This piece is part of our series “Reimagining the Future of Digital Health.” Read more here.

By Mira D. Vale

May 20, 2026

As AI tools in mental healthcare proliferate — chatbots for therapeutic dialogue, apps to track behavior, algorithms to speed diagnosis and predict future crises — ethical dilemmas have come to the fore. Can people meaningfully consent to continuous, passive monitoring? What does privacy mean when seemingly banal smartphone metadata might forecast a relapse? Who is responsible when something goes wrong? The people developing AI mental health tools are a diverse bunch, including developers, clinician-scientists, entrepreneurs, and academics. For them, tackling these questions requires ethical as well as technological innovation. 

Studying this world over the past six years, I have seen that work take various forms. Mental health AI innovators assemble multidisciplinary teams of experts, bringing together ethicists, philosophers, coders, and clinicians to address concerns beyond any one group’s domain. They commission studies of new strategies for informed consent and data privacy. They iterate safety features both proactively and in response to controversies. And they sound the alarm about the field’s lack of regulation, decry actors they perceive as less-than-scrupulous, and press, repeatedly, for serious oversight. 

But alongside these efforts, there is a second-order process underway that gets less attention. As these groups work to resolve the ethical dilemmas of AI mental healthcare, they are also building what I call the field’s moral infrastructure: the norms, standards, and assumptions that will govern what AI mental healthcare is, what counts as adequate evidence, who has the authority to evaluate it, and what kinds of problems the field is responsible for. If consent protocols, safety features, white papers, and oversight bodies are the visible architecture of ethical solutions, moral infrastructure is what holds everything up and makes it work. It is the foundation poured first, the wiring routed through the walls.

Consider one norm that has been solidifying in academic digital psychiatry: the idea that passively sensed behavioral data — output derived from digital sensors that describe people’s movements and sentiments — is more reliable than what patients say about themselves. In my research, I encountered this preference repeatedly, expressed confidently. One psychologist explained to me that she prefers not to take patients’ self-reports at their word, because “people are very unreliable.” If the goal is to understand mental distress, she argues, individual experience without sensor validation is too variable, too messy, too subjective.

That argument has intuitive appeal for researchers committed to generalizable insight. But it is also a moral choice dressed as a methodological one. Deciding that sensed data should take precedence over a patient’s self-report is a decision about whose account of mental distress is authoritative. It is a very different paradigm for mental healthcare than classic therapeutic modalities, which focus overwhelmingly on patients’ experiences and feelings. It has consequences for what gets treated, by whom, and how.

A version of the same process appears to be unfolding beyond the world of purpose-built mental health tools, in general-purpose AI systems. Anthropic reports that 2.9 percent of user interactions with its chatbot Claude are “affective” conversations, “personal exchanges motivated by emotional or psychological needs.” OpenAI reports about the same for ChatGPT.  At the scale of their use, this amounts to hundreds of millions of conversations per week. Analyzing use patterns, companies are starting to make decisions about how their AI assistants should respond to a user who expresses thoughts of self-harm: what language to use, whether to promote a crisis line, when to disengage. These decisions encode answers to questions that mental health ethics has spent decades debating: What is a non-clinical provider’s duty of care? When is referral adequate? What does it mean to take someone’s distress seriously?

What matters is not just what the answers are, but how they’re arrived at. These decisions are made by product and safety teams, sometimes in consultation with clinicians, and rarely through anything resembling a public process. If they converge across companies — and in all likelihood, they will — they will constitute a de facto standard for how AI handles mental health crises.

This second-order process is happening now, largely outside formal regulatory routes, and mostly without deliberate public input. Federal regulation has focused on removing barriers to AI’s development, and state laws are an uneven, haphazard patchwork. Existing laws that might otherwise apply, like HIPAA, were designed before the current generation of technologies existed, and they cover only a fraction of the tools and contexts now in use. Consumer-facing mental health apps, for instance, largely fall outside HIPAA’s definition of a covered entity. In the wide space this leaves, the people building AI — both the tools designed for mental healthcare and the tools used for mental healthcare regardless of design — are settling not just the field’s ethical norms. They are also setting the scope of what AI ethics in mental healthcare is taken to be.

This is how moral infrastructure gets built: through the accumulation of choices made by people doing their jobs, trying to build things that work, and presenting those choices to funders, regulators, colleagues, and the public. The choices may be reasonable. But they are choices nonetheless, about what knowledge counts, who bears responsibility, and what kinds of harm are foreseeable and therefore preventable.

Though mental health AI often gets described as a “Wild West,” the field is already converting decisions into defaults and defaults into common sense. The questions now being resolved will not wait for regulators. By the time formal oversight arrives, this infrastructure will be a solid foundation, and the practical cost of disturbing it will be high. In other words, the moral infrastructure established for mental health AI today forms the scaffolding for whatever safeguards may exist in the future. 

Most importantly, as social scientists who study technology have long observed, infrastructure becomes hard to see once it is installed. The standards and assumptions being established in AI mental healthcare now will not always look like decisions. They will just be how things are done. That is why the moment of construction matters, as these decisions are made without being named, in ways that cannot be revisited. The wiring is going in now. This is the moment to see what’s being routed where, before it disappears behind drywall.

Mira D. Vale is an assistant professor of sociology at Washington University in St. Louis. She examines how emerging technologies shape the practice and provision of health care, drawing on approaches from economic sociology, medical sociology, and science and technology studies. Her current research focuses on the ethics of digital mental health as AI and machine learning tools are adapted for clinical use.