Short answer: yes, in some situations a human can see your conversation with AI, but usually nobody reads it in real time. According to OpenAI’s documents, authorized employees and trusted vendors have access in four cases: investigating abuse, helping the user, legal matters, and improving the model (unless you turn that off). Google admits that reviewers read and annotate some Gemini chats, and Anthropic sets retention depending on whether you agree to model training. The most effective protection is privacy settings, temporary modes, and not typing anything you wouldn’t want someone else to see.
You type lab results, a draft contract, a fight with your partner or a business idea into a chat. The answer comes back in seconds, and it’s hard to believe anyone else is watching on the other side. And yet your chat history sits on the company’s servers, and sometimes it reaches people too: moderators, quality reviewers and, in extreme cases, a court. This guide explains when and why that happens, how the three most popular chatbots differ, and what you can do to reduce the risk.
In short: when a human can see your chat
| Scenario | Can a human see the conversation? | What it means for you |
|---|---|---|
| A normal conversation, no reports | Usually not in real time; some samples may go to quality review unless you turn that option off | The risk depends on your account settings |
| The conversation looks like a policy violation | Yes, to investigate abuse and safety issues | The conversation may be kept longer than usual |
| You ask for help or report a bug | Yes, a support agent can see the excerpt you share with them | Don’t paste sensitive data into support tickets |
| A lawsuit or court order | Yes, data can be preserved and handed over | Deleting a chat doesn’t always mean the data is gone |
Four reasons a human may see your conversations
In its help center, OpenAI describes so-called “authorized human access” to user content. Access is supposed to be limited to authorized staff on a “need-to-know” basis, and OpenAI says it monitors and logs every such access. It lists four reasons: investigating abuse or a security incident, helping the user, legal matters, and improving model performance (unless the user has opted out). Selected snippets may go to service providers “for data annotation and safety purposes.” You’ll find similar logic at the other providers, though the details differ.
| Reason | Example | What you can do |
|---|---|---|
| Safety and abuse | A conversation about creating malware or harming people | Don’t break the rules; on difficult topics, ask for a safe framing |
| Quality and model training | Randomly selected conversations rated by reviewers or used to fine-tune the model | Turn off training in settings or use temporary mode |
| Technical support | You report a bug and attach a screenshot of the chat | Remove personal data from the conversation before sending |
| Law and court orders | A court orders the company to preserve chat logs | Don’t type anything you don’t want to leave in any archive |
ChatGPT, Claude and Gemini: policy comparison (as of November 21, 2026)
The table below covers consumer accounts (free and personal paid plans). Business accounts and the API work differently, as we explain below. We describe the rules based on providers’ official documents, not their marketing claims.
| Issue | ChatGPT (OpenAI) | Claude (Anthropic) | Gemini (Google) |
|---|---|---|---|
| Training on conversations | Possible by default for Free, Go, Plus and Pro accounts; turn it off in Settings → Data Controls → “Improve the model for everyone” | Since August 28, 2025 the user chooses whether to consent; the “Help improve Claude” option was on by default; change it in Settings → Privacy | Depends on “Keep Activity” (on by default for adults); turning it off stops saving and use for improvement |
| Human review | Authorized staff and vendors in four cases (see above) | Conversations flagged as policy violations may be kept longer than the standard 30 days | Reviewers (including external ones) read and annotate some conversations |
| Temporary mode | Temporary Chat: outside your history, no memory, no training; a copy kept up to 30 days for safety | Incognito chats are not used to improve Claude | Temporary chats and chats with Keep Activity off are kept for up to 72 hours |
| After deleting a chat | Removed from systems within 30 days, except data that is de-identified or required for legal/security reasons | Deleted conversations are not used for future training | Conversations already reviewed by humans are not deleted with your activity; they may be kept for up to three years |
| Retention period | Usually until the user deletes it | 30 days without training consent; up to 5 years with consent | Depends on activity settings; human-reviewed conversations may be kept for up to 3 years |
Real-life cases: when a “private” chat stopped being private
Case 1. Shared chats in Google results (2025)
In summer 2025 it turned out that search engines were indexing chats that users had shared via a link and marked as “discoverable.” Fast Company found close to 4,500 public ChatGPT conversations in Google results at the time. OpenAI’s chief information security officer Dane Stuckey announced on July 31, 2025 that the feature was being removed, calling it a “short-lived experiment” that created too many opportunities for accidental exposure. The option was voluntary (you had to choose a chat and tick a box), but many people didn’t realize the consequences. Takeaway: a link to a conversation is a publication, not a private message.
Case 2. A court orders “deleted” conversations preserved (NYT v. OpenAI)
On May 13, 2025, the court in the lawsuit brought against OpenAI by The New York Times ordered the company to preserve ChatGPT output logs that would normally have been deleted. Lawyers pointed out that this also covered chats deleted by users and temporary mode. The court held that users were not parties to the case, so their attempts to intervene were rejected. The obligation ended on September 26, 2025 (order of October 9), but logs already preserved remained available to the plaintiffs, and data of users from the EEA, Switzerland and the UK was excluded from the obligation. Takeaway: “delete” means removing it from your history, but the law can force data to be kept.
Case 3. Google admits: reviewers read and annotate Gemini conversations
In its privacy notice, Google states plainly that reviewers (including external ones) “read, annotate, and process” Gemini conversations to improve quality. According to earlier versions of the notice, the data is first disconnected from the Google account. Conversations reviewed by humans are not deleted when you delete your activity and may be kept for up to three years. Even if you turn off “Keep Activity,” Google keeps chats for up to 72 hours to provide the service, process feedback and protect against abuse. Takeaway: deleting your history is not the same as every copy disappearing.
Case 4. Claude: from 30 days to 5 years depending on a single consent
On August 28, 2025, Anthropic changed the rules for consumer accounts (Free, Pro, Max). Previously, chat data was deleted after 30 days and was not used for training. After the change, the user chooses: if they agree to training, their data may be kept for up to five years; if not, it stays at 30 days. Deleted conversations are not used in future training. Conversations flagged as violating the rules may be kept longer: previously up to two years, and according to The Register up to seven years. The change does not apply to Team and Enterprise plans. Takeaway: one click in a pop-up window changes the lifetime of your data by an order of magnitude.
Case 5. When a chat ends up in the case file: an Italian judge and a chatbot
A conversation with AI can also become material in legal proceedings. We wrote about a case in which traces of a judge’s conversation with a chatbot were found in a ruling (see our analysis). It’s a reminder that a pasted chat excerpt can be quoted or attached to documents.
Case 6. An office scenario (illustrative, not a specific company)
An HR specialist pastes a spreadsheet with employees’ names and performance ratings into a free chatbot to write a summary faster. The account is personal, so consumer rules apply: the data may be used for training if the option wasn’t turned off, and the company has no control over it. Depending on the circumstances, this may breach data protection law. Good practice: a business account with a data processing agreement, anonymized data and a company AI use policy.
What not to type into an AI chat
| Category | Why it’s risky | Safe alternative |
|---|---|---|
| Passwords, codes, API keys | They may be saved in logs and leak | A password manager; paste only code snippets without keys into the chat |
| Social Security number, ID number, payment card | Identity theft, fraud | Replace the data with fictional details: “John Smith, 1 Example Street” |
| Health data and therapy | A special category of data; no professional confidentiality | Ask in general terms, without names or dates; take serious matters to a doctor |
| Legal documents and client contracts | Trade secrets, confidentiality | Anonymize the parties; use a business account with an agreement |
| Data about clients, employees, students | GDPR and privacy law, liability toward third parties | Pseudonymization, aggregated data |
| Other people’s secrets and private conversations | It’s not your data to disclose | Describe the situation without identifying details |
| Ideas you want to patent or keep secret | Stored on a third-party company’s server | A business account or a local model |
Settings step by step: how to reduce the risk
The steps below are based on providers’ official documents. Option names may change, so if you can’t find something, search Settings for “Privacy,” “Data controls” or “Activity.”
ChatGPT: turn off training on your conversations
Go to Settings, then Data Controls, and turn off “Improve the model for everyone.” Once it’s off, new conversations won’t be used to train models.
ChatGPT: use Temporary Chat for sensitive topics
A temporary chat doesn’t go into your history, doesn’t create memories and isn’t used for training. Keep in mind that OpenAI may keep a copy for up to 30 days for safety purposes.
Claude: check the “Help improve Claude” toggle
Go to Settings, then Privacy, and check the “Help improve Claude” setting. If you don’t want longer retention, turn it off. For sensitive conversations, use an incognito chat, which isn’t used to improve Claude.
Gemini: turn off “Keep Activity” and set auto-delete
Go to myactivity.google.com/product/gemini and turn off activity saving. Be aware that Google still keeps chats for up to 72 hours, and conversations already reviewed by humans won’t disappear along with your activity.
Review your shared conversation links
Check the list of shared chats in settings and delete links you don’t need. Anyone who has the link can see the conversation.
Anonymize before you paste
Replace names, addresses, numbers and company names with labels (Person A, Company X). You keep the meaning of the task while reducing the risk.
Keep personal and work accounts separate
Don’t use a personal chatbot for company matters or vice versa. Business accounts follow different rules (see below).
Check your settings once a quarter
Providers change their policies and option names, and a change can reset your choice (as with Anthropic’s 2025 pop-up). Set a calendar reminder.
Business accounts and work: how they differ
Consumer rules don’t cover every plan. OpenAI says that by default it doesn’t use content from ChatGPT Business, Enterprise and Edu for training unless the organization opts in. Anthropic notes that the August 2025 changes don’t apply to Team and Enterprise plans or the API, which are covered by its Commercial Terms. Remember that with business accounts your employer (the administrator) may have access to conversations, so a “private” chat at work isn’t private from your company.
- Ask your IT department: which AI tool is allowed and what data you may paste.
- Make sure there’s a data processing agreement when handling client data.
- Don’t work around bans with a personal account — that’s a typical source of leaks.
Related topics: AI as a search engine, EXIF data in photos and personalized results.
Frequently asked questions (FAQ)
Not in real time, and usually not all of them. According to OpenAI, authorized staff or trusted vendors may access content to investigate abuse, provide support, handle legal matters and improve models (unless you turn off training). Access is limited and logged.
Google admits that reviewers, including external ones, read and annotate some conversations to improve quality. Conversations reviewed by humans may be kept for up to three years and don’t disappear when you delete your activity.
Conversations flagged as violating the rules may be kept longer. For consumer accounts, Anthropic introduced a choice in August 2025: agreeing to training means retention of up to 5 years, not agreeing means 30 days. Check your privacy settings.
Not fully. Temporary Chat in ChatGPT doesn’t go into your history and isn’t used for training, but OpenAI may keep a copy for up to 30 days for safety. In Gemini, temporary chats are kept for up to 72 hours. Incognito chats in Claude are not used to improve the model.
It usually disappears from your history, and from the provider’s systems within about 30 days (OpenAI), but there are exceptions: data that is de-identified, legally required or already reviewed by humans (Gemini). The court order in NYT v. OpenAI also covered deleted conversations, although the obligation ended on September 26, 2025.
Yes. Providers reserve the right in their privacy policies to disclose data in legal proceedings or when required by law. One example is the order to preserve logs in NYT v. OpenAI from May 2025.
Yes, unless you opt out. At OpenAI, training is possible for Free, Go, Plus and Pro accounts, and you turn it off in Data Controls. At Anthropic it depends on the choice you made in 2025. At Google it depends on the Keep Activity setting.
No. Anyone with the link can see the conversation. In 2025, shared chats with the discoverability option ticked ended up in search results, and Fast Company found close to 4,500 of them. OpenAI removed the feature.
They usually have stricter default rules (e.g., no training at OpenAI for Business, Enterprise and Edu), but the organization’s administrator controls the data and may have access to conversations. The rules depend on the contract.
Passwords, API keys, Social Security or ID numbers, card numbers, confidential documents, client and employee data, and health details along with a name. Replace data with labels before you paste text.
Providers usually don’t show individual cases. Check your settings and consent history (e.g., the 2025 pop-up in Claude, Data Controls in ChatGPT, myactivity.google.com at Google) and assume earlier conversations may have been used if the option was on.
Not completely. Turning off training limits the use of your data to improve the model, but staff or vendors may still have access for other purposes, such as investigating abuse or legal matters.
It requires supervision. We cover this in separate guides about kids and the internet. Set up family accounts, turn off activity saving where possible, and talk about what not to type.
If it runs on your computer without sending data to the cloud, yes: conversations don’t leave your device. However, it requires more powerful hardware and setup, and answer quality is often lower than with the largest models.
Often. In 2025 both Anthropic (August) and OpenAI (removal of discoverable links, changes to log obligations) changed theirs, and the Gemini page has many versions with update dates in 2026. We update this post, but always check the original documents.
Get the “How to Search?” guide
Practical methods for searching and verifying information, step by step, with dozens of examples.
See the guide → ☕Support us on Buy Me a Coffee
We create these materials in our spare time. Every bit of support helps us write more, and more often.
Support → ▶️Watch us on YouTube
The same topics in a shorter video format, with concrete examples on screen.
Subscribe →More from our blog
Sources: OpenAI: How OpenAI handles data in consumer services (human access, Temporary Chat, deletion) · OpenAI: How your data is used to improve model performance · Anthropic: Updates to Consumer Terms and Privacy Policy (August 28, 2025) · The Register: Anthropic and five-year chat retention · Anthropic Privacy Center: is my data used for training (incognito chats) · Anthropic: Privacy Policy · Google: Gemini Apps Privacy Notice · Notebookcheck: Gemini keeps reviewed chats for three years · TechCrunch: Google saves your Gemini conversations for years · Search Engine Land: ChatGPT kills indexable chats · Fortune: OpenAI and chats in Google Search · Engadget: the end of the ChatGPT log preservation order · Simon Willison: OpenAI on the court order (NYT v. OpenAI) · TechRound: court lifts log preservation order (EEA, Switzerland, UK). We describe the rules set out in providers’ documents as of November 21, 2026; they may change. The office scenario (case 6) is an illustration, not a description of a specific company. This article was created with the help of AI tools for research and editing, including a model from Anthropic, whose policies we also discuss. That’s why we rely on documents rather than statements.