How to opt out of AI training on ChatGPT, Google, Meta and more

Many AI features built into the apps you already use are improved using data from the people who use them. On several of the services in this guide, your prompts, posts, streams, or documents can end up shaping how a model responds to someone else unless you opt out.

This guide covers how to opt out of AI training on 14 widely used services: ChatGPT, Google Gemini, Meta, GitHub Copilot, Atlassian, Grok, LinkedIn, Pinterest, Canva, Tumblr, WordPress.com, Perplexity, Grammarly, and Twitch. We also checked some of these settings directly in our own accounts to verify the toggle, rather than relying on public documentation alone.

Opting out of AI training means turning off the setting that lets a platform use your activity, content, or conversations to improve its AI models going forward. It’s usually found under privacy or data settings, and for most services it only affects data collected after you turn it off, not data the platform has already used.

Want cloud storage that doesn‘t look inside your files? MEGA‘s zero-knowledge encryption means your files are encrypted before they even reach us.

Quick reference

Service	Default	What’s used	Opt-out setting
ChatGPT	On (individual accounts)	Conversations	Data controls
Google Gemini	On (Keep Activity)	Chats, files, photos	Gemini Apps Activity
Meta AI	Applies unless you object (confirmed for EU users); check your region	Public posts, Meta AI chats	Privacy Centre objection form
GitHub Copilot	On (Free, Pro, Pro+, Max)	Copilot interactions, code	Copilot settings
Atlassian	On for Free/Standard in-app data; off for Premium/Enterprise. Metadata always on for Free/Standard/Premium	Metadata, in-app data	Admin data contribution settings
Grok / X	Opt-out confirmed; current default not stated by X	Posts, Grok interactions	Privacy and safety settings
LinkedIn	On	Profile, posts, comments, Groups	Data privacy settings
Twitch	On (confirmed by Dutch regulator)	Streams, VODs, clips	Security and privacy settings
Pinterest	Opt-out model; adult default not explicitly stated. Auto-excluded under 18	Pins, descriptions	Privacy and data settings
Canva	Governed by Privacy Settings (general users, no stated universal default); opt-out (Contributors)	Content, activity data	Privacy control
Tumblr	Opt-out (auto-excluded if search-indexing already discouraged)	Public blog content	Blog visibility settings
WordPress.com	Opt-out (auto-excluded if search-indexing already discouraged)	Public site content	Site visibility settings
Perplexity	On	Search queries	Account preferences
Grammarly	On (individual plans)	Written content	Privacy/data settings
Twitch	On (confirmed by Dutch regulator)	Streams, VODs, clips	Security and privacy settings

Where a row doesn’t say a flat “on,” it’s because the platform’s own documentation doesn’t state the current default in one clear sentence, even though the opt-out itself is confirmed.

Does opting out delete data that’s already been used?

For most of the services below, no: opting out stops eligible data from being used for future AI training, but it doesn’t necessarily stop the service from collecting or processing the same data for other purposes, and there’s usually no way to have your specific contribution removed from an already-trained model. Atlassian is one notable exception as it says it removes your data from its improvement datasets and retrains the affected models after you opt out, though it can still retain de-identified, aggregated patterns for up to seven years. Check each provider’s specific policy rather than assuming one rule covers all of them.

ChatGPT (OpenAI)

Default: On for individual Free, Plus, and Pro accounts. Off by default for Business, Enterprise, Edu, and the API.
What’s collected: your conversations and interactions with ChatGPT.
How to opt out: go to Settings > Data controls and turn off Improve the model for everyone, or use OpenAI’s privacy portal. New conversations are then excluded, and Temporary Chat is excluded by default.
Worth knowing: a thumbs-up or thumbs-down on a response can still let that conversation be used, even with training turned off.

Google Gemini

Default: On, tied to the Keep Activity setting.
What’s collected: chats, and files, photos, or videos you share with Gemini.
How to opt out: go to Gemini Apps Activity and turn off Keep Activity. Future chats won’t be saved or used for training unless you separately send feedback.
Worth knowing: Google still processes your chats for up to 72 hours for service and safety purposes, even with the setting off.
Worth knowing: audio recordings and Gemini Live video or screen-share have a separate AI-training setting that’s off by default, so it isn’t controlled by Keep Activity. Turning Keep Activity off also limits which Connected Apps are available, so this can affect features you already use.

Default: training applies unless you object, confirmed directly by Meta for EU users. Meta uses an objection process rather than a simple on/off toggle, and the same process may not work identically in every region, so check what applies where you are.
What’s collected: public posts, photos, and captions from adult users, plus your interactions with Meta AI. Private messages aren’t used unless you share them with an AI feature.
How to opt out: open Settings & Privacy > Privacy Centre, find the notice about AI training, and select the object link. Add your email and submit; Meta confirms by email once it accepts the objection.
Worth knowing: European regulators pushed for this route. The UK ICO raised concerns about Meta’s approach, and Meta acknowledged Ireland’s Data Protection Commission’s request to pause its EU rollout in 2024 before relaunching with clearer objection rights.

GitHub Copilot

Default: On for Free, Pro, Pro+, and Max plans, effective 24 April 2026.
What’s collected: your Copilot interactions such as inputs, outputs, code snippets, and associated context.
How to opt out: in Copilot settings, turn off Allow GitHub to use my data for AI model training.
Worth knowing: Business and Enterprise customer data isn’t used this way, regardless of this setting.

Twitch

Default: On, confirmed directly by the Dutch data protection authority on 20 August 2026, which stated the setting starts on and must be actively disabled. It was also confirmed by Twitch Chief Product Officer Mike Minton bluntly stating, ”If it was opt-in, nobody would opt in.”
What’s collected: streams, VODs, clips, stream chats, images, and channel text, used to train Amazon’s generative AI content models.
How to opt out: Settings > Security and Privacy > Opt-out Preferences, then turn off Training for Generative AI.
Worth knowing: this setting doesn’t cover chat messages you post in someone else’s stream since that streamer’s own setting decides whether your chat is used. The Dutch regulator described faces, voices, and other identifying details captured on stream as personal data that can’t be removed once used, and specifically flagged facial imagery as sensitive personal data.

Looking for cloud storage that stays out of your files? MEGA‘s zero-knowledge encryption means we can‘t scan or read what you store with us.

Grok (X and standalone)

Default: X’s current documentation confirms public posts and your Grok interactions can be used for training and confirms the opt-out, but doesn’t explicitly state the setting’s current starting position. When the control first appeared in July 2024, TechCrunch reported X enabled the new setting by default.
What’s collected: your public X posts, plus your interactions, inputs, and results with Grok.
How to opt out on X: Settings and privacy > Privacy and safety > Data sharing and personalisation > Grok & Third-party Collaborators, then uncheck the Grok training and fine-tuning option.
How to opt out on grok.com/the app: Settings > Data > Improve the Model off, or use Private Chat.
Worth knowing: Ireland’s Data Protection Commission confirmed X had already processed EU/EEA public-post data for Grok training between 7 May and 1 August 2024, and later took court action over it before X agreed to suspend the practice.

Default: On, confirmed directly by LinkedIn.
What’s collected: profile data, posts, articles, comments, Groups activity and Group messages, and some resume or job-application data. Private Inbox and InMail messages are excluded.
How to opt out: Settings & Privacy > Data privacy > Data for Generative AI Improvement, then turn it off. For a formal request, use LinkedIn’s Data Processing Objection form.
Worth knowing: LinkedIn expanded generative-AI training to the EU, EEA, UK, Switzerland, Canada, and Hong Kong on 3 November 2025. Separately, LinkedIn also shares certain data with its affiliate Microsoft for AI training purposes; reporting on LinkedIn’s own wording says this sharing won’t apply to members in the EEA, Switzerland, or the UK without further notice, though secondary coverage of this specific point has been inconsistent, so treat it as a detail worth double-checking in LinkedIn’s current settings rather than a fixed rule. The UK ICO raised concerns about LinkedIn’s approach, after which LinkedIn improved its notices, simplified opting out, and extended the objection window.

Atlassian (Jira, Confluence, Jira Service Management, Rovo)

Default: in-app data is on for Free and Standard plans, off for Premium and Enterprise. Metadata contribution is always on for Free, Standard, and Premium and it can’t be turned off on those plans; only Enterprise can disable it, effective 17 August 2026.
What’s collected: metadata and in-app data, used to improve Atlassian’s products and AI experiences.
How to opt out: as an admin, go to Security > Data contribution. Free, Standard, and Premium plans can turn off in-app data; only Enterprise can also disable metadata contribution.
Worth knowing: unlike most services on this list, Atlassian says it removes your data from its improvement datasets and retrains affected models after you opt out. It removes in-app data within 30 days and content attributes within 90 days, though de-identified, aggregated patterns can still be retained for up to seven years. One of the few genuinely retroactive opt-outs here.

Default: an opt-out exists for eligible adult accounts, and users under 18 are automatically opted out. But Pinterest’s own help page doesn’t explicitly state what the starting position is for adult accounts, so don’t assume it’s confirmed either way.
What’s collected: public Pins, descriptions, and metadata you save or upload, used to train Pinterest Canvas, its image-generation model.
How to opt out: Settings > Privacy and data, then turn off Use your data to train Pinterest Canvas (Profile > Privacy and data on mobile).
Worth knowing: Pinterest says Canvas has been trained on around 500 million rows of public Pins, descriptions, and metadata, including data collected as early as 2009.

Canva

Default: for general users, this is governed by your Privacy Settings. Canva says it won’t use your user content to improve AI-powered features unless this is consistent with your Privacy Settings, but Canva’s public documentation doesn’t state a universal starting default for ordinary accounts. In the MEGA team’s own account, this setting was on on 27 August 2026. Contributors are different: Canva’s Contributor Agreement says eligible Stock Media can be used to develop AI/ML models unless the contributor opts out.
What’s collected: for general users, usage data and User Content where your Privacy Settings permit it; for Contributors, eligible Stock Media unless they’ve opted out.
How to opt out: general users manage this Settings > Privacy controls; Contributors opt out through their Privacy Settings or by contacting Canva, though this forfeits the additional compensation Canva pays for AI/ML use of their content. Education accounts aren’t used for training.
Worth knowing: the two programmes work in opposite directions. Canva’s own documentation doesn’t guarantee your everyday designs are excluded by default, so it’s worth checking your Privacy Settings directly rather than assuming, even though its Contributors’ stock content follows a clearer opt-out-by-default rule.

Tumblr

Default: opt-out model. Blogs already set to discourage search engines from indexing them were automatically opted out.
What’s collected: public blog content, eligible to be shared with Automattic’s licensed AI and content partners unless you opt out.
How to opt out: in blog settings, under Visibility, enable Prevent third-party sharing. This also covers reblogs of your opted-out posts.
Worth knowing: Automattic, Tumblr’s parent company, only confirms working with “select AI companies.” 404 Media reported, based on internal documents, that the prospective partners included OpenAI and Midjourney.

WordPress.com

Default: opt-out model, with the same automatic exclusion as Tumblr for sites already discouraging search indexing.
What’s collected: public site content, eligible for sharing with third parties, including AI-training platforms, unless you opt out.
How to opt out: Settings > Reading > Site Visibility, then check Prevent third-party sharing. This also adds known AI bots to your site’s robots.txt disallow list.
Worth knowing: WordPress.com and Tumblr are both owned by Automattic, so the same policy and the same 404 Media reporting on OpenAI and Midjourney applies here.

Perplexity

Default: on for Free, Pro, and Max.
What’s collected: your search and query data.
How to opt out: Account > Preferences, then turn off AI data retention. Enterprise accounts are excluded, and turning the setting off only affects data collected after that point.

Grammarly

Default: On for Free, Premium, single-user Pro, and website-purchased multi-user Pro. Off by default for Sales-purchased, Enterprise, and Education accounts.
What’s collected: the content you write inside apps or sites where Grammarly is active, including Grammarly Go.
How to opt out: in your privacy or data settings, turn off Product Improvement and Training. Non-content data, like writing statistics, may still be collected.

What else opting out doesn’t do

A few patterns repeat across most of the services above:

Usually not retroactive. Turning the setting off typically stops future use of your data, not past use. Atlassian, above, is the notable exception.
Feedback can be an exception. Rating a response can let that specific interaction be used even with general training off.
Safety review still applies. Moderation and abuse detection continue regardless of your training settings.
Defaults split by account type. Business, education, and enterprise accounts are usually excluded automatically; individual consumer accounts usually aren’t. Canva is a partial exception: its general-user default isn’t stated as a fixed company policy, while its Contributor programme runs on a clearer opt-out-by-default basis.
Rights vary by region. EU, UK, and Swiss users generally get clearer, stronger opt-out rights under GDPR-style regulation.

In short, these settings usually decide what happens next rather than what already happened, but as Atlassian shows, it’s worth checking the specific policy rather than assuming.

A habit worth keeping

New AI features tend to launch with data collection switched on and the opt-out a few menus deep. It’s worth checking your settings again each time a platform you use announces a new AI feature, rather than assuming an earlier opt-out still covers it.

Want cloud storage that doesn‘t look inside your files? MEGA‘s zero-knowledge encryption means we can‘t scan or read what you store with us.

Frequently asked questions

Does opting out delete data that’s already been used to train a model?

Usually not. Most providers only stop future use, and there’s typically no way to remove your specific contribution from an already-trained model. Atlassian is a notable exception: it says it removes opted-out data from its improvement datasets and retrains the affected models.

Is opting out of AI training the same as deleting my account data?

No. Opting out controls how your data is used; account and data deletion is usually a separate setting in the same privacy section.

Will turning off AI training affect the features I already use?

It depends on the service. Most of these settings only control whether your data trains future models, not whether current features work for you. Google Gemini is an exception worth knowing about: turning off Keep Activity also limits which Connected Apps are available.

Does a thumbs-up or thumbs-down still count as training data?

On several platforms, yes. Rating a specific response can let that conversation be used even if you’ve turned off general training.

Why do individual accounts default to training while business accounts often don’t?

This is a common pattern rather than a universal rule: business, education, and enterprise customers often have contractual or compliance requirements that individual consumer accounts don’t, which is one likely reason many providers exclude them by default. Canva doesn’t fit neatly into this pattern. Its general-user default isn’t stated as fixed policy, while its separate Contributor programme uses a clearer opt-out-by-default structure.