How to Opt Out of AI Training Data for Freelance Work

Short answer: most major AI tools now use your prompts to train their models by default, and you have to switch the setting off yourself, one platform at a time. Our verdict for freelancers is USE WITH CAUTION, because opting out genuinely shrinks the data pipeline but never fully closes it. As of June 2026, an opt-out toggle reduces what a vendor keeps and how long, yet documented carve-outs (safety reviews, thumbs-up feedback, business-tier defaults) mean client material you paste can still leave your control. If you draft contracts, edit client PDFs, or paste code under NDA, the opt-out is step one, not the finish line. This guide walks through the real settings, what they do and do not cover, and how we vet privacy claims before trusting a toggle.

What “AI training opt-out” actually does with your data

An opt-out is a forward-looking switch. It tells the vendor to stop feeding your future conversations into the next model and usually shortens how long those conversations sit on their servers. On Claude, opting out reportedly cuts standard retention to roughly 30 days instead of the multi-year window that applies when you leave training on (per Tom’s Guide coverage, retrieved 2026-06-14). On ChatGPT, the Data Controls switch stops future chats from joining the training set, and temporary chats skip storage altogether (per TrustScan’s 2026 opt-out guide, retrieved 2026-06-14).

What it does not do matters more for paid work. It cannot pull anything out of a model that already trained on it. It rarely lets you pick where data is stored. And reporting in June 2026 flagged a carve-out in Anthropic’s updated terms: conversations flagged for safety review can still be used regardless of your preference, with no public definition of what triggers a flag (per techcoffeehouse analysis of the June 8 2026 policy, retrieved 2026-06-14). Treat the toggle as a meaningful reduction in exposure, not a guarantee.

The defaults also moved against users over the past year. Anthropic shifted consumer plans so that Free, Pro, and Max users now have to actively opt out, where earlier the safer setting was closer to the default (per Anthropic’s consumer terms update, retrieved 2026-06-14). The same coverage notes that leaving training enabled can extend retention to roughly five years in de-identified form, while opting out brings it back down toward a 30-day window (per Tom’s Guide, retrieved 2026-06-14). For a freelancer, that is the difference between a client draft living on a vendor’s servers for a month versus half a decade. The pattern repeats across the industry: bundled AI in tools like Slack, Adobe, and Zoom often ships with its own training default that is looser than the standalone app you already trust, so a single account-wide assumption does not hold.

What this means for solo freelancers

The gap between “I turned the setting off” and “my client’s data is safe” is where freelancers get burned. Three concrete scenarios show why.

• If you paste a client’s contract to rewrite a clause before opting out, that text may already sit in a retention window or a training batch. Based on the policies as written, you cannot retract it later, which carries a confidentiality-breach risk if your client agreement bars third-party processing.
• If you click thumbs-up on a great answer, several vendors treat that as consent to use the whole conversation for training even when your global opt-out is on. Based on the policy as written, that single click can re-expose the client material in that thread.
• If you handle EU client data, the controller-versus-processor question stays murky on consumer tiers. A standard opt-out is weaker than a formal objection; for EU/EEA work, Article 21 of the GDPR gives a stronger right to object to processing for training, which carries less ambiguity than a UI toggle alone.

There is also a reputational cost that does not show up in any privacy policy. If a client later asks, in writing, whether their material ever touched a tool that trains on user input, “I opted out afterward” is a weak answer. Freelancers who win regulated or enterprise work increasingly face data-handling questionnaires, and a clean answer is “client material never entered a training-eligible consumer account in the first place.” That is a workflow decision you make before the project starts, not a setting you flip after a problem surfaces.

The practical takeaway: the opt-out protects future conversations, but your exposure is set the moment you paste. For NDA-bound or regulated work, assume the consumer tier is the wrong tool and reach for a no-training business tier or a tool that keeps data local. We weigh each tool’s default behavior, its documented carve-outs, and the gap between marketing language and the actual policy text before we call anything safe for client use.

How to opt out safely on every major platform

1. ChatGPT: profile picture, then Settings, then Data Controls. Turn off “Improve the model for everyone.” Use Temporary Chats for anything client-sensitive, and skip the thumbs-up button on those threads.
2. Claude: profile picture, then Settings, then Privacy. Turn off “Help improve Claude for everyone.” Use Incognito Mode for client work; those chats are excluded from training even with the global switch on.
3. Gemini: Settings and help, then Activity, then set the “Keep Activity” drop-down to off. New chats are held briefly for service operation but kept out of training.
4. Everything else (Slack AI, Adobe, Zoom, Notion AI): check each tool’s admin or workspace settings individually. Bundled AI features often inherit a separate, looser default than the standalone app.

Two workflow habits beat any single toggle. First, redact client names, account numbers, and identifying details before you paste; the model does not need the real names to help you draft. Replace “Acme Corp’s Q3 contract” with “the client’s contract” and the output is just as useful while the identifying detail never leaves your machine. Second, keep a separate, opted-out workspace or browser profile for client work so a stray personal chat never lands in the same training-eligible bucket. If your tool offers a no-training business tier, that one decision removes most of the per-click risk above.

Finally, re-check these settings on a schedule. Vendors revise their terms often, and an opt-out you set in 2025 can be reset by a major terms update or a new feature that ships enabled. A quick quarterly pass through each tool’s data controls takes five minutes and catches the silent re-opt-in that has burned users before. If you onboard a new AI tool mid-project, run its opt-out before the first client paste, not after the deadline crunch when you forget.

Privacy-friendlier alternatives for client work

When the consumer tier’s opt-out is not enough, these privacy-first tools change the default instead of asking you to fight it. Each one removes a specific risk the toggle leaves open.

• Proton — what it gives you that a consumer chatbot does not: an AI assistant (Lumo) and encrypted mail/drive built so the vendor cannot read or train on your content by design, not by toggle. Pricing band: free tier available, paid plans roughly €10–13/month. Target user: freelancers who want privacy as the default for both client comms and AI drafting.
• 1Password — what it gives you that an AI tool does not: a vaulted place to store the client credentials, NDAs, and API keys you should never paste into a chatbot in the first place. Pricing band: around $8/month for the business tier. Target user: any solo worker juggling logins across multiple client accounts.
• Bitwarden — what it gives you that a chatbot does not: open-source, audited secret storage at the lowest price point, so sensitive strings stay out of training-eligible prompts. Pricing band: free for individuals, around $3/month for premium. Target user: budget-conscious freelancers who want auditable security.

For hardware-level account protection on the logins behind these tools, a physical security key such as a YubiKey 5 NFC blocks the account-takeover route that no software opt-out covers. None of these replace a chatbot’s drafting speed, but they shrink the surface where client data can leak.

The verdict

[INTERNAL_LINK_TO_CLUSTER_freelancer-cybersec]

Frequently asked questions

Does opting out delete data the AI already trained on?

No. An opt-out is forward-looking only. It stops future conversations from joining the next training run and usually shortens retention, but anything a model already absorbed cannot be pulled back out. Based on the policies as written, this is why the timing of your opt-out matters: switch it off before you paste anything client-sensitive, because retroactive removal is not offered by the major vendors as of June 2026.

Is opting out enough for GDPR-bound freelance work?

Not on its own. For EU or EEA client data, a UI opt-out is weaker than a formal objection. Article 21 of the GDPR gives a stronger right to object to processing for training, and the controller-versus-processor status of consumer AI tiers stays ambiguous. Based on the policy as written, freelancers handling EU client data should prefer a business tier with a data processing agreement rather than relying on a toggle in a personal account.

Can a thumbs-up rating override my opt-out?

On several platforms, yes. Submitting feedback with a thumbs-up or thumbs-down can be treated as consent to use that entire conversation for training, even when your global Data Controls switch is off. For client threads, the safest habit is to skip the rating buttons entirely, so a single click never re-exposes material you meant to keep out of the training pipeline. If you want to flag a bad answer, do it in a throwaway prompt that contains no client detail rather than rating the original thread.

Do business and team tiers train on my data?

Usually not by default. Enterprise and team tiers from the major vendors generally exclude customer content from training as a baseline, which is the opposite of the consumer-tier default. That makes a paid business plan the cleaner lane for freelancers handling sensitive client work. Always confirm the specific tier’s terms, since defaults differ between the free app and its bundled AI features.

Where do I opt out on each platform?

ChatGPT: Settings, then Data Controls, then turn off “Improve the model for everyone.” Claude: Settings, then Privacy, then turn off “Help improve Claude for everyone.” Gemini: Settings and help, then Activity, then set Keep Activity to off. Bundled tools such as Slack AI, Adobe, Zoom, and Notion AI each have separate settings that often inherit a looser default than the standalone product, so check them individually.

Sources

• TrustScan, “How to Opt Out of LLM Training Data” (2026) — https://trustscan.dev/blog/opt-out-llm-training-data-2026 — retrieved 2026-06-14
• Tom’s Guide, “Your Claude chats are being used to train AI” — https://www.tomsguide.com/ai/claude/your-claude-chats-are-being-used-to-train-ai-heres-how-to-opt-out — retrieved 2026-06-14
• techcoffeehouse, “Claude Training Data Opt-Out Carve-Out” (June 9, 2026) — https://techcoffeehouse.com/2026/06/09/claude-training-data-opt-out-carve-out/ — retrieved 2026-06-14
• Built In, “How to Opt Out of AI Training Data” — https://builtin.com/articles/ai-training-data-opt-out — retrieved 2026-06-14
• Anthropic, “Updates to Consumer Terms and Privacy Policy” — https://www.anthropic.com/news/updates-to-our-consumer-terms — retrieved 2026-06-14

Reviewed by Jérémy, founder of AidTaskPro and GreenBudgetHub. Based in central France. Privacy posture sourced from public policies and vendor documentation as of 2026-06-14.