Skip to content
Independent researchSIGNAL & SELF

Hypotheses

Several stories can be true at once

Observed increases in psychiatric identification online have many candidate explanations. The scientific job is to say what each predicts, what would strain it, and how it sits with the others—not to pick a team.

Moderate evidence

Awareness hypothesis

People with previously unrecognized disorders can finally identify their symptoms because information is more available.

Predicts

  • Self-identification should be higher where historical under-recognition is documented
  • Clinical confirmation rates should be relatively high for some conditions (e.g., autism in adults)
  • Help-seeking should increase, not only self-labelling

Would be strained by

  • Self-labels systematically fail structured assessment even in previously under-recognized groups
  • Identification rises for conditions without evidence of prior under-detection while staying flat for historically missed groups

Strongest where services historically missed people. Weakest as a universal explanation for every online diagnostic trend.

Compatible with: access, destigmatization, selection. Supporting: Lewis, 2016; Gulliver, 2010; Naslund, 2020; Pretorius, 2019.

Moderate evidence

Access hypothesis

People self-diagnose because professional assessment is inaccessible, delayed, expensive, or culturally unsafe.

Predicts

  • Self-diagnosis higher where wait times and cost are higher
  • Formal diagnosis follows when access improves
  • Self-diagnosis functions as a stopgap, not an identity endpoint

Would be strained by

  • Self-diagnosis rates remain high in well-resourced systems with short waits
  • People with access still prefer community labels to assessment

Explains self-diagnosis as rational under constraint. Does not explain functional symptom waves.

Compatible with: awareness, destigmatization. Supporting: Gulliver, 2010; Pretorius, 2019; Lewis, 2016.

Moderate evidence

Destigmatization hypothesis

People are more willing to discuss conditions that previously remained hidden; apparent rises partly reflect disclosure.

Predicts

  • Public stigma declines as identification rises
  • Rises appear first in self-report, later in service data
  • Effects should span many conditions rather than a few media-salient ones

Would be strained by

  • Stigma remains stable or worsens while identification of specific media-linked presentations spikes
  • Condition-specific spikes track viral content more than general destigmatization

Stigma change is real but uneven across diagnoses. It is a poor explanation for FTLB phenomenology.

Compatible with: awareness, broader-distress. Supporting: Pescosolido, 2021; Naslund, 2016; Gulliver, 2010.

Moderate evidence

Selection hypothesis

People already experiencing symptoms seek related content and communities; algorithms then look like causes because they follow demand.

Predicts

  • Content exposure follows symptom onset more than it precedes it
  • Within-person analyses show reverse or bidirectional paths
  • Recommended content mirrors prior search and dwell

Would be strained by

  • Randomized or quasi-random exposure to diagnostic content raises new self-diagnosis in previously asymptomatic people

The default alternative to simple causation. Must be tested, not assumed.

Compatible with: algorithmic, awareness, broader-distress. Supporting: Heffer, 2019; Coyne, 2020; Odgers, 2020.

Mechanistically plausible

Algorithmic amplification hypothesis

Recommendation systems progressively increase exposure to disorder-related material beyond what users would choose unassisted.

Predicts

  • Recommended mental-health content exceeds searched content over a session
  • Repetition predicts outcomes better than total screen time
  • Similar users diverge when ranking objectives change

Would be strained by

  • Logged recommendation traces show little concentration of diagnostic content
  • Time-use predicts outcomes as well as recommended-topic exposure

Mechanically plausible and under-measured. Platform data needed. Must not leap from ranking mechanics to psychiatric causation.

Compatible with: identity, reinforcement, social-learning. Supporting: Narayanan, 2023; Cinelli, 2021; White, 2009; Yeung, 2022.

Emerging

Identity hypothesis

Diagnostic labels can become integrated into adolescent identity, especially where they confer meaning, community, and a story of the self.

Predicts

  • Label use becomes more central to self-description over time
  • Effects concentrate in adolescence
  • Community language migrates into offline self-concept

Would be strained by

  • Labels remain instrumental and are readily dropped after assessment or treatment
  • No developmental window

Illness identity is not inherently pathological. The question is when identity organizes around impairment versus recovery and support.

Compatible with: reinforcement, awareness, social-learning. Supporting: Yanos, 2010; Blakemore, 2014; Cruwys, 2014; Lewis, 2016.

Mechanistically plausible

Reinforcement hypothesis

Social rewards—likes, comments, belonging, status—reinforce particular forms of presentation or signalling.

Predicts

  • Posts that display symptoms or identity markers receive more engagement in some communities
  • Subsequent posting tracks prior reinforcement
  • Expression changes more readily than underlying severity

Would be strained by

  • Engagement does not predict later posting or expression after baseline
  • Communities reward recovery narratives at least as much as symptom display

Keep posting, interpretation, expression, and severity separate. Most evidence, where it exists, is about posting.

Compatible with: identity, algorithmic, social-learning. Supporting: Nesi, 2018; Marchant, 2017; Cruwys, 2014.

Emerging

Social-learning hypothesis

People can learn symptom vocabulary, interpretive frames, or behaviour through observation of models.

Predicts

  • Phenomenology of new presentations resembles high-reach models
  • Onset follows exposure rather than only following stress
  • Symptoms can be condition-specific rather than generic distress

Would be strained by

  • New presentations do not resemble available models
  • Exposure is unrelated to phenomenology after stress is controlled

Best evidenced, still imperfectly, for functional tic-like behaviours and some self-harm content. Not a licence to treat all diagnoses as learned.

Compatible with: functional, algorithmic, reinforcement. Supporting: Müller-Vahl, 2022; Olvera, 2021; Hull, 2021; Arendt, 2019; Hatfield, 1993.

Clinical observation

Functional symptom hypothesis

Some susceptible individuals may develop genuine functional symptoms influenced by social context and available symptom models.

Predicts

  • Rule-in functional features on examination
  • Demographic and phenomenological mismatch with the imitated disorder
  • Symptoms can improve with explanation and reduced reinforcement

Would be strained by

  • Cases meet classic neurodevelopmental criteria and course
  • No excess of functional signs relative to historical baselines

Functional ≠ feigned. This hypothesis is condition-bound until shown otherwise.

Compatible with: social-learning, pandemic, broader-distress. Supporting: Espay, 2018; Pringsheim, 2021; Müller-Vahl, 2022; Bartholomew, 2002.

Moderate evidence

Broader distress hypothesis

Changes reflect increasing psychological distress—economic, educational, ecological, social—rather than social-media effects specifically.

Predicts

  • Parallel rises in distress among light users
  • Macro indicators (loneliness, sleep, economic insecurity) track symptoms
  • Platform-specific phenomena are a small slice of the overall rise

Would be strained by

  • Symptom rises concentrate in heavy, recommended diagnostic-content users after confounders
  • Condition-specific functional waves have no analogue in general distress measures

Twenge treats the rise as real and attributes it partly to phones; critics treat the rise as real or artefactual and dispute the cause. Distress and media are not mutually exclusive.

Compatible with: pandemic, destigmatization, awareness. Supporting: Racine, 2021; Odgers, 2020; Twenge, 2018.

Strong evidence

Pandemic hypothesis

COVID-era disruption independently changed adolescent mental health through isolation, uncertainty, bereavement, and service collapse.

Predicts

  • Discontinuities at 2020
  • Partial reversal as schools reopened—unless scarring occurred
  • Similar trends in low-social-media contexts that shared lockdowns

Would be strained by

  • Phenomena continue to accelerate after disruption recedes, tracking platforms more than disruption

The pandemic is both a confounder and a natural experiment. It cannot be ignored, and it cannot automatically explain every subsequent pattern.

Compatible with: broader-distress, functional, algorithmic. Supporting: Racine, 2021; Orben, 2020; Heyman, 2021.