Gamma's help page says two things on the same screen.
For individuals on Free, Plus, Pro and Ultra: "Data is allowed to be used to train AI models." For Business and Enterprise accounts, the same data is "automatically excluded" and the setting is locked so nobody can switch it back on.
Same product. Same models. Opposite default. The only variable is what you pay.
I spent today reading the privacy policies, terms and help pages of 21 AI tools to find out how common that is. It is very common.
The one question I asked
For each tool: can a reader find out whether their own work is being used to train the model, without logging in and without emailing support?
Not "is it possible in theory." Not "what does the company say about security." Just: does the page tell you what is happening to your work right now.
Of the 21, I could read 19. OpenAI and Midjourney returned 403 from every page I tried, which is a bot wall rather than a dead site, so ChatGPT, the most used tool on the list, is missing from every count below. I am not filling that gap from memory.
Eight of the nineteen answer the question. Eleven do not.
The eight that answer
| Tool | What the page actually says |
|---|---|
| Grammarly | "Product Improvement and Training is on for individual users by default" |
| Gamma | "Default setting: Individuals (Free, Plus, Pro, and Ultra): Data is allowed to be used to train AI models" |
| Microsoft Copilot | trains "except for certain categories of users or users who have opted out" |
| DeepL | free version is used "to train and improve our neural networks and algorithms"; Pro texts "will not be used" |
| Notion | "By default, your customer data is not used to train models" |
| Cursor | "We do not use Inputs or Suggestions to train our models" |
| Zoom | "Zoom does not use any of your audio, video, chat, screen sharing, attachments" to train its or third-party models |
| Adobe | "Adobe does not analyze your content to train generative AI models, unless you choose to submit content to the Adobe Stock marketplace" |
Note what the clear answers have in common. Four of them are a no. The companies that do not train on you tend to say so in plain words, early, and without conditions. That turns out to be a useful signal on its own.
The eleven that do not
Gemini, Claude, Perplexity, Canva, Jasper, ElevenLabs, GitHub Copilot, Otter.ai, Descript, Runway and Suno.
Most of them are not hiding anything. They simply stop one sentence short. Runway's terms say "Inputs and Outputs may be used by the Company to train and improve its AI models." May. Suno says its data is used "to train and enhance the models that power our Services." Otter.ai describes "training our proprietary AI technology on de-identified audio recordings and on transcriptions." All true, none of it tells you whether it is happening to your account.
Gemini is the sharpest example. Google states clearly that with the Keep Activity setting on, "Google uses your activity to provide, develop, and improve its services (including training generative AI models)," and that with it off, future chats "won't be used to train our AI models." Both halves are documented. What Google never says, anywhere on that page, is whether Keep Activity is on or off when you start.
The part nobody has written down
Seven companies publish a different default for consumers than for business customers, on their own pages, in writing.
| Company | If you are a freelancer | If you are an enterprise |
|---|---|---|
| Gamma | on by default | "automatically excluded", setting locked |
| Grammarly | "on for individual users by default" | off, "with no toggle visible" |
| ElevenLabs | not stated | "By default, we don't train on any data from our Enterprise customers" |
| Anthropic | used "unless you opt out" | "By default, we will not use your inputs or outputs from our commercial products" |
| Descript | used "unless you have opted out" | Enterprise drives "disabled by default" |
| DeepL | free text trains the networks | Pro text "will not be used" |
| OpenAI | could not read | API data "is not used to train or improve OpenAI models" |
Seven out of seven run the same way round.
This is not a scandal and it is not illegal. Enterprise contracts are negotiated, and the people negotiating them ask this question. The point is narrower and worse: the protection exists, it works, it is already built, and whether you get it depends on which invoice you are on. A freelancer pasting a client's confidential brief into the Free tier and a company pasting the same brief into Enterprise are covered by opposite rules.
If you have ever signed an NDA and then used a free AI tool to do that work, this is the table to read twice.
Two companies contradict themselves
Anthropic. The privacy policy says: "We may use your Inputs and Outputs to train and improve Anthropic AI models, unless you opt out through your account settings." The consumer terms repeat it. Both read as on unless you act.
Anthropic's own consumer help article says the opposite. Training happens if "You choose to allow us to use your chats and coding sessions to improve Claude."
Opt out and choose to allow are opposite defaults. Both pages were live on 2 November 2026.
Descript. Its terms say content is used to train "unless you have opted out of such use." Its help article says internal research models use "only data from users who have opted into data sharing." Same contradiction, same day.
In both cases the answer probably exists inside the product, in a settings screen I would have to log in to see. That is the problem. You should not have to create an account to find out what the company will do with your work.
Six tools use your content and give you no switch
Perplexity, Otter.ai, Runway and Suno describe no training opt-out of any kind on the pages I read. Perplexity's privacy notice does carve out one thing: "We do not use the content of emails to create, train, improve, or fine-tune AI models." Your emails are protected. Your searches and uploads are not mentioned.
Jasper has an opt-out, but only if you live in the EU or UK, and it is not a setting. Its terms say you "submit a request by contacting [email protected]," after which Jasper "will cease using your Customer Property for model training and improvement."
For DeepL, the opt-out is to buy Pro.
Read Jasper's trust page slowly
It says: "Your company data and outputs are never used to train third-party LLMs."
Its terms say Jasper uses your input and output "for the purpose of developing and enhancing Jasper's artificial intelligence models."
Both sentences are true at once. The trust page is about other people's models. Jasper's own are a different sentence, in a different document, that the trust page does not link to. Nothing here is false. The word doing the work is "third-party," and it is doing a great deal of work.
Opting out is not an off switch
Two companies say so themselves.
Anthropic: even after you opt out, material is still used when a conversation is flagged for safety review or when you submit feedback, and feedback conversations are stored "for up to 5 years."
Google: "Your Gemini settings don't control processing of your chats to create anonymized data." And chats already seen by a human reviewer "are retained for up to three years" even after you delete your activity. Turning the setting off stops future chats. It does not reach back.
An opt-out narrows what happens to your work. On the evidence of these pages, it does not stop it.
Sometimes you cannot read the policy at all
Notion gave the clearest no-training statement in the whole audit. It is on a help-center page with no date on it. Notion's actual privacy policy, at notion.com/privacy, redirects to app.notion.com/privacy and returns 401 Unauthorized. You need an account to read the terms you are agreeing to.
Nine of the 21 companies blocked at least one page from automated reading. Gamma's privacy policy and terms both returned 403, so everything I know about Gamma comes from a help article. And then there is the staleness. Google's Generative AI Additional Terms page is still served, still says "Last Modified: August 9, 2023," and says on its own face that it stopped applying on 22 May 2024. GitHub's Copilot privacy documentation returns 404, and its Copilot product terms are marked "deprecated effective 5 March 2026" and are still the page you land on.
The two minute check, for any tool
Search the privacy policy for "train". Then search the terms for "train" separately. The answer often lives in only one of them. Jasper's privacy policy is silent on model training while its terms are explicit. Runway is the same. Reading one and not the other gets you the wrong answer.
Then search both for "default". This is the word that separates the eight from the eleven, and almost nobody uses it.
Check whether the sentence says "third-party". If a company promises not to train third-party models, it has told you nothing about its own.
If you cannot find a default stated anywhere, assume you are being trained on. The four companies that do not train on customer content all say so plainly and early. Silence is not neutral here.
Sourcing
Twenty-one AI tools, read on 2 November 2026. For each I read the privacy policy, the terms of service and the relevant help-center articles on the company's own domain, and asked the same question of every one, so the counts are comparable. Every quotation above came off a company page, not from reporting about it.
Read and counted (19): Claude, Gemini, Microsoft Copilot, Grammarly, Perplexity, Canva, Notion, Jasper, ElevenLabs, Cursor, GitHub Copilot, Otter.ai, Descript, Runway, Suno, Adobe, Zoom, Gamma, DeepL.
Could not be read (2): OpenAI and Midjourney returned 403 from every page tried, across help.openai.com, openai.com, docs.midjourney.com and midjourney.com. These are live pages behind bot walls, not dead ones. They are excluded from every count rather than guessed at. If you want ChatGPT's answer, the help article is titled "How your data is used to improve model performance" and opens fine in a normal browser.
Partial reads. Gamma's privacy policy and terms both returned 403, so its entry rests on a help article dated 24 September 2026. Notion's privacy policy returned 401, so its entry rests on an undated help page. GitHub's Copilot-specific privacy page returned 404, which is why GitHub Copilot sits in the eleven: the document that would answer the question is gone.
Policy dates, as published. Gemini 26 October 2026. Perplexity and Runway 23 October 2026. Cursor 11 October 2026. ElevenLabs 9 October 2026. Anthropic 10 September 2026. Zoom 17 September 2026. Microsoft September 2026. Canva 25 August 2026. Grammarly, now published under Superhuman Platform Inc., 6 July 2026. Suno 23 June 2026. Otter.ai 16 June 2026. Adobe 24 October 2025. Jasper 30 May 2025. Descript 14 April 2025.
Policies change, and the ones above were read on a single day. The method in the two minute check will outlast the table.
One last thing. The clearest, most complete, most plainly worded answer in this entire audit came from Notion, and you cannot read Notion's privacy policy without an account.