Which platforms train AI on your data?
Of the 8 platforms on file, 5 are graded: 3 train on your data by default and 2 don't. The other 3 have a policy that doesn't say either way. The best graded is Apple (B) and the worst is Google (F).
Do they train on what you search, buy, say and listen to?
The companies on this page aren’t apps with one kind of data. Between them they hold what you search, what you buy, what you say to a speaker, what you listen to and what you publish. Each is graded here on a narrower question than the usual one about who sells it: does that data train a model, and how hard is it to stop. Read together, they divide by what their documents say. Some say they train, and offer a route out or nothing at all. Some promise not to, in a sentence with edges worth reading. And the ones that hold recordings of you never settle the question.
The ones that say they train
Google’s answer depends on where you read it. The European text says it uses “your interactions with AI models and technologies like Gemini Apps to develop, train, fine-tune, and improve these models”. The text served in the United States names only publicly available information, and neither repeats nor denies the rest. No control over AI training is named in either version. What the European reader has is a right to object under data-protection law, exercised by writing to Google, not by moving a setting.
Spotify’s clause is about listening, not anything you make: “development and training of algorithmic and machine learning models, which help us to keep our service safe and enhance our users' experience”. That’s narrower than most entries on this site, and the record says so. Every control the policy names is for tailored advertising or for withdrawing consent, and consent isn’t the basis it gives for the training use. The document graded is the one Spotify serves in the United States. The European text is unverified.
Microsoft says it “may use your data to develop, train, and fine-tune our AI models, including large language models (LLMs)”, and that sentence names no control. The one opt-out in the statement is narrower: Copilot conversation data “can help train our AI models in Microsoft Copilot unless you opt out”, and it opens “In some markets” without ever naming them. The switch itself, Training on conversation activity under Privacy in Copilot, is set out on the verdict, and covers future conversations only.
The ones that promise not to, with an edge
Apple’s sentence is short: “Your private personal data is not used to train our foundational AI models.” Two qualifiers in it do the work. “Private personal data” isn’t defined in the capture, and “foundational” is narrower than all models. The sentence doesn’t speak to feature-level ones. It’s a clear commitment as far as it goes. The one form the policy offers is about Applebot’s crawl and research datasets, not about training on your account.
Substack’s undertaking is in its help centre, not its privacy policy, which is silent: “Neither Substack nor Pangram uses publisher content to train generative AI models.” Pangram is its AI-detection vendor. The sentence covers posts and newsletters, and generative models. Readers’ data and other kinds of model aren’t addressed. The toggle in publication settings signals outside crawlers not to train on your content, and speaks only to the ones that respect it. Substack was written to and didn’t reply.
The ones that hold your voice and never settle the question
Amazon’s privacy notice contains no form of “train”, no “machine learning” and no bare “AI”. The nearest it comes is a legitimate-interests line about “when we use your voice, video, or camera input to improve services”, which names no model. It links onward to a “Generative AI Development Disclosure”, which describes the datasets behind its generative services as “licensed and proprietary datasets, synthetic datasets, open-source datasets, and publicly available content”. It never says whether your orders, searches or recordings are among the proprietary ones.
Samsung’s document is silent almost everywhere: no machine learning, no artificial intelligence, and every “model” is the model number of your handset. One phrase isn’t silent. Samsung records your video chats with support “for training and quality control purposes”. That’s the standard formula for staff review, and read that way it says nothing about AI. Read the other way it’s video and audio of you, used for training. The document doesn’t choose between the readings, and neither does the verdict.
Sonos’s newest statement is longer than the last and has more to be silent about. A voice upgrade now sends “voice recordings, transcripts, AI-inferred command intent, and command metadata” to its cloud for “AI-assisted processing”, where the classic voice control kept everything on the device. A section on AI partners shares that voice data with them, and ambient noise is collected “to help us improve our speech recognition technology”. None of it says train, model or machine learning. Improving a technology isn’t saying your recordings are the material.
Side by side
4 pairs of platforms people choose between, with both verdicts on one page.