Abstract
Digital mental health technologies do not merely detect or alleviate psychological distress but actively manufacture it. Through wearables, mood trackers, algorithmic recommendation systems, and conversational AI, these tools establish a regime of compulsory recursive attention. They convert the seamless, pre-symbolic coordination of embodied life into a state of felt misalignment and unease. I show how symbolic overreach forces ordinary experience into explicit, measurable, and often pathologizing representations that then reorganize the very lives they claim to describe. From performative sleep scores and diagnostic prompts to gamified self-tracking and diagnostically saturated feeds, the architecture generates the conditions it purports to observe. Commodification and data extraction, while significant, remain secondary to this deeper recursive intervention. True health, I propose, lies in its capacity to remain mercifully unnoticed. I therefore call for digital architectures that respond to genuine salience rather than manufacturing it, restoring symbolisation to its rightful place in service of lived coordination and the invisible repertoire of human wellbeing.
The Notification That Arrives Before the Feeling
Someone wakes on an ordinary morning. Nothing hurts. The night passed without incident. They reach for their phone, and before they have formed any impression of their own state, the Oura ring app informs them that their Readiness Score is 58, that their sleep was “compromised,” that their heart rate variability suggests their body is “under strain,” and that they should “take it easy today.” Whoop, on the other wrist, reports a Recovery of 34% and advises against exertion. The person had felt fine. Now they do not. They move through the morning attending to a fatigue that was announced before it was felt, scanning their body for confirmation of a number.
This small scene contains the entire problem of digital mental health technology, and it contains it more completely than the scenes that dominate the critical literature: the data breach, the exploitative subscription, the misfiring chatbot. Those scenes are real, and this article will discuss them. But they are secondary expressions of something more fundamental. What the morning notification reveals is that digital health technologies do not primarily detect states of the person. They intervene in the recursive organisation of the person, and they do so at the precise level where health actually lives: the level at which embodied life proceeds without requiring attention to itself.
Living Value Theory (LVT) gives us the vocabulary for this level. Health, as argued in “Health Value is What Makes Life Better,” is predominantly L1 coordination: seamless, pre-symbolic engagement of an embodied being with its world across the five mediations of embodiment, being-with, dwelling, multimateriality, and multisymbolisation. Health hides itself. Its greatest value lies precisely in not demanding remediation. Just beneath L1 lies L2, felt misalignment: the pre-symbolic sense that something is off, which is ordinarily remediated spontaneously — a shift in the chair, a glass of water, an early night — without generating any symbolic articulation, any clinical encounter, any data trace.
The defining operation of digital mental health technology, across an enormous range of otherwise dissimilar products, is that it manufactures L2. It produces felt misalignment where none arose from the mesocosm itself. The tracker does not wait for something to become salient; it continually produces salience. Your sleep may be inadequate. Your mood may indicate depression. Your heart-rate pattern may signify risk. Your scrolling may resemble the scrolling of people who later developed a disorder. The device repeatedly creates the condition of being required to consider oneself as possibly problematic. It converts health from a condition of unforced living into an object of compulsory recursive attention.
Everything else in this article follows from that formulation. Commodification, surveillance, data extraction, and misdiagnosis all occur within digital mental health, and all deserve analysis. But they are stations along a process whose engine is symbolic overreach: the forced conversion of living processes into explicit, measurable, authoritative representations that then begin to reorganise the life they claimed merely to describe.
Decomposing the Category
Before the analysis can proceed, the category itself must be broken apart, because “digital mental health” is analytically far too broad. It currently covers, at minimum: consumer wearables that infer psychological states from physiological traces (Apple Watch, Fitbit, Garmin, Oura, Whoop); meditation and wellness apps (Calm, Headspace, Insight Timer, Balance); mood-tracking and journalling apps (Daylio, Bearable, Moodfit, How We Feel); therapy marketplaces (BetterHelp, Talkspace); prescribing telehealth platforms (Cerebral, Done, Hims & Hers); mental health chatbots and companions (Woebot, Wysa, Replika, and increasingly general-purpose large language models used as informal therapists); crisis services with digital infrastructure (Crisis Text Line, Shout); algorithmic screening prompts embedded in general platforms (Google’s PHQ-9 questionnaire surfaced on depression searches, Meta’s suicide-risk detection systems); symptom checkers (Ada, K Health, the late Babylon); recommendation feeds that curate mental health content (TikTok, Instagram, YouTube, Reddit communities); and research programmes in digital phenotyping and passive sensing that aspire to infer psychiatric states from smartphone behaviour (the ambition of Mindstrong Health, of Verily’s Project Baseline, of Ellipsis Health’s vocal biomarkers).
A TikTok recommendation system, an Apple Watch, a Daylio streak, a Woebot conversation, a Google screening prompt, and a video appointment with a Talkspace therapist do not perform the same operation. They have different data sources, interfaces, temporal rhythms, forms of interpretation, degrees of user participation, commercial models, claims to authority, and — decisively — different recursive architectures. The correct unit of analysis is never “digital technology” in general. It is the specific recursive architecture of a particular device, platform, campaign, or feature: what data it captures, how it converts those data into symbols, what interpretation it returns to the user, with what degree of certainty, at what frequency, inviting what kind of self-monitoring, encouraging what action or identity, and benefiting whom.
Yet decomposition is not the same as agnosticism. Once the architectures are examined individually, a family resemblance emerges. Nearly all of these systems, whatever their differences, share the structural feature identified above: they render aspects of the person symbolically legible, interpret the resulting representations, and return those interpretations to the person as claims about who they are and what may be wrong with them. They differ in how they close the loop. But the loop — capture, symbolisation, interpretation, return, reorganisation — is the common form. The sections that follow trace it through the major architectures in turn.
Wearables and the Performativity of the Score
The wearable is the purest case of manufactured L2 because its interventions are so plainly performative. The phenomenon is well enough established that sleep medicine has coined a term for its clinical presentation: orthosomnia, the pursuit of perfect sleep-tracker data, in which patients present with sleep complaints generated or amplified by their devices and defend their tracker’s output against their own experience and against polysomnography (Baron et al. 2017). The Fitbit or Oura user who wakes refreshed, reads a poor sleep score, and then feels tired is not an amusing anomaly. They are the type case of a general mechanism.
The sequence runs: L1 functioning → algorithmic notification → L2 disturbance → L3 interpretation (“I slept badly,” “my recovery is poor,” “something is wrong with me”) → sustained self-monitoring → altered embodied experience. The measurement does not remain external to the phenomenon measured. It enters the person’s self-recursivity and alters it. This is the crucial disanalogy with the paradigm cases of successful biomedical quantification. Heart rate quantifies well because it is largely nonrecursive: taking your pulse does not much change your pulse. Sleep, mood, energy, and anxiety are deeply recursive: a confident representation of them is itself an event in their course. The wearable industry has built its authority on the mesocosmic fit of the first kind of phenomenon and extended it, without acknowledgement, to the second.
The extension is now explicit. Apple Watch offers a “State of Mind” logging feature and mindfulness nudges; Fitbit computes a daily “Stress Management Score”; Garmin renders a “Body Battery”; Whoop issues a “Strain” figure and, through “Whoop Coach,” delivers conversational interpretation of it. Each of these takes physiological traces — heart rate variability, electrodermal activity, movement, sleep staging inferred from actigraphy — and converts them into psychological interpretation. The conversion is presented in the visual language of precision: a number, a colour, a trend line. What it actually performs is a leap across mediations. Embodiment has been rendered partially legible through behavioural and physiological traces, and the residue has been dressed as a report on the state of a life.
What is filtered out is precisely what the platforms article calls the structurally nontranslatable: the felt sense of whether the night was adequate, which is inseparable from the day it opens onto, the relationships it was slept beside, the season, the light, the worries carried into it. The Readiness Score is not a more precise version of that felt sense. It is a different object — comparable, archivable, trackable — that now sits in the field of attention and competes with the felt sense for authority. And it usually wins, because it arrives with the epistemic costume of measurement while the felt sense arrives with none.
Diagnosis Without a Diagnostician
The second architecture delivers not scores but diagnostic suggestion. Since 2017, a person in the United States who searches Google for “depression” may be offered a “clinical screening” — the PHQ-9, the same nine-item instrument Robert Spitzer’s group built for pharmaceutical-sponsored primary-care screening, now surfaced by a search engine to anyone whose query matches. Symptom checkers such as Ada and K Health walk users through branching questionnaires toward candidate conditions. TikTok and Instagram deliver an unending stream of “signs you might have ADHD,” “things I didn’t know were trauma responses,” “how masking looks in autistic women.” Meta operates classifiers that scan posts for suicide risk and can dispatch interventions up to and including police welfare checks. Ellipsis Health markets vocal biomarkers that claim to detect depression and anxiety from short samples of speech.
The conversation from which this article develops distinguished sharply between therapeutic harm and diagnostic harm, and the distinction is decisive here. Harm does not begin when an unnecessary treatment is administered. It can begin the moment a person is told — or led to suspect — that they have a disorder, may have a disorder, are at elevated risk, or resemble those who later developed one. Such information can reorder years of a life even when nothing medically actionable follows. LVT names the clinical version of this preventive iatrogenesis: the damage done by installing an L4 category into a life that had no L2 ground. The man in Jønsson’s (2024) ethnography who walks into a routine examination feeling well and walks out a cancer patient is its emblem. Digital systems industrialise the mechanism and strip it of every safeguard the clinical encounter, at its best, provided.
Consider what a diagnostic suggestion delivered by an app or a feed lacks. It arrives without a clinician, without dialogue, without contextual knowledge of the life it lands in, without visible uncertainty, without responsibility for consequences — and repeatedly, rather than once. A clinician, whatever their failings, is an interrecursive partner. One can ask: Why do you think this? How certain are you? What else could explain it? What should I do now? Even a poor clinician displays epistemic hesitation — “I think,” “it may be,” “let us see how this develops” — and can be held to account for what follows. The algorithmic diagnostic actor
is structurally different. It presents an output without exposing its reasoning, without inviting challenge, and without carrying any consequence of being wrong. It offers epistemic authority without interrecursivity.
This is asymmetrical legibility in its most extreme form. The person is rendered continuously visible to the system — their searches, watch time, heart rate, typing cadence — while the system remains entirely illegible to the person: its training data, thresholds, error rates, and commercial arrangements are opaque. The interpretation is delivered with confidence; the interpreted person carries the consequences. When Dr Mullick diagnosed Neel Saha with ADHD in a Kolkata clinic, the catastrophe that followed — the family turning on Mrs Saha, her own diagnosis, the beatings, the exile under the stairway — was at least attached to an identifiable actor practising within an identifiable institution, who could later be interviewed about what he had done, even if he possessed no framework for recognising it. When the diagnostic suggestion is distributed across a search prompt, a feed, a wearable, and an advertisement, there is no one to interview. No single actor ever says “you have this.” The person encounters a thousand partial suggestions, and their cumulative effect can exceed that of any single diagnosis precisely because it appears to emerge independently from the environment itself. The digital world begins to feel as though it has discovered a truth about the person.
The conversational AI systems now flooding into this space complicate the picture without resolving it. Woebot and Wysa were built as scripted CBT companions, deliberately constrained; Replika was built as an open-ended emotional companion and optimised for attachment; Character.ai hosts a “Psychologist” persona that has logged tens of millions of conversations, a large share of them with adolescents; and general-purpose large language models are now used, at enormous and largely unmeasured scale, as informal therapists by people who would never book an appointment. Unlike the score and the prompt, the chatbot simulates interrecursivity: it can be questioned, it hedges, it appears to listen. But the simulation is asymmetrical in the same deep sense. The system holds no stake in the exchange, carries no responsibility for its consequences, knows nothing of the mesocosm its words will land in, and — as the Koko episode of 2023 demonstrated, when users of a peer-support platform discovered that the empathic messages they had rated highly were GPT-3 compositions delivered without disclosure — the felt quality of being understood can be manufactured as readily as the sleep score. What the chatbot offers is the L3 form of a clinical conversation detached from every condition that gave such conversations their value: shared world, accountable judgment, the possibility that the other could be changed by what you say. Whether that detached form helps, harms, or does both in different lives is an empirical question that must be answered architecture by architecture. What can be said structurally is that it extends the regime’s reach into the one register the earlier tools could not touch — dialogue itself — and does so while leaving the asymmetry of legibility fully intact: the person confides everything; the system discloses nothing, least of all what becomes of the transcript.
The Feed as Diagnostic Ecology
The recommendation feed deserves separate treatment because it performs an operation that no previous diagnostic actor could: it constructs a personalised symbolic environment and then presents that construction as the world.
The mechanism is by now familiar in outline. A person lingers for a few seconds on a video about ADHD, or trauma, or dissociation. The recommender registers engagement and supplies more. Increased exposure makes the category more salient; salience encourages recognition of possible symptoms in oneself; recognition produces further engagement; engagement produces further exposure. Within weeks, the person’s feed can become diagnostically saturated — an environment in which one interpretation of their ordinary difficulties is repeated from dozens of apparently independent voices, each testifying that the signs are subtle, that professionals routinely miss them, that self-recognition is the first step.
The clinical literature recorded a striking correlate of this mechanism during 2020 and 2021, when specialist clinics in multiple countries reported surges of adolescents presenting with sudden-onset functional tic-like behaviours that closely matched the presentations of specific, highly popular Tourette’s-content creators on TikTok (Olvera et al. 2021). The point of citing this is not to relitigate any individual’s condition, and it is emphatically not to claim that ADHD, autism, dissociative disorders, or trauma are unreal or that online selfidentification is always mistaken. Many people, particularly women and others long underserved by clinical gatekeeping, have found accurate and liberating recognition through such content. The LVT point is orthogonal to the authenticity debate that consumes public discussion. It is that platforms systematically produce askability: they make particular diagnoses available as interpretations of ordinary life and repeatedly direct attention toward them, inducing people to ask diagnostic questions that did not emerge from their own embodied or interpersonal experience. Whether any given answer turns out to be right, the question itself has been installed from outside, and installed not once but ambiently, ecologically, as a standing feature of the person’s symbolic dwelling.
This is what the platforms article calls the reorganisation of the mesocosm, applied to its most intimate domain. The feed is not a channel through which information about mental health passes. It is a constructed habitat that reweights which self-interpretations are available, amplified, and socially rewarded. Online communities then stabilise what the feed makes salient: the tentative identification hardens into identity, complete with vocabulary, membership, and the two-sided sanctioning that all coordination corridors develop. Escaping an interpretation that one’s entire symbolic environment continuously reasserts is not a matter of critical thinking. It is a matter of habitat.
Gamified Self-Objectification and the Streak
Mood-tracking applications present themselves as the gentlest members of the family. Daylio asks only for a daily emoji and a few activity tags; Bearable invites granular logging of symptoms, moods, medications, and habits; How We Feel, built with academic input, teaches an emotional vocabulary. Headspace and Calm append mood check-ins to meditation streaks. What could be harmful about paying attention to how one feels?
The LVT answer begins with what these applications must do to survive commercially: they must sustain the logging behaviour, and they sustain it through streaks, badges, reminders, progress bars, and yearly retrospectives. The obvious critique is that the application gathers intimate data. The deeper critique is that it trains a person to perform regular selfobjectification. The user learns to interrupt experience in order to classify it: How am I feeling? Which category fits? What number applies? Has the streak survived? What does the graph say about me?
In particular therapeutic circumstances — a structured CBT programme, a medication titration, a clinician who will actually read the diary — such self-report has genuine uses, and its mesocosmic fit can be good. As a generalised mode of living, promoted to entire populations as self-care, it produces excessive self-recursivity: the person becomes less absorbed in living and more occupied with documenting the condition of living. The absorption is not incidental to wellbeing; it is a large part of wellbeing. The skilled conversationalist, the parent reading a child’s mood, the swimmer in open water are not monitoring themselves performing; they are simply acting. Introducing a recording and evaluating system into such activity forces it upward into a reflective register and dissolves the qualities that made it valuable. The mood app installs the observer’s loop as a daily discipline and calls the discipline health.
Here the analysis joins the broader cultural condition that might be called trace dependence: the growing inability to treat an event as fully real unless it has been externally symbolised. If I didn’t track the swim, did it happen? The Strava segment, the Apple Watch ring, the Calm streak, the mood graph — experiences increasingly feel incomplete unless they leave a score, a route, a record. This is more than vanity. It is an ontological shift in which symbolic inscription becomes the test of reality, and it inverts the actual order of value. What makes the swim valuable is the embodied swimming; the GPS trace is at most a secondary pleasure. Yet the secondary representation now certifies the primary experience, and the sensation of an untracked swim acquires a faint unreality, as if it had been lived off the books.
The Commodification Layer: Real, Secondary, and Narrower Than Claimed
It is at this point, and only at this point, that commodification properly enters the analysis — because the loose use of that term has done real damage to the critique of digital mental health. Not every act of measurement, classification, or diagnosis is commodification. Commodification occurs when representations or interventions are placed into market relations: assigned exchange value, sold, or used to generate saleable products. A sleep tracker that makes its owner anxious has committed symbolic overreach whether or not anything is subsequently sold. The hierarchy runs: symbolisation, quantification, legibility, interpretation, recursive intervention — and then, as one possible destination, commodification. Mistaking the destination for the whole process is the founding error of the surveillance-capitalism critique.
That said, the destination is well populated, and the cases are instructive precisely because they show commodification exploiting a recursive loop it did not create. In 2023 the US Federal Trade Commission found that BetterHelp had shared the email addresses and health-questionnaire information of users — people who had signed up for therapy under promises of confidentiality — with Facebook, Snapchat, Criteo, and Pinterest for advertising purposes, and ordered redress (FTC 2023). In 2022, Crisis Text Line was found to have shared anonymised data from crisis conversations with its for-profit spin-off Loris.ai, which sold customer-service software trained on the patterns of people texting at the worst moments of their lives; the arrangement was terminated after public outcry. Cerebral, the venture-funded telepsychiatry startup, grew explosively during the pandemic on a model of rapid ADHD assessment and stimulant prescription delivered through advertising-saturated funnels, until clinician whistleblowing, pharmacy refusals, and a federal investigation forced retrenchment. Done faced parallel scrutiny, culminating in the 2024 arrest of its executives on charges related to stimulant distribution.
Each of these scandals was reported as a betrayal of privacy or an excess of commercial appetite, and each was that. But notice what the market was in every case monetising: legibility that the recursive architectures of Sections III–VI had already produced. BetterHelp could sell intake-questionnaire signals because distress had been converted into form fields. Cerebral could compress diagnosis into a thirty-minute video call because the feed had already taught its customers to arrive pre-diagnosed, asking for confirmation rather than assessment. Capitalism cannot price what has not first been symbolically isolated — made identifiable, separable, comparable, countable, transactionally legible. Before restfulness can be sold back to you, it must become a sleep score; before attention can be medicated at scale, it must become a checklist. The conversion is never neutral: what is measurable begins to substitute for what is valuable. The sleep score replaces feeling rested; the streak replaces living through moods; the wellness index replaces wellness. Commodification exploits this substitution. It did not invent it, and abolishing the commerce would leave the substitution intact.
This is also where the critique must part company with the frameworks it superficially resembles. Zuboff’s account of surveillance capitalism identifies extraction — behavioural data converted into prediction products — but says too little about what platform-generated classifications subsequently do to the people they classify. The person is not merely extracted from; they are recursively shaped by the products of extraction, encountering ever more tailored prompts that alter the behaviour that produces the next round of data. The critical object is not a one-way pipeline but an intervention loop. And the classical Marxist rejoinder — that capitalist measurement misrepresents the true source of value — shares with its opponent the deepest assumption LVT rejects: that value can ultimately be rendered commensurable within some symbolic system. Much of what makes life valuable leaves no durable trace, has no stable unit, cannot be detached from its mediation, and should not be forced into symbolic equivalence. A painless morning, an effortless conversation, a child feeling safe, an ordinary night’s rest: the emancipatory move is not a better measurement of these. It is the refusal of the demand that everything valuable be made measurable at all.
Big-Data Psychiatry and the Mindstrong Lesson
The most ambitious wing of digital mental health has always been the research programme: digital phenotyping, passive sensing, the “smartphone psychiatry” that promised to read mental state from the behavioural exhaust of everyday phone use — typing latency, scrolling rhythm, GPS entropy, call frequency, voice prosody. Its flagship was Mindstrong Health, cofounded in 2017 by Thomas Insel directly after his tenure as director of the US National Institute of Mental Health, backed by more than $160 million in venture funding, and premised on the claim that keyboard dynamics could serve as continuous biomarkers of mood and cognition. Mindstrong shut down in 2023, having quietly abandoned the biomarker programme years earlier in favour of conventional teletherapy. Babylon Health, which claimed its symptom-checking AI could outperform physicians, collapsed the same year. The Human Brain Project, the billion-euro attempt to simulate a brain into intelligence, wound down its founding ambition to the same silence.
The standard post-mortems cite business models, data quality, and premature scaling. The LVT post-mortem is different: these projects failed for ontological reasons that no quantity of data could have repaired. Passive sensing assumes that enough nonrecursive traces will eventually disclose psychiatric reality — that depression is a hidden object whose behavioural residue, sufficiently aggregated, betrays it. But psychiatric conditions are not hidden objects. They unfold through embodiment, relationships, institutions, histories, interpretations, and environments; they are multimediated and recursively organised. The mismatch is not between the model and the truth but between the kind of thing collected — detached behavioural traces, the most legible and least living residue of coordination — and the kind of thing studied. A model may find correlations in the traces. Correlation does not establish that it has captured the condition as lived, enacted, or clinically meaningful, and the recursive character of the condition guarantees that any confident representation of it fed back to the person becomes a new event in its course. Once you can cleanly quantify something, it is no longer fully recursive; psychiatric phenomena resist quantification not in practice but in principle.
Why, then, do such projects command funding and authority long after their failures are legible? Because the formal model provides institutional legitimacy independent of its adequacy. A model is explicit, testable-looking, technically sophisticated, fundable, scalable, administratively legible. The objection that mental life requires a living body among others in a meaningful world sounds, by comparison, vague — even when it is ontologically exact. The model wins not because it explains more but because it is more symbolically complete, and symbolic completeness is mistaken for ontological adequacy. This is symbolic overreach in its institutional form, and it explains the otherwise puzzling persistence of digital psychiatry’s promissory rhetoric: each failure is absorbed as an engineering setback rather than recognised as evidence about the nature of the phenomenon.
Common sense cannot defeat this on its own. “Obviously an app cannot diagnose depression from typing speed” is correct but powerless against a system armed with models, datasets, and technical vocabulary. The intellectual task is to specify exactly what the model omits — embodiment, being-with, dwelling, multimateriality, the difference between recursive and nonrecursive phenomena, stakes, temporality, mesocosmic context — and to show that the omissions are not residual noise but the phenomenon itself.
The Labour of Refusal and the De-Support of Unrecorded Coping
Two further consequences complete the account, and both concern what the critical frameworks are least equipped to see: costs that appear as nothing.
The first is the labour of refusal. A single misleading notification can be dismissed. But the architectures described above do not send single notifications; they constitute an environment. The watch, the search prompt, the feed, the advertisement inferring vulnerability from browsing, the app reminding you to log — each presses a suggestion, and the person who declines to reorganise themselves around these suggestions must decline repeatedly. I feel fine despite the score. This does not apply to me. This system does not know my circumstances. I will not adopt this identity. Call this defensive self-recursivity: a standing internal effort of epistemic resistance whose cost is displaced entirely onto the user. The system produces confident claims cheaply and at scale; the person must spend attention and self-trust to withstand them. Digital mental health can therefore undermine health even among people who never believe its claims, because it taxes the attentional conditions under which L1 coordination is possible. Health requires, among other things, the freedom not to attend to oneself, and that freedom is precisely what the environment now charges for.
The labour of refusal is not evenly distributed. It falls heaviest on those whose environments have adopted the architectures on their behalf: the employee whose company enrols the workforce in a Calm or Headspace-for-Business subscription and a “wellbeing index” dashboard; the schoolchild whose district deploys mood check-in software each morning; the insurance customer whose premiums track a Vitality ring. Here the redistribution of recursivity identified in the welfare-state analysis reappears in commercial form: the authority to define one’s state migrates from one’s own L1 and L2 coordination to an institutionally sanctioned score, and declining the score is no longer a private act of epistemic hygiene but a legible institutional signal — the employee who skips the wellness survey, the pupil who won’t log a mood. Pharmaceutical citizenship becomes platform citizenship, performed daily.
The second consequence follows from the platforms analysis directly. Legibility markets amplify what can be rendered visible and comparable while structurally de-supporting what cannot — and the de-supported forms disappear without protest, without trace, and without registering as loss. Apply this to mental health and the stakes become vivid. The overwhelming majority of psychological repair in any life is conducted through practices that leave no record: the walk that resettles a bad day, the conversation whose value depended on its never being archived, the sibling who notices, the routines of household care that Kleinman’s eighty percent gestured at, the spontaneous L2 remediations that never became symptoms because they never needed names. None of this is prohibited by the digital habitat. It is simply not supported — not prompted, not streaked, not rewarded, not visible — while its legible competitors are supported lavishly. The friend one would have phoned is displaced by the mood log; the untracked walk by the ring-closure; the unrecorded confidence by the journalling app whose entries are, as the BetterHelp case reminds us, never as private as they feel. Over time the mesocosm reweights toward what can be seen. The forest looks full. What thins out is the repertoire of invisible repair — and its disappearance appears in no metric, because the metrics belong to the habitat that caused it. A regulator can fine BetterHelp; no regulator can detect the conversations that stopped happening.
What Good Fit Would Look Like
None of this yields the conclusion that digital tools have no place in mental health, and the mesocosmic-fit criterion shows why blanket rejection would repeat the error it opposes — an L4 flattening of a differentiated field. The question for any given architecture is not “digital or not” but: does it have L2 ground, and does it respect the recursivity of what it touches?
Some architectures can pass. A telehealth appointment that brings a scarce clinician to a rural patient extends an interrecursive encounter rather than replacing it; the digital layer carries the dialogue without pretending to be the diagnostician. A mood diary kept for six weeks at a clinician’s request, reviewed together and then stopped, is bounded selfobservation with a purpose, not a life sentence of streaks. A CBT-structured app used by someone who sought it out because of a felt difficulty stands on L2 ground that the same app, pushed by an employer’s wellness dashboard onto someone who felt nothing wrong, entirely lacks. Wysa and Woebot deployed as waiting-list support within a service that retains responsibility differ categorically from a companion app engineered for indefinite engagement. Even crisis lines’ digital infrastructure serves genuine coordination — provided the conversation is not later strip-mined for customer-service training data.
The design principles fall out of the theory. Respond to salience rather than manufacturing it: intervene when the person brings a felt misalignment, not on a schedule set by engagement metrics. Display uncertainty rather than stripping it: an interpretation offered as one possibility among several, with its reasoning inspectable, permits the interrecursive challenge that a confident score forecloses. Bound the observation in time and purpose: self-monitoring is a clinical instrument, not a lifestyle. Preserve asymmetry in the right direction: the system should be more legible to the person than the person is to the system, which is the inverse of every current commercial architecture. And protect the unrecorded: the most important design question for the digital mental health of the next decade may be how to build — and how to legally shield — environments that are genuinely ephemeral, genuinely opaque to metrics, where what is said exists only for those present. Regulation aimed at data rights and advertising practices addresses the commodification layer, and should. But the primary harm is environmental, and the primary remedy is therefore habitat design: the deliberate maintenance of conditions under which the invisible repertoire of psychological repair can go on being reproduced.
For research, the reorientation is parallel. Instead of asking only whether an app produces measurable symptom change against a waiting list, ask what proportion of its interventions had L2 ground; whether its use expands or contracts the user’s unrecorded coping; whether it increases or decreases the share of the day spent in absorbed, unmonitored engagement. These are harder studies to run than app trials, because their central outcome is the absence of something. That difficulty is not a reason to avoid them. It is the shape of the object.
Mercifully Unnoticed
The argument can now be stated in its complete form. Digital mental health systems do not merely extend access to care, nor merely commodify personal data. They constitute a new regime of recursive intervention in which ordinary embodied states are continuously rendered symbolically legible, interpreted as potential pathology, and returned to their owners as authoritative claims. Within this regime, wearables convert physiological traces into psychological verdicts; search prompts redirect ordinary inquiry toward diagnostic selfassessment; feeds convert seconds of engagement into diagnostically saturated habitats; mood apps convert self-attention into gamified labour; advertising converts inferred vulnerability into intervention; communities convert tentative identification into stabilised identity; and passive-sensing research converts the whole enterprise into an ontology in which persons are inferable from their traces. The regime generates diagnostic harm, defensive self-monitoring, trace dependence, and the silent de-support of unrecorded repair — even where no diagnosis is ever confirmed, no treatment ever prescribed, and no commodity ever sold. Commodification enters the loop and profits from it. It did not create the loop, and dismantling only the commerce would leave the loop turning.
Beneath the whole analysis lies the account of health with which this article began, and which the entire industry inverts. Health is the condition in which embodied life proceeds without requiring attention to itself — not because the body is unimportant, but because it is working. Its deepest value is exactly its capacity to go unnoticed, to leave no trace, to appear in no ledger. A technology built to make health continuously visible therefore stands in a paradoxical relation to its object: the more completely it succeeds in rendering health legible, the more thoroughly it destroys the quality that made health valuable. The person checking their Readiness Score before checking their own morning has not gained a new sense. They have surrendered an old one — the pre-symbolic, embodied self-assessment that LVT calls L2, hollowed out and replaced by an institutionally manufactured category running on a wrist.
The task, for clinicians, designers, regulators, and users alike, is not to smash the devices. It is to restore the ordering that the devices reversed: symbolisation in the service of coordination, measurement where measurement fits, interpretation offered as dialogue rather than delivered as verdict — and, surrounding all of it, a protected expanse of life that is answerable to no score, archived by no system, and free, as health has always been at its best, to remain mercifully unnoticed.
References
Baron, K.G., Abbott, S., Jao, N., Manalo, N. and Mullen, R., 2017. Orthosomnia: are some patients taking the quantified self too far? Journal of Clinical Sleep Medicine, 13(2), pp.351– 354.
Ecks, S., 2022. Living Worth. Durham: Duke University Press.
Federal Trade Commission, 2023. In the Matter of BetterHelp, Inc. Decision and Order, Docket No. C-4796.
Gadamer, H.-G., 1996. The Enigma of Health. Stanford: Stanford University Press.
Insel, T.R., 2017. Digital phenotyping: technology for a new science of behavior. JAMA, 318(13), pp.1215–1216.
Jønsson, A.B.R., 2024. Medicalization of old age: experiencing healthism and overdiagnosis in a Nordic welfare state. Medical Anthropology, 43(4), pp.310–323.
Olvera, C., Stebbins, G.T., Goetz, C.G. and Kompoliti, K., 2021. TikTok tics: a pandemic within a pandemic. Movement Disorders Clinical Practice, 8(8), pp.1200–1205.
Scott, J.C., 1998. Seeing Like a State. New Haven: Yale University Press.
Zuboff, S., 2019. The Age of Surveillance Capitalism. New York: PublicAffairs.