A language model just 280 megabytes in size, running directly on an Android smartphone, can carry on a therapeutic conversation at 17 tokens per second — without a single byte of information ever leaving the device. This is not a concept deck or a conference slide: the system, called MoPHES, was built by researchers Wei, Zhou, and Wang and published in an IEEE journal in 2025. It is a working answer to a question that sits under every mental-health app: where does your most private data go?

Why therapy data is the most sensitive data you produce

Think about what you tell a therapist — the thoughts you don't say out loud to anyone else. Now add what a mental-health app can observe: not just your words, but potentially your location traces, your sleep, your communication patterns. Digital phenotyping research treats exactly this stream — GPS, communication, sleep, and social activity — as among the most sensitive data a person generates, and we cover how much of it a phone can read in our piece on digital phenotyping. It is far more revealing than a credit-card number, and unlike a password, you can't change your history after a breach — a leaked therapy transcript is exposed permanently.

The fear of that exposure is itself a barrier to care. According to the WHO, more than 70% of people with mental-health conditions never seek help, and stigma and fear of leaks are among the reasons. That fear is well-founded: even in research settings, ethics boards restrict the sharing of real therapeutic session data. A study by Song and colleagues (2024) found that people are willing to open up to an AI conversational partner — but only when they're confident their words won't be read by a third party. Privacy isn't a nice-to-have here; it's a precondition for the honesty that makes therapy work at all.

Cloud vs. on-device: what actually differs

Traditional online mental-health services, including chatbots that have demonstrated effectiveness in clinical trials, rely on the cloud. Every message travels to a remote server, is processed, and comes back. Even with end-to-end encryption in transit, the data is stored somewhere — and what is stored can, in principle, be breached, subpoenaed, or repurposed. An on-device model runs on your phone: no server, no cloud, sometimes no internet connection at all.

Dimension	Cloud AI	On-device AI
Data location	Remote servers	Stays on your phone
Breach surface	Every stored conversation	Only your own device
Model size / capability	Very large, most capable	Compact, narrower
Offline use	Needs a connection	Works with no internet

MoPHES makes the on-device side concrete. It uses two compact 0.5-billion-parameter modules run through the llama.cpp framework; after Q4_K_M quantization each takes about 280 MB — less than an average mobile game. On the test device, a Xiaomi 13 Ultra, it generates responses at 17.3 tokens per second, and a mental-state assessment takes 4.2 seconds — fast enough that the user feels no lag. Splitting the work in two matters: one module handles empathetic, flexible dialogue while the other stays rigorous and structured, saving assessment results to a local file and retrieving session history from local memory. All of it happens without a single server call. The trade-off is real, though: a 0.5-billion-parameter model is far less capable than a large cloud model, so on-device tools handle screening and protocol-based support well but aren't yet sufficient for complex psychotherapeutic work.

There's an access dimension too, and it cuts both ways. The WHO estimates that nearly a billion people worldwide need mental-health support and most never receive it — because of a shortage of professionals, cost, distance, or fear of judgment. A private, offline tool that runs on a phone already in someone's pocket lowers several of those barriers at once. But not every phone has the memory MoPHES was tested with, so mass adoption likely means even smaller models or a hybrid design: basic functions on-device, with heavier capabilities in the cloud only with explicit consent. The privacy win is strongest when the sensitive core stays local, whatever runs elsewhere.

What regulators require today

Where your data goes is increasingly a legal question, not only an engineering one. The EU's AI Act (Regulation 2024/1689, in force since 1 August 2024) classifies health AI as high-risk, requiring conformity assessments, technical documentation, human oversight, and risk management — with fines reaching €35 million or 7% of global annual turnover. In the United States, the picture is looser and less reassuring: the FDA has cleared roughly 97% of AI devices through the 510(k) pathway, but a 2025 analysis in npj Mental Health Research found that many cleared devices lack direct evidence of efficacy.

The mental-health track record is sobering. Rejoyn (March 2024) became the first FDA-cleared prescription app for depression, yet its trial showed no statistically significant difference from control in the intention-to-treat analysis. Pear Therapeutics — maker of reSET, the first FDA-authorized digital therapeutic — went bankrupt in 2023, raising hard questions about whether these business models are sustainable. And as of late 2025, no generative-AI mental-health device has full FDA approval at all. On the data-rights side, GDPR grants a right to erasure and data portability, while the US HIPAA framework governs health data differently; reconciling the two is a genuine technical and legal challenge, which is part of why researchers propose "dynamic consent" models that let people continuously choose what they share. An on-device architecture sidesteps much of this by design — if the data never leaves the phone, there is far less to govern, leak, or subpoena, and far less that a future change of ownership or policy can put at risk.

Questions to ask any AI therapy app

You don't need to read a model card to protect yourself. A few plain questions separate a careful product from a careless one:

Where is my data processed — on my device, or on the company's servers?
What is stored, and for how long? Transcripts, mood assessments, location?
Can I delete everything, and does deletion actually remove it from backups?
Is my data ever used to train models or shared with third parties?
Does it work offline? If not, everything you type is leaving the device.
Is the app honest that it's an AI, and does it have a plan for a crisis?

If an app can't answer these clearly, that opacity is itself an answer. The broader ethics of building these tools responsibly — beyond privacy alone — is the subject of our piece on ethics in AI psychotherapy, and the safety mechanisms every serious product needs are covered in our guide to guardrails for mental-health AI.

MoPHES is the first fully autonomous AI mental-health support system that runs on a mobile device, and it demonstrates that privacy and accessibility don't have to be in conflict. As quantization and mobile chips keep improving, on-device models will grow more capable — and the line between a laboratory experiment and an everyday self-care tool will keep blurring. Services like Nearby use evidence-based approaches to support mental health, with the honest caveat that a phone-based tool complements, rather than replaces, a human professional.

FAQ

Is my AI therapy chat private?

It depends entirely on where the model runs. With a cloud service, your messages travel to and are stored on remote servers, so "private" means "as private as that company's security and policies allow." With a true on-device model like MoPHES, the conversation never leaves your phone — there is no server copy to leak. The single most useful thing you can do is check whether processing happens on-device or in the cloud, and read the privacy policy for what's stored and shared.

Can AI therapy apps sell my data?

Some can, depending on their policy and jurisdiction — which is exactly why the policy matters. Regulations like GDPR (right to erasure, data portability) and HIPAA constrain this, and the EU AI Act now treats health AI as high-risk with heavy fines for violations. But protections vary by country and by app. On-device processing is the strongest technical guarantee: data that never leaves your device is data that can't be sold.

Does on-device AI mean worse quality?

For now, somewhat — a 280 MB model is far less capable than a large cloud model at open-ended, nuanced conversation. But for the tasks these tools are actually suited to — structured mood assessment, protocol-based supportive dialogue — compact models are already good enough to be useful, and they're improving quickly. The realistic framing is a trade: you give up some conversational depth in exchange for a privacy guarantee the cloud can't match.

Nearby is a support tool that uses evidence-based psychology. It does not replace a psychologist, psychotherapist, psychiatrist, or emergency service.