Which education apps train AI on your data?
Of the 7 education apps on file, 4 are graded: 3 train on your data by default and 1 doesn't. The other 3 have a policy that doesn't say either way. The best graded is Udemy (B) and the worst is Chegg (F).
Do they train on your coursework, your answers and your voice?
Every story about AI in education is about students handing in what a chatbot wrote. The apps here hold the other side of the exchange. The coursework you submit for help, the assignments and forum posts you hand in, the study sets you build, and recordings of you speaking a language you don’t yet speak well. They’re graded on one narrow question: does that material train a model, and how hard is it to stop. Read together, they divide by what their own documents say. The ones that say they train and offer no switch, or half of one. The ones that say plainly they don’t. And the ones whose only word on training is about somebody else, or nothing at all.
The ones that say they do
Chegg’s clause reaches past Chegg: “We and/or our business partners” may use what students submit through AI Functions “to train or fine-tune our artificial intelligence and machine learning models”. The material is coursework, from people usually paying for help with it. The policy says you “may have the right to opt out of such use depending on your jurisdiction”. In the regional section that’s the GDPR objection, for the EEA, UK and Switzerland. Everyone else gets no setting, no address and no plan on which this is off.
Babbel says learner data “may also be used to train our AI and to further develop and optimize our services”. The basis is legitimate interest, so it runs unless a reader objects, and there’s no setting for that. Voice recordings are used “in anonymized form for training purposes”. That’s a limit on the form the data takes, not a control.
Duolingo may record “the text and audio you submit” and use it to train and run its own AI models. A switch exists: you “may also choose not to share your audio with us for product improvement purposes within the app Settings”. It reaches audio only, the policy never connects it to training, and it doesn’t say which way it ships. For the text you type there’s no setting, only the right to object. This entry was corrected after Duolingo replied. The reply is printed on the verdict.
The ones that say they do not
Udemy’s is a plain denial for learners. Inputs to its GenAI features are processed to answer them, “but we do not use your personal data to train or fine tune GenAI models”. Off by default is what a B is on this site. The purposes table separately mentions improving the Services with AI consistent with an Instructor GenAI Policy, a separate document this site hasn’t read, which concerns the course material instructors publish. It isn’t a use of learners’ data.
The ones whose documents never say what they do
Coursera’s privacy notice devotes a section, “How We Use AI”, to naming its features, dubbing, flashcards, roleplay, proctoring and chatbots, and discloses that AI may affect your grade and that you can elect a human grader instead. What it never says is whether the assignments, forum posts and answers you submit train those models. The only clause on training, in its terms, points at you: learners may not “develop or train any application, software, code, or data models” on Coursera’s content. The obligation runs one way.
Khan Academy’s one training statement is a real protection, and it points outward. Its service agreements with the providers of the AI models behind its features “do not permit the providers to use Input data to train their models”. It’s scoped to those providers. Whether Khan Academy itself uses learner input to train or fine-tune anything, the policy never says, and an assurance about a supplier is not an assurance about the party giving it. Grading it good would credit the document with a promise it doesn’t make.
Quizlet sells an AI tutor and AI-generated study material, and its privacy policy doesn’t contain the words train, model, machine learning or artificial intelligence anywhere. The closest it comes is that service providers “help us improve our services”. A policy that doesn’t raise the subject can’t be read either way, so whether your study sets and chats train anything is a finding of silence, not an inference.
Side by side
One pair of education apps people choose between, with both verdicts on one page.