Which SaaS Tools Train AI on Your Data? 13 Vendors' Defaults and How to Opt Out (2026)

The Short Answer: It Depends on the Plan You Bought, Not the Brand

Whether a SaaS tool trains AI on your data is almost never a property of the company. It is a property of the plan. In the vendor documents checked for this guide, the same product routinely has two answers: OpenAI says it may use content from ChatGPT Free, Plus and Pro to train its models while content from ChatGPT Business, Enterprise, Edu and the API is not used by default; Grammarly switches its training control on for individuals and off for teams bought through sales; Figma trains on Starter and Professional teams by default and has the setting disabled on Organization and Enterprise; Gamma allows training on individual workspaces and locks it off on business ones; Vidyard's terms license Free-plan content for AI training and do not extend that clause to paid plans.

That makes the practical question narrower and more answerable than 'is this tool safe'. For each tool, find out which plan your account is on, what the documented default for that plan is, where the control lives and who is allowed to change it, and what the control does not cover. The last part matters most: at several vendors a thumbs-up or thumbs-down sends the whole conversation to training even after you opt out, and opting out stops future use without undoing training that already happened.

This guide documents thirteen vendors from their own help centres, privacy pages and terms, checked on November 12, 2026, and gives a 45-minute audit with a free CSV. No account was created, no setting was toggled and no vendor was contacted, so the paths quoted are the vendors' written instructions rather than observed screens. Nothing here is legal advice, and a signed enterprise agreement or data processing addendum overrides the public pages quoted below.

Download the AI Training Settings Audit Checklist

Download the free 22-step AI training settings audit checklist CSV. Each row has an ID, phase, task, owner role, where to check, a dependency, acceptance evidence and a status column. It is an original Softora worksheet, not an export from any privacy or compliance product, and a spreadsheet will not enforce the dependencies, so start a row only when the rows it depends on have evidence.

The checklist is built for a team of roughly three to fifty people where one person ends up owning tools. It complements the adoption steps in our AI tools implementation checklist, which covers choosing and rolling out an assistant; this one covers a single question for every tool already in use, including the ones nobody thinks of as AI tools.

Four Patterns That Explain Almost Every Vendor

The first pattern is on by default with a self-serve switch, and it is the norm for individual and low-tier plans. ChatGPT's individual plans, Grammarly's individual accounts, Figma's Starter and Professional teams, Gamma's individual workspaces and LinkedIn's member setting all work this way: the vendor documents that your content may be used, and a toggle in settings stops it going forward. Anthropic's consumer plans are framed as a choice, with chats used if you choose to allow it, and Google's consumer Gemini app ties improvement to your activity settings.

The second pattern is off by default on business tiers. ChatGPT Business, Enterprise and Edu, Claude for Work and the Anthropic API, Notion AI, Zoom, Google Workspace with Gemini, Microsoft 365 Copilot, Figma Organization and Enterprise, Gamma's business workspaces and Grammarly accounts purchased through its sales team all state that customer content is not used to train models by default. These are the tiers where the commitment is usually also written into a contract rather than only a help article.

The third and fourth patterns are rarer and easier to miss. Slack separates generative AI, which it says it does not train on customer data without affirmative opt-in, from predictive global models for things like search ranking and emoji suggestions, which do analyse customer data unless a workspace owner emails Slack to opt out. And some vendors document training without documenting a switch: Otter.ai says it de-identifies user data before training its models, and Vidyard's Self-Serve Terms license Free-plan content for training as a condition of the free tier. For those, the control is the plan you choose or a written request.

AI Assistants: ChatGPT, Claude and Gemini

OpenAI's help article on model improvement states that when you use its services for individuals, Free, Plus and Pro, it may use your content to train its models. To opt out you turn off 'Improve the model for everyone' under Settings, Data controls in ChatGPT, or select 'Do not train on my content' in the Privacy Portal; either is sufficient and new conversations are then excluded. Temporary chats are not used while they remain temporary. Codex has a separate 'Include environments' setting that the ChatGPT toggle does not change. For ChatGPT Business, Enterprise, Edu and the API, inputs and outputs are not used by default. Our ChatGPT review covers the plans themselves.

Anthropic splits its documentation in two. For consumer plans, Claude Free, Pro and Max, its privacy article, dated March 16, 2026, says chats and coding sessions are used to improve models if you choose to allow it, if a conversation is flagged for safety review, or if you explicitly opt in to a programme. The switch is 'Help improve our AI models' under Settings, Privacy, per the settings article, and Incognito chats are excluded even when the setting is on. For Claude for Work and the API, the commercial article, dated August 18, 2026, says inputs and outputs are not used to train models by default. See our Claude review for plan differences.

Google also has two regimes. The Gemini Apps Privacy Hub, updated October 26, 2026, covers the consumer Gemini app: Google says it uses the data to improve its services including generative AI models, that human reviewers review some of it, and asks you not to enter confidential information you would not want a reviewer to see. Gemini Apps Activity is where you control whether your data is used to improve Google AI, the auto-delete default is 18 months and can be changed to 3 months, 36 months or indefinite, and chats reviewed by human reviewers are retained for up to three years even if you delete your activity. Google adds that even with Keep Activity off it still uses chats to respond and to protect its services, including with human reviewers.

The Workspace version is different. Google's Generative AI in Google Workspace Privacy Hub, updated October 6, 2026, says that for qualifying Workspace editions your content is not human reviewed or otherwise used for generative AI model training outside your domain without permission, and that prompts are retained from 90 days to indefinitely as admins decide for Gemini in Workspace and up to 36 months for the Gemini app. The same page says Gemini Notebook, formerly NotebookLM, does not retain prompts and responses after the session ends, which matches the account-edition caution in our NotebookLM review and Gemini review. The point for a small team is that a personal Gmail login and a Workspace login to the same assistant are governed by different pages.

Workspace Suites: Microsoft 365 Copilot, Slack, Notion and Zoom

Microsoft's data, privacy and security article for Copilot, last updated November 4, 2026, states twice that prompts, responses and data accessed through Microsoft Graph are not used to train foundation LLMs, including those used by Microsoft Copilot, and that the product follows Microsoft's existing commitments to Microsoft 365 commercial customers. It also names Anthropic and OpenAI as subprocessors for some experiences. There is no training switch to find because there is no training default to turn off; the control that matters is who has permission to see which files, since Copilot surfaces anything a user can already view.

Slack's privacy principles page is the one most often misread. It says Slack does not develop generative AI models using Customer Data, that AI in Slack uses third-party LLMs that do not retain or train on that data, and that Slack will not use Customer Data to train generative AI models without affirmative opt-in consent. It also says that to develop predictive models for features such as emoji and channel recommendations, its systems analyse Customer Data including messages, content and files. To exclude a workspace from those global models, an Org Owner, Workspace Owner or Primary Owner must email feedback@slack.com with the workspace URL and the subject line 'Slack Global model opt-out request'. There is no toggle, and an ordinary member cannot do it. Our Slack review covers the plans.

Notion's AI security and privacy practices page says that by default neither Notion nor its AI providers use Customer Data to train any models, that its contracts prohibit providers from doing so, and that using Notion AI grants Notion no licence to train its own. What differs by plan is retention rather than training: on Enterprise most models offer zero data retention, while on Free, Plus and Business the AI provider retains data for 30 days or fewer before deleting it, and embeddings are deleted within 60 days of a page or workspace being deleted. Our Notion review covers the tiers.

Zoom's AI whitepaper, last updated November 10, 2026, states that Zoom does not use any customer audio, video, chat, screen sharing, attachments or other communications-like customer content such as poll results, whiteboard and reactions to train Zoom's or its third-party AI models, and that it has zero data retention arrangements with third-party model providers with limited trust and safety exceptions. Administrators can also disable or zero-out transcript retention for meeting summaries. The remaining exposure in a meeting is usually not Zoom but whatever note-taker a participant has invited, which our Zoom review and AI note-taker guide discuss.

Writing, Design and Presentation Tools: Grammarly, Figma and Gamma

Grammarly's Product Improvement and Training control article is unusually explicit about defaults by purchase route. For Grammarly Free, Premium and single-user Pro accounts the control is on by default and can be switched off in account privacy settings. For multi-user team accounts bought directly through the website it is initially on and an admin can opt the whole organisation out. For Grammarly for Education, Business, Pro and Enterprise accounts purchased through its sales team it is automatically off. A user who joins a team account adopts the admin's setting, and even with the control off Grammarly says it may still use non-content data such as writing statistics. Because Grammarly runs wherever you type, this one setting covers email, documents and chat at once; see our Grammarly review.

Figma's AI settings article says content training is turned on by default for Starter teams and for Professional teams, and that admins of those teams can toggle it off at any time in team settings, after which new content and edits are not used to train AI models. On Organization and Enterprise plans content training is off by default and the article notes it is currently disabled and cannot be toggled on. The setting is separate from whether AI features are enabled at all, so a team can keep the features and decline the training.

Gamma's data control article, dated September 24, 2026, follows the same split. On individual Free, Plus, Pro and Ultra workspaces data is allowed to be used to train AI models by default, with an opt-out under Settings, Data Controls. On Business and Enterprise workspaces content is automatically excluded and the setting is locked. Our Gamma review notes the consequence for a freelancer building a client deck on a personal plan: the opt-out exists, but it is the user's job to find it before uploading the client's material.

Meeting and Video Tools: Otter.ai and Vidyard

Otter.ai's privacy and security page answers the training question in its FAQ: Otter uses a proprietary method to de-identify user data before training its models so that an individual user cannot be identified, the process is automatic so recordings and transcripts are not manually reviewed by a human, and the training data is encrypted. The page also says explicit consent is required before any human at Otter reviews a recording. What it does not document is a self-serve switch to exclude an account from that training, which is why our Otter.ai review advises buyers who need recordings excluded to ask Otter in writing or evaluate its Enterprise terms.

Vidyard handles the question in its contract rather than in settings. Section 4.3.b of its Self-Serve Terms of Service, last updated November 28, 2025, has Free-plan customers grant Vidyard a licence to use their content for internal analysis and to train, test and improve AI models, and, subject to consent for customers outside the United States or opt-out choices for those inside it, to share content with third parties for their AI training. The clause is specific to the Free plan; paid-plan content is licensed for operating the service. Our Vidyard review sets out the plans, and the practical reading is that the opt-out for a recording that names a customer is the $59 Starter seat.

Meeting tools carry a second-party problem the other categories do not. A recording contains the words of people who never accepted the vendor's terms, and Otter's own page tells users they must comply with local law and ask for consent when recording. A team policy that covers only its own accounts misses the client or candidate on the other side of the call, so the audit below treats 'which tools record other people' as its own row.

LinkedIn: The Setting That Is On by Default

LinkedIn's setting article states plainly that the Data for Generative AI Improvement member setting is set to on by default unless you turn it off. Turning it off means LinkedIn and its affiliates will not use the data and content you provided to train models that generate content going forward, and does not affect training that has already taken place. The setting covers content-generating models only; models used for personalisation, security and anti-abuse are outside it and need a separate Data Processing Objection form.

Its generative AI FAQ lists what is used: profile data, content entered into generative AI features, saved resumes and screening answers, group activity and group messages, posts, articles and comments, and feedback. It lists what is not: private messages including InMail, login credentials, payment data and salary or job-application data attributable to a member. It says the data of members it believes are under 18 is not used even if the setting appears on, and that its European Regional Privacy Notice was updated on November 3, 2025 to cover this use for members in the EEA, Switzerland and the UK. For a business, the relevant accounts are the personal profiles of staff who post on its behalf, which no company admin console controls.

The Feedback Trap: When a Thumbs-Up Overrides Your Opt-Out

The most common way an opted-out account still contributes to training is the feedback button. OpenAI's article says that even if you have opted out, if you submit feedback on a response on an individual plan, for example a thumbs up or thumbs down, the entire conversation associated with that feedback may be used to train its models. LinkedIn says the same of its setting: if you choose to provide feedback, the information in it may still be used to train content-generating models even if you opted out. Anthropic says feedback submitted with the thumbs buttons stores the entire related conversation for up to five years and may be used to train its models, on consumer and commercial products alike.

Business tiers narrow this but do not always close it. OpenAI states that for ChatGPT Business, Enterprise and Edu, response feedback does not permit training on the associated conversation, and workspace admins can disable response feedback altogether. Anthropic says owners of Team and Enterprise plans can disable the 'Rate chats' setting under Organization settings, Data and Privacy, which removes the button for members. Zoom's whitepaper says that for ZoomMate feedback users choose whether to attach the prompt, response, screenshots and diagnostics, are shown what will be shared before submitting, and that account administrators can control whether that context may be included.

The rule for a team is short: on any plan where you have opted out, treat the rating buttons as a consent form for that conversation, and where an admin switch exists to remove them, use it. A member who rates a response in a thread containing a customer contract has, under the vendors' own wording, shared that thread.

Opting Out Does Not Undo Training or Stop Retention

Opt-outs are prospective. Anthropic's settings article says that after you turn the setting off, new chats are not used for future training but your data will still be included in model training that has already started and in models that have already been trained. LinkedIn says opting out does not affect training that has already taken place. OpenAI says that after you opt out it will not use your new conversations. None of the three describes removing past contributions, so the date you change the setting is the date that matters, and the audit records it.

Training and retention are separate controls. Anthropic's retention article, dated July 1, 2026, says that if you allow model improvement your data may be retained in de-identified form for up to five years in training pipelines, that deleted conversations are removed from back-end storage within 30 days, and that flagged content is kept for up to two years. Google keeps human-reviewed Gemini chats for up to three years regardless of deletion. Notion's providers retain Free, Plus and Business AI data for up to 30 days even though nothing is trained. A tool that never trains on your data can still hold it for a month or more.

Scope is the third limit. OpenAI's ChatGPT toggle does not change the Codex environments setting. LinkedIn's setting does not cover non-generative models. Grammarly's control does not stop use of non-content data. Google says Gemini settings do not control processing of chats to create anonymised data to improve its services. Each opt-out is narrower than its name, and the vendor's own sentence about what it excludes is the part worth copying into your record.

What Privacy Costs: When the Safe Default Is a Paid Tier

For several tools the documented way to get a locked no-training default is to buy a higher plan, and that has a price a small team can compute. Gamma's Business workspace, where training exclusion is locked, is quoted in our Gamma review at $40 per member per month billed annually, so five members cost 5 × $40 = $200 a month, or $2,400 a year, against an individual plan where the same protection is a free toggle each person must remember. Vidyard's Starter seat, which takes recordings out of the Free-plan training licence, is $59 per seat a month on annual billing per our Vidyard review, so five seats cost 5 × $59 = $295 a month, or $3,540 a year.

The cheaper route is usually the toggle, and for most vendors in this guide it is free: ChatGPT's data control, Claude's privacy setting, Grammarly's control, Figma's team setting, Gamma's individual opt-out and LinkedIn's member setting cost nothing but the minutes to find them. The paid tier buys three things the toggle does not, which are a default nobody can forget to change, a control one admin sets for everyone, and in most cases a contractual commitment rather than a help-centre statement. Whether that is worth $2,400 a year depends on what goes into the tool, which is the judgement our SaaS evaluation checklist asks buyers to make before choosing a tier.

A workable rule is to classify by content rather than by tool. Material that names customers, employees or unreleased plans belongs on a tier with an admin-enforced default or stays out of the tool; everything else can live on an individual plan with the opt-out switched on and recorded. That keeps the paid seats to the people who handle the sensitive material, which for a five-person team is often one or two rather than five.

A 45-Minute Audit for a Small Team

Minutes 0 to 15, inventory and classify. List every tool where staff type, upload, record or design, including the ones bought for another purpose: the chat app, the meeting recorder, the grammar extension, the design tool, the deck builder and personal LinkedIn profiles used for work. For each, write the plan name from the billing page, because the plan decides the default, and whether the login is a personal or organisation account. Mark the tools that hold customer, employee or confidential material, and the tools that record people outside the company.

Minutes 15 to 35, read and configure. For each marked tool open the vendor page linked in this guide and copy the sentence that states the default for your plan, with the page's date. Then change what can be changed: the individual toggles for ChatGPT, Claude, Grammarly, Gamma and LinkedIn; the admin settings for Figma teams and Grammarly teams; the feedback switches in ChatGPT workspaces and Claude organisations; and, for Slack, the owner's email to feedback@slack.com with the documented subject line. Where there is no switch, as with Otter.ai's de-identified training or Vidyard's Free-plan licence, decide between a paid tier, a written request to the vendor, or keeping sensitive material out.

Minutes 35 to 45, verify and maintain. Record the date each setting was changed, since opt-outs are prospective, and take a screenshot of the setting in its off state. Write two lines of team policy: no rating buttons on conversations containing confidential material, and no confidential material in tools on personal plans. Then set a reminder to repeat the audit in six months and whenever a plan changes, a new tool is added or a person joins, because a new member's individual account starts on that vendor's default, not yours. The one-page usage policy in our AI tools for small business guide is the natural home for those two lines.

Methodology, Limits and What This Guide Does Not Claim

This guide is documentary. Every statement about a vendor comes from that vendor's own help centre, privacy page, whitepaper or terms as read on November 12, 2026, and each is linked where it is described. No account was created, no setting was changed, no vendor was contacted and no claim about what a vendor does in practice is made beyond what it has written. Settings paths are quoted from the documents and may differ by app version, region or account type. Canva was researched for this guide but is not included, because its official help page on this topic could not be retrieved and verified on the day.

The guide does not claim that training on customer data is improper, that any vendor fails to honour its stated settings, or that an opt-out makes a tool suitable for regulated data. It is not legal advice: data-protection law in your jurisdiction, a signed data processing addendum or an enterprise agreement can impose stricter or different terms than the public pages quoted here, and several vendors note regional differences, including LinkedIn for the EEA, Switzerland and the UK. These pages change often, several of them within the month before this guide was written, so recheck the linked source before relying on any default.