customer list, an internal contract or unreleased source code is not — that content leaves your device, is stored on a provider's servers, and on consumer plans may be reviewed by humans or used to improve future models unless you have explicitly turned that off.

This guide explains what actually happens to pasted text in 2026 across ChatGPT, Claude, Gemini and GitHub Copilot, the incidents that made companies clamp down, and a practical rule set you can apply today without giving up AI tooling.

What Actually Happens When You Paste Data Into an AI Chatbot

ChatGPT is not the only option, and the providers differ more than people assume — we compared ChatGPT, Claude and Gemini on actual privacy policies, including who trains on your chats by default.

An AI chatbot is not a local text editor. When you press send, your prompt is transmitted over TLS to the provider's infrastructure, tokenised, processed by a model running in their data centre, and then written to your conversation history so the assistant has context on your next message. That means three copies of your data exist almost immediately: in transit, in the model's short-term context, and at rest in your account history.

• In transit: encrypted, so a hostile network cannot read it — this is the part that is genuinely well protected. If you are on hotel or cafe WiFi you can add another layer; see our What Is a VPN? explainer.
• In the model context: your text is combined with the system prompt and prior turns. Nothing is memorised by the model itself in that moment — memorisation only happens if the text later enters a training set.
• At rest: stored under your account. Providers state that deleted conversations are purged from their systems within roughly 30 days, but abuse-monitoring and legal-preservation processes can keep copies longer.

The second thing worth understanding: retention is not the same as training. Retention means a copy exists. Training means the content may influence a future model. Both matter, and they are controlled by different settings.

Does ChatGPT, Claude, Gemini or Copilot Use My Input to Train Their Models?

It depends entirely on which tier you are on, and the pattern in 2026 is consistent across vendors: consumer tiers lean toward training with an opt-out, business tiers exclude training by default.

ChatGPT

On free and Plus accounts, OpenAI may use your conversations to improve its models unless you disable the relevant data controls in settings (or use temporary chats). ChatGPT Team, Enterprise and API traffic are excluded from training by default under OpenAI's business terms.

Claude

Anthropic historically did not train on consumer conversation data by default, and has since moved to a model where consumer users choose whether their chats can be used for model improvement — the choice materially changes how long chats are retained. Claude for Work and API usage are not used for training by default.

Gemini

Google's consumer Gemini experience saves your activity by default and human reviewers may see selected conversations; that activity is retained on a rolling window unless you shorten or disable it in your Google account activity controls. Gemini in Workspace is covered by Workspace terms and is not used to train general-purpose models.

GitHub Copilot

Copilot for Individuals historically allowed code snippets to be used for product improvement with an opt-out toggle; Copilot Business and Enterprise do not retain your code for training. This is the setting most developers forget to check, because Copilot reads context from files you never consciously "pasted".

Policies change. Before you commit a workflow to a tool, open the vendor's current data-usage page for your exact plan rather than relying on a blog post — including this one.

Real Incidents Where Pasted Data Caused Problems

The best documented case is Samsung. In 2023, Bloomberg reported that Samsung Electronics restricted employee use of generative AI tools including ChatGPT after engineers pasted internal material — reportedly source code and meeting content — into the chatbot. Coverage by Forbes, The Register and Fortune described an internal memo warning staff that misuse could result in termination, and Samsung began building internal tooling instead.

Two more recent, verifiable patterns are worth knowing:

• In June 2026, TechCrunch reported that Google filed suit against an alleged Chinese cybercrime network that used AI to run scam operations against very large numbers of victims — evidence that AI-assisted fraud infrastructure is now industrial in scale.
• In May 2026, Microsoft's security researchers documented a multi-stage phishing campaign that ended in adversary-in-the-middle token theft — the kind of attack where a single pasted session token or credential is enough to compromise an account.

Note what is not in that list: there is no credible public case of a mainstream chatbot spontaneously reciting one company's pasted source code to another user. The realistic risk is more boring and more likely — a third party now holds your confidential data, your account credentials protect it, and a contract you did not read governs it.

What You Should Never Paste Into an AI Tool

Use this as a hard list. If the content matches any line here, do not paste it into a consumer AI account, no matter how convenient it would be.

• Passwords, recovery codes and session tokens. If you already pasted one, rotate it — then generate a strong replacement rather than reusing a variant of the old one.
• API keys, cloud access keys, private keys and .env file contents.
• Customer or employee PII: names with emails, addresses, phone numbers, national IDs, health or payment data.
• Unreleased or proprietary source code, especially security-relevant logic, licensing checks and internal cryptography.
• Unpublished financial data: draft results, forecasts, valuations, M&A material — this can also create regulatory exposure, not just privacy risk.
• Anything under NDA, legal privilege or a data-processing agreement that does not list your AI vendor as an approved sub-processor.

A useful mental test: would you be comfortable pasting this into a public support forum under your real name? If not, it does not belong in a consumer chatbot.

How to Use AI Tools Safely

You do not have to choose between productivity and confidentiality. In practice, four controls cover almost every scenario.

1. Pick the right tier. Business, Team or Enterprise plans exclude your data from model training by default and give admins retention controls. For any company use, this is the single highest-value change.
2. Turn training off on consumer accounts. In ChatGPT, disable model improvement in data controls; in Gemini, shorten or disable activity retention; in Copilot, uncheck code-snippet collection. Do it once per device-independent account, not per chat.
3. Redact before you paste. Replace real identifiers with placeholders — ACME_CLIENT_1, user@example.com, sk-REDACTED. Where you need consistent-but-meaningless identifiers, a one-way hash of the original value works well and is reproducible across prompts.
4. Run a local model for the genuinely sensitive work. Small open-weight models running on your own machine never send tokens anywhere. They are weaker than frontier models, but for refactoring internal code or summarising confidential documents they are usually good enough.

Two habits matter as much as the settings: keep AI chats out of your incident-response and secrets-handling workflows entirely, and write a one-page internal rule for your team. Most leaks are not sophisticated — they are a rushed engineer pasting a whole config file at 11pm.

The Bottom Line

Pasting code into ChatGPT is safe when the code is generic and carries no secrets, and unsafe the moment it contains credentials, customer data or unreleased intellectual property. Assume everything you send is stored, check the training setting for your exact plan, redact by default, and reserve local models for the material that genuinely cannot leave your control.