Microsoft’s consumer Copilot stores your chat conversations for 18 months by default. Opt out of model training if you want — it won’t change that retention. In fact, none of the major AI chatbots fully purge your data the moment you hit delete, and a new guide from ZDNET lays out exactly how long your chats linger on company servers.
What the latest fine print actually says
A ZDNET report published on July 31, 2026, digs into the privacy controls of ChatGPT, Microsoft Copilot, Google Gemini, and others. The key finding: toggling off model training, deleting chat history, or using a “temporary chat” mode does not mean your conversations vanish instantly from the provider’s servers. Each service keeps copies for safety, security, and compliance reasons, with very different timelines.
OpenAI’s ChatGPT may hold temporary chats for up to 30 days. Google Gemini retains data for up to 72 hours even when activity tracking is disabled. Microsoft’s consumer Copilot stores conversations for 18 months by default. These durations are not affected by training opt-outs. In fact, the “do not train on my data” setting only stops future conversations from being used to improve AI models, but it does nothing to shorten how long the company hangs onto the raw dialogues.
Even feedback mechanisms can override privacy preferences. OpenAI says submitting a thumbs-up or thumbs-down on a ChatGPT response may allow that conversation to be used for training, regardless of your opt-out setting. Anthropic’s Claude has a similar caveat for feedback submissions.
The retention timelines you’re never told about
Here’s how long the major consumer AI services hold onto your chats after you’ve hit “delete” or ended a session, according to their published policies:
- Microsoft Copilot (consumer): 18 months by default. You can manually delete individual conversations or your entire history, but the data may persist in logs for security, service improvement, advertising, or compliance.
- OpenAI ChatGPT: Up to 30 days for temporary chats or deleted history, retained for safety review. Regular chats stay in your account history until you remove them.
- Google Gemini: Up to 72 hours for chats excluded from Gemini Apps Activity. The company reserves the right to retain data for service provision, feedback processing, and platform protection.
- Anthropic Claude: Retains full conversations associated with feedback for analysis and model improvement, even if normal privacy settings prohibit training.
These figures don’t appear in the apps’ onboarding screens or in the flashy “privacy center” dashboards. They are buried in support pages and terms of service — and they contradict the growing expectation that an AI chat is ephemeral.
What this means for Windows users
For the millions of people using Copilot baked into Windows 11 or via the Edge sidebar, the 18-month retention is the starkest number. Every query you type, every file you upload, every screenshot you share could sit on Microsoft servers for a year and a half. That timeline applies even if you later delete the conversation and toggle off all training and personalization.
Home users: If you’ve discussed financial details, health symptoms, or unreleased creative projects with Copilot, that information is not gone. It’s archived alongside your Microsoft account. The same goes for any Windows user who signed into ChatGPT or Gemini in a browser while logged into their Microsoft account.
IT administrators and enterprise users: Microsoft says organizational accounts using Entra ID are excluded from consumer Copilot model training, but that does not mean the data is instantly erased. More critically, if employees access personal AI chatbots from their work PCs — via Edge, for instance — corporate data can leak out of the managed environment. A staffer pasting a confidential spreadsheet into a personal ChatGPT tab is not covered by your company’s compliance controls. The account boundary, not the device, defines the privacy scope.
Developers: Pasting proprietary source code into any cloud-based assistant is a gamble. Even a temporary chat can be retained for safety purposes, and a misconfigured custom GPT with third-party actions could forward your code somewhere else entirely.
How we got here: from disposable search boxes to persistent digital companions
Three years ago, most people treated chatbots like fancy search engines. You asked a question, got an answer, and closed the tab. Conversations were not saved, and model training was a distant, abstract concern. That changed rapidly.
Today’s assistants remember your name, your writing style, your projects, and even your dietary restrictions. They accept file uploads, voice recordings, and screen captures. They can browse the live web, connect to your cloud storage, act through plugins, and summon third-party services. Each of these features creates another copy of your data or hands it to another organization. A single prompt to a connected bot might touch OpenAI’s servers, a third-party action provider, your OneDrive, and a billing system — each with its own retention policy.
Privacy controls, meanwhile, splintered. A user can be opted out of training but still have memory enabled, or be in a temporary chat that a custom action forwards externally. The result is a confusing matrix of toggles, none of which offer the simple guarantee users want: “Delete everything now.”
Six steps to tighten your AI privacy right now
1. Sanitize every prompt before you send it
No privacy setting can protect data you’ve already handed over. Replace names, addresses, account numbers, and unique identifiers in documents. For code, isolate only the relevant snippet. Never paste full logs, HR records, or legal documents into a consumer chatbot. If you wouldn’t put it in a public support ticket, don’t put it in an AI chat.
2. Lock down Copilot on Windows
Open Copilot’s settings (gear icon in the top right of the Copilot pane) and go to Privacy. Disable both “Model training on conversation activity” and “Model training on voice conversations.” Then review the Personalization and Memory sections separately — disabling training does not disable memory. If you’ve used Copilot for anything sensitive before tightening these controls, delete those individual conversations manually.
3. Rethink ChatGPT’s data controls
In the ChatGPT web or app settings, under Data Controls, turn off “Improve the model for everyone.” This stops future chats from training the model. For a disposable session, switch to Temporary Chat from the model dropdown. But remember: OpenAI can still retain the transcript for up to 30 days. If you need absolute confidentiality, don’t send the data at all.
4. Don’t overlook Gemini and Claude
For Google Gemini, head to Gemini Apps Activity and toggle off “Keep Activity.” This prevents future chats from appearing in your Google account and from training models — but they are still retained for 72 hours. In Anthropic’s Claude, avoid using the feedback buttons (thumbs up/down) on any conversation that contains sensitive material, as that can preserve the entire chat for review.
5. Disconnect unnecessary apps and plugins
Review which services have access to your AI accounts. In ChatGPT, under Connected Apps, revoke anything you don’t actively use. Never connect a chatbot to your email, cloud storage, or code repository unless it’s essential for the task. A manual upload of a cleaned excerpt is safer than granting persistent access.
6. Consider local processing for truly sensitive tasks
For work that involves personal, regulated, or proprietary data, a local AI model running entirely on your Windows PC avoids sending prompts to any cloud server. Tools like LM Studio or local deployments of Llama and Mistral models can handle summarization, drafting, and code assistance without an internet connection. The tradeoffs are speed, capability, and management overhead — but for absolute data sovereignty, they’re the only consumer-friendly option.
The outlook: more integration, more exposure
Microsoft is weaving AI deeper into the operating system with features like Recall and seamless Copilot integration. While enterprise-grade protections exist for managed tenants, the vast majority of Windows users are on personal accounts with consumer-level defaults. Until regulation mandates universal, transparent retention policies — or forces a real “delete means delete” standard — the responsibility sits with the person at the keyboard.
In the meantime, treat every AI conversation like a piece of email stored on someone else’s server. Opt-out of training, delete your history regularly, and when in doubt, don’t send it.