Anthropic’s decision to switch Claude’s default settings and begin tapping consumer chat data for model training has jolted the AI world, forcing users to take immediate steps if they want to keep their conversations private. The move, which places the burden on users to opt out, mirrors similar privacy toggles already in place at OpenAI and Google, but with a critical difference: Anthropic’s policy now defaults to data sharing, with a retention window of up to five years. This shift exposes the new normal in generative AI—words exchanged with a chatbot are not confidential by default, and staying private requires deliberate action.

The New Default: Opt-In Data Sharing

Until recently, Anthropic stood apart by not using consumer chat content to train Claude, except when users submitted explicit feedback. That changed abruptly. New and returning Claude users now face a pop-up with a pre-checked box that says “Help improve Claude”. Unless a user unticks that box, future conversations—and any resumed chats—can be swept into training datasets. The same goes for OpenAI’s ChatGPT and Google’s Gemini, which already had “Improve the model for everyone” and “Gemini Apps Activity” toggles, respectively, but the Anthropic reversal has reignited scrutiny over how these companies handle consent.

The practical effect is stark. A casual user who clicks through without reading—behavioral research shows most people accept defaults—could see their medical queries, legal drafts, proprietary code, or intimate personal details stored for up to five years and used to train future models. Anthropic has carved out narrow exemptions for enterprise, government, and API accounts, which are handled differently, but for the millions of consumers on free or personal paid plans, the default is now “share first, ask later.”

How to Lock Down Your Privacy on ChatGPT, Gemini, and Claude

Each provider offers settings that prevent your chats from being used for model improvement. The steps are simple, but the caveats are many.

OpenAI / ChatGPT

  • Open ChatGPT on web or mobile and click your profile icon.
  • Navigate to Settings > Data Controls.
  • Turn off “Improve the model for everyone.”
  • For individual sessions, choose “Temporary Chat” from the model dropdown—these disappear from history and are not used for training.

Watch out: OpenAI retains unsaved chats for 30 days under certain configurations. Feedback you submit (thumbs up/down with comments) can still be used for training even if the toggle is off. Temporary Chat and training controls may not behave identically across devices and browsers.

Google Gemini

  • Sign into Gemini at gemini.google.com or open the app.
  • Tap the menu and go to Activity (or “Gemini Apps Activity,” soon to be renamed “Keep Activity”).
  • Turn the activity setting off to stop your chats from being used for model improvement.
  • Use “Temporary Chat” mode to keep a specific conversation out of history and training.

Watch out: Gemini retains operational copies of chats for about 72 hours for service continuity unless you use Temporary Chat. Chats escalated for safety or human review can be held longer and may still feed model improvements. Workspace/enterprise accounts often have different defaults set by admins.

Anthropic / Claude

  • When the pop-up appears, uncheck “You can help improve Claude” before proceeding.
  • If you’ve already passed the prompt, go to Settings > Privacy and turn off “Help improve Claude.”
  • New users must make the choice at signup; existing users get the pop-up and a deadline to decide.
  • Opting out applies to new and resumed chats; data already used for training cannot be retroactively purged.

Watch out: The default is to share. If you do nothing, you’re in. Opted-in data may be retained for up to five years—far longer than earlier consumer practices. Even if you opt out, conversations flagged for trust and safety review or explicitly reported can still be used to improve Claude’s safety systems.

Why Companies Want Your Conversations

There are legitimate reasons providers are so hungry for chat logs. Real, messy human interactions are training gold. They teach models to handle follow-ups, correct errors on the fly, and grok implicit intent—skills that static text from books or websites cannot impart. Chat transcripts also help build safety classifiers that catch abuse, phishing, and other misuse. Access to proprietary conversational data gives a competitive edge: a model fine-tuned on millions of real dialogs often performs better in dialogue than one trained only on crawled web pages.

When users opt in, providers can also personalize responses and memory systems, delivering more useful long-term assistance. These benefits matter, and for some users they are worth the trade-off. But they depend on safeguards that remain inconsistent and, at times, opaque.

Privacy Risks and Unresolved Weaknesses

The opt-out framework has gaping holes that privacy advocates and regulators are only beginning to probe.

Default opt-in exploits inattention. Most users never change defaults. Presenting data sharing as the path of least resistance—especially with a prominent “Accept” button and pre-checked box—borders on dark pattern design, a practice consumer protection agencies have flagged in other contexts.

Retention windows vary wildly. Gemini’s operational copies live for days; opted-in Claude data can linger for five years. Long retention multiplies exposure to data breaches, legal subpoenas, and secondary uses never envisioned at the time of the chat.

Human review is a black box. All three providers reserve the right to send content to human reviewers. Who those reviewers are, how they’re trained, and how long they keep the data is seldom disclosed in detail. This lack of transparency erodes trust.

Deletion is not erasure. Providers say they won’t use your chats for training after you opt out, but they cannot remove your data from models that have already ingested it. There is no mechanism to “unteach” a model.

No independent verification. There’s no way for you to audit whether your conversation really was excluded. You have to trust the company’s word and self-reported compliance.

Enterprise gets a privacy premium. Business customers often enjoy stronger defaults against training, creating a two-tier system where only paying organizations get robust protection by default.

Practical Steps to Regain Control

Beyond toggling the training switches, a few habits can significantly reduce your exposure:

  1. Turn off model training on every device and account you use for ChatGPT, Gemini, and Claude.
  2. Use Temporary Chat / ephemeral modes when discussing sensitive topics like medical symptoms, financial details, or proprietary code.
  3. Never paste truly sensitive identifiers—Social Security numbers, full contracts, patient records—into any cloud chatbot. Prefer private, offline tools or enterprise options with explicit non-training commitments.
  4. Delete history regularly and file a formal privacy request through the provider’s portal if you want an additional administrative record.
  5. Opt for paid business or enterprise plans if your use case demands guaranteed non-training terms; these tiers often exclude your data from model improvement by default.
  6. Keep local copies of important prompts and outputs instead of relying on chat history that could be deleted or retained inconsistently.
  7. Watch for policy deadlines. If a mandatory decision window pops up (like Anthropic’s new prompt), act before the cutoff to avoid being auto-enrolled.
  8. Audit your accounts periodically for default changes or synchronization quirks across devices.

What Good Privacy Practice Should Look Like

If providers truly want to balance innovation with trust, they need to adopt clearer, more privacy-forward standards:

  • Default to opt-out of training for consumer accounts, or at minimum make opt-in a separate, unbundled choice.
  • Offer fine-grained controls: per-conversation opt-in, separate toggles for media vs. text, and an explicit reviewer opt-out.
  • Shorten retention windows for training data, and publish concrete timelines for human review retention.
  • Release regular transparency reports and independent audits that verify privacy commitments and quantify human review.
  • Where feasible, provide cryptographic proofs or verifiable logs that a conversation was excluded from training.
  • Design consent flows without dark patterns—clear language, balanced button prominence, and meaningful friction for data sharing.
  • Extend enterprise-grade protections to paid consumer tiers at a reasonable cost, not just to high-value business customers.

Regulatory Pressure Looms

Regulators in the U.S., EU, and elsewhere are paying attention. Consumer protection agencies and privacy enforcers are starting to scrutinize whether default opt-in model training meets the bar for informed consent. Key legal questions are unresolved: How should consent be defined for training on conversational data? What obligation do vendors have to purge data from models once it’s ingested? Should human reviewers be subject to stricter access logs and controls?

Pending enforcement actions and new legislation will likely determine whether opt-out systems become the permanent norm or whether regulators mandate stricter baseline protections—perhaps eventually requiring opt-in by default.

A Fragile Truce

User-facing privacy controls are a step forward, but they are not the finish line. The current model asks everyday users to navigate a patchwork of toggles, fine print, and retention policies that shift without much warning. The burden should not rest solely on the individual.

For now, the rule of thumb is simple: assume any cloud chatbot is listening for more than just your immediate query. If privacy matters, speak up—by toggling the right switches—because silence is increasingly treated as consent.