How AI Tutoring Platforms Track Your Thinking (And What Their Privacy Policies Don't Say)

The most striking finding is a universal silence. We asked each platform a simple question: can cognitive process records be reconstructed after a session ends? In other words, after a learner closes their session, can the platform rebuild a picture of what that person was struggling with, where they hesitated, what they explored and abandoned?

Not one of the eight platforms addresses this question in their privacy policy.

This matters because AI interaction data is fundamentally different from traditional web browsing data. When a student works through a problem with an AI tutor, the conversation itself is a record of their cognitive process: the confusion, the false starts, the moments where understanding breaks down. This is not equivalent to a search history. It is a map of how someone thinks.

Magee, Ienca, and Farahany (2024), writing in Neuron, define cognitive biometric data as information that can be processed to infer mental states. By that definition, AI tutoring conversations qualify. Every prompt a struggling learner sends contains signals about their cognitive state: what they understand, what confuses them, where their reasoning breaks down. The platform captures all of it. None of the eight policies we reviewed address what happens to that cognitive layer after the session ends.

Two platforms (Duolingo and Coursera) are transparent about the depth of their behavioral tracking. Duolingo discloses the use of FullStory, a session-replay tool that logs clicks, mouse movements, scrolling, and typing. Coursera similarly acknowledges session-replay tools that capture mouse movements, page visits, and console logs.

The other six platforms use umbrella language. ChatGPT collects "how you interact… features you use and the actions you take." Khanmigo logs "information about your use of our Service and your learning activity." Google Gemini collects "interaction logs, performance metrics." Microsoft Copilot captures "interactions with bots… performance, usage, and device data."

What none of these vague descriptions clarify is whether the platform measures granular behavioral signals during AI sessions: typing speed, backspace frequency, pause duration, revision patterns. These are exactly the signals that indicate cognitive state. A student who types a question, deletes it, retypes it differently, and pauses for thirty seconds before submitting is telling the system something about their confidence, their confusion, and their reasoning process.

Whether platforms are capturing these signals is genuinely unclear from their policies. And that ambiguity is itself the problem.

Nearly every platform we reviewed admits to using interaction data to "personalize" the experience. Khanmigo uses data to understand how usage "impacts learning outcomes." Duolingo tailors lessons to "include more words that a user is struggling with." Google Gemini generates "personalized insights" from chat history. Character.AI explicitly "generates and analyzes inferences… such as your preferences and interests."

All of this involves inference about cognitive and psychological states. The platform is deducing what the learner knows, what confuses them, what engages them, and what they're likely to do next. These are inferences about mental states derived from behavioral data.

Of the eight platforms reviewed, only one (Character.AI) explicitly classifies these inferences as protected personal data. The other seven do not address whether inferred cognitive states receive the same protection as directly provided information like a name or email address.

This gap maps onto a structural problem in privacy law that Solove (2025) identifies in the Florida Law Review: current frameworks regulate input (data collected) rather than output (inferences made). The practical consequence for learners is that the most sensitive information a platform holds about you (what you struggle with, what confuses you, how your reasoning works) is also the least protected.

Every platform we reviewed has some form of age-gating. Anthropic bans all users under 18. ChatGPT and Coursera ban users under 13. Character.AI prohibits users under 13 (16 in the EU/UK). Khanmigo and Microsoft Copilot offer specific child account structures with parental controls.

But age-gating is a blunt instrument when the risk is cognitive, not just informational. Standard COPPA compliance protects children's personally identifiable information (name, email, location). It was not designed to address what happens when a child's developing reasoning process becomes training data for an AI model.

Only two platforms go beyond age-gating toward substantive protections. Microsoft Copilot for Education explicitly forbids building student profiles or targeting ads, making it the strongest policy position in the group. Khanmigo warns users against including personal information in AI inputs and prohibits its AI vendors from using student input data to train models.

The other six platforms default to standard compliance without addressing the specific risks of AI interaction data from minors. None address the documented tendency of young users to form emotional bonds with AI systems. Research from Internet Matters (2025) found that 40 percent of teens who use AI companions trust their guidance without question. Yang and Oshio (2025), writing in Current Psychology, have documented how users form measurable attachment bonds with AI systems using the same psychological framework applied to human relationships.

A platform that captures a child's cognitive struggles, uses that data to personalize engagement, and retains it across sessions is not just collecting information. It is building a persistent model of a developing mind. Current privacy policies do not acknowledge this distinction, and current privacy law does not require them to.