The Day I Realised My Phone Knew More About My Mental Health Than I Did

A few months ago, I was reviewing research on digital phenotyping for one of our projects when something stopped me cold. A study showed that passive smartphone data, things like typing speed, screen time patterns, and movement frequency, could predict the onset of a depressive episode days before the person themselves felt the shift. The algorithm saw it coming before the human did.

That finding simultaneously excited and unsettled me. Excited because I understand how transformative early detection can be. Unsettled because it forced me to confront a question I had been orbiting for a long time: just because we can monitor someone's mental state passively, should we?

This is the story of how digital phenotyping has shaped my thinking about what it means to build ethical health AI, and why I believe the decisions we make in the next two years will define whether this technology liberates or surveils.

What Digital Phenotyping Actually Is

For those unfamiliar with the term, digital phenotyping refers to using passively collected data from smartphones and wearables to quantify human behaviour and infer health states. Your phone's accelerometer tracks your movement. Your screen usage reveals your sleep patterns. Your typing cadence can indicate cognitive load or emotional distress. Your GPS data shows whether you have been isolating.

None of this requires opening an app or answering a survey. It happens continuously, invisibly, in the background. That is both the power and the peril.

The science is compelling. A 2025 NIH-backed review analysed 42 peer-reviewed studies using machine learning on passively collected data to diagnose mental health conditions. The Accelerating Medicine Partnership Schizophrenia Study is collecting smartphone phenotyping data from over 40 global sites. By 2026, AI platforms for passive mental health monitoring are expected to become standard clinical tools.

As a founder in this space, I cannot ignore this trajectory. But I also cannot ignore what it means for the people whose data fuels it.

The Ethical Tightrope

My background in healthcare taught me that every clinical tool has a risk-benefit profile. A medication that saves lives also has side effects. A diagnostic test that detects disease also generates false positives. The question is never whether a tool is perfect. It is whether the benefits outweigh the risks, and whether the risks are managed responsibly.

Digital phenotyping's risk profile is unlike anything we have seen before. A 2025 survey found that nearly half of mental health apps share personal health information with third parties. Many users have no idea this is happening. The regulatory frameworks that should protect them are fragmented at best: the EU AI Act is imposing high-risk requirements, but the US has no comprehensive federal protection for wellness app data.

The proposed HIPRA Act would extend privacy protections to health and fitness apps, but until it passes, millions of people are generating deeply personal mental health data in a regulatory grey zone. As someone building products in this space, this keeps me up at night.

And it gets more complicated in the workplace. Employers are increasingly deploying AI-powered wellness tools that passively monitor employee communications, physiological data, and behavioural patterns. Research from the European Agency for Safety and Health at Work shows that this monitoring can actually worsen mental health, creating the exact problem it claims to solve. Employees under surveillance report feeling stressed, mistrusted, and psychologically unsafe.

When the Phone Knows Where the Missiles Fall

To understand the profound intimacy of mobile data, we don't even need to look at mental health apps first. We just have to look at how we survive in a crisis.

Recent events here in the UAE have made this incredibly literal. As regional tensions escalated into active conflict with Iran, UAE residents experienced the National Early Warning System in real-time. Designed by the Ministry of Interior and NCEMA, this system uses cellular broadcast technology to send high-tone, unignorable alerts directly to mobile phones.

But the most striking thing about these alerts is their precision. The system doesn't just broadcast blindly; it targets devices in the specific geographic footprint of incoming threats, guided by the trajectory of intercepted missiles and drones. Your phone knows exactly what geographic grid you are standing in, and the authorities use that precise location data to instruct you to seek immediate shelter while air defenses engage overhead.

In that context, our phone's unblinking awareness of our physical reality is a life-saving feature. We are deeply grateful that the network knows where we are. But this very same infrastructure, the sensors, the constant tethering to cellular towers, the relentless background tracking, is the exact foundation of digital phenotyping. The sensors that can save your life from a falling drone are the same sensors that can trace the shrinking geographic radius of a depressive episode when you stop leaving your house.

How I Think About Building in This Space

When we build products at CIGMA and when I advise other founders through EthicaLabs, I keep coming back to four principles that digital phenotyping forces us to confront:

Consent must be active and ongoing, not assumed. "Terms of service" consent is not real consent, especially for mental health data. If someone agrees to use a wellness app, they are not necessarily agreeing to have their behavioural patterns analysed by machine learning algorithms and shared with third parties. We need consent mechanisms that are specific, understandable, and genuinely voluntary.

The user must be the primary beneficiary. If a system collects data about someone's mental state, the person whose mental state is being monitored should be the first and primary beneficiary of that insight. Not their employer. Not an advertiser. Not an insurance company. If the business model requires monetising user mental health data, the business model is wrong.

Transparency is not optional. When we build AI that makes inferences about someone's psychological state, we owe them visibility into what was observed, what was inferred, and what the limitations of that inference are. This is not just good ethics. It is the standard that the EU AI Act, the FDA, and emerging state regulations are going to enforce.

First, do no harm is not just a medical principle. Every technology that touches mental health has the potential to cause harm through misinterpretation, privacy violation, discriminatory profiling, or simply by creating anxiety through surveillance. The threshold for deploying these systems should be clinical-grade rigour, not "move fast and break things."

The Builder's Responsibility

I will be honest: there have been moments when the commercial logic of passive monitoring was tempting. Continuous data collection enables better models. Better models enable better predictions. Better predictions enable better products. The technical flywheel is undeniable.

But I have seen enough in healthcare to know that the line between monitoring and surveillance is not drawn by technology. It is drawn by values. And once you cross it, the trust you lose is nearly impossible to rebuild.

The FDA's Digital Health Advisory Committee is currently evaluating AI-based mental health tools. State legislatures are passing AI mental health regulations. The HIPRA Act is working its way through Congress. The governance infrastructure is being built right now, in real time, alongside the technology.

As founders, we have a choice. We can build systems that respect human autonomy and treat mental health data as sacred. Or we can build systems that extract maximum value from the most intimate data people generate. The technology supports either approach. The market will reward either approach, at least in the short term.

But the founders who will build companies that last, companies they are proud of, are the ones who choose the harder path. The path that treats every person's mental health data as if it were their own.

The Ramyfications

Digital phenotyping will transform mental healthcare. I am convinced of that. The evidence is too strong, the potential too significant, the unmet need too vast. Within the next few years, your phone will be able to detect a mood shift before you consciously feel it. The question is whether that capability will serve you or someone else.

I choose to build for the former. And I believe the founders, clinicians, regulators, and users who shape this technology over the next two years will determine whether digital phenotyping becomes the most profoundly beneficial or the most profoundly invasive technology in mental health history.

This is not a choice we can defer. The infrastructure is being built now. The norms are being set now. If you are building in this space, if you are investing in it, if you are using these tools: your voice matters. The conversation is happening. Make sure you are part of it.