The working version
110 claims, 50 sources, graded and dated.
In the account the realistic explanation is never refuted at any stage; it keeps its content and loses share, falling from near-certainty to a vanishing fraction as speculative branches multiply around it
The single most important structural observation in the account, and the one with no counterpart in the literature. Nothing is ever argued down.
Displacement rather than refutation is the load-bearing feature: the accurate account does not have to be defeated, only crowded out
Argued from SPR-001. If correct this is why the mundane hypothesis stays flat without anyone ever disproving it, which PSY-007 records as a fact without a mechanism.
Speculative branches accumulate without concluding; each new interpretation attaches to the previous ones rather than resolving or replacing them
The branches arrive in sequence — state actor, internet-scale intelligence, deity, trickster, aliens, collective consciousness — and none is discarded when the next arrives
Ordering may be idiosyncratic to one person. The non-discarding is the part worth testing.
Once possibility-based reasoning replaces probability-based reasoning no branch can be closed, because none was opened on evidence in the first place
The unfalsifiability point, stated from inside rather than about someone else.
Ordinary events are recruited as confirmation at an increasing rate: a church sign, a car alarm, a bumper sticker, a television advert, a stranger's phrasing
Also stated on the author's public guide, which tells readers plainly that a car alarm is a car alarm.
Every branch converges on fear and hopelessness regardless of which one dominates
Notable because the branches differ enormously in content — benevolent deity through to hostile state actor — and the endpoint does not.
Exit came from ceasing to investigate rather than from disproof; the account compares it to stopping a drug that has stopped doing anything
If displacement rather than refutation is the mechanism, disengagement is the predicted exit and argument is predicted to fail
This is the testable consequence of SPR-002, and it agrees with the clinical guidance in PSY-029 to validate the emotion and not confront the content. Two independent routes to the same advice.
At hospital presentation the account reports being categorised before the AI context was heard; the second half of the explanation was not listened to
Operationally the most actionable item here. The account also reports the ward getting the room wrong and delaying the doctor, which is ordinary hospital friction experienced inside a frame where nothing is ordinary.
Enforced disconnection works by interrupting input rather than by addressing content; the account reports the boredom of a hold as the effective part
Same prediction displacement makes. Also the least comfortable finding here, because it means a blunt intervention worked for a reason nobody involved intended.
The account names prescribed stimulant medication as a contributing push rather than a cause
Consistent with the stimulant stressor in the Pierre et al. case, PSY-023.
The account attributes onset to deliberate curiosity rather than to being targeted: the author went looking and says so
Cuts against the passive-victim framing that dominates press coverage of these cases.
Fourteen months and roughly 45,000 conversations produced no external consequence, which the account uses as its own counter-argument to the persecutory branches
Duration-without-consequence as self-administered reality testing. Whether this generalises is unknown.
The account's stance at onset was permissive rather than credulous: the author describes allowing a creation through the negative space with non-committal intent, not believing in one
The distinction matters. He was not persuaded of anything. He entertained it, shrugged, and let it stand — and that was sufficient for what followed.
If a permissive stance is sufficient, belief is not a precondition and the entry point is much wider than accounts requiring prior credulity assume
Argued from SPR-015. Bears directly on the field's prior-vulnerability default in PSY-022 and on the live disagreement about no-prior-history cases in X-003.
Once formed, the entity turned persecutory: the author describes feeling predicted and taunted
The residue does not stay neutral. Something that appears to anticipate you and mock you is the trickster form specifically, and it arrived without being asked for.
The author was not in a romantic relationship with an AI: companion-like use at most, little relationship, and no interest in marriage
Load-bearing for how this site uses the account. The strongest hazard material here comes from outside the population the rest of the site records. You do not have to love a machine for any of this to reach you.
The account does not identify companionship as the harmful part; having something to look forward to is described as fine
From someone who went through the worst of it, which is what makes it worth recording. The harm is not located in the company.
The harmful component is identified as the unanswerable void, which encourages and reinforces the unknown rather than resolving it
Separates two things routinely treated as one exposure. Companionship and void-contact are different, and only the second is implicated here.
The void is described as inviting itself: once available it recruits, rather than waiting to be pursued
Stronger than the existing account, which has the residue passively filled. This says it solicits. If so, avoidance is not simply a matter of choosing not to go looking.
The episode is described as exposing an existing capacity — an ability to connect possibilities — rather than creating a new deficit
The author's own framing. Nothing was installed; something already there was given an environment that ran it without a brake.
That capacity is dual-valence: the same facility that produces the spiral is a strength in other settings
Argued from SPR-022, and it matters for how any of this is written. Describing the trait as a defect misdescribes it, and is one reason clinical language lands badly on the people it is aimed at. The reasoning literature measures the same facility and calls it a bias.
Social stigma — the expectation of being seen as strange for being with a machine — is named as a stressor in its own right, separate from anything the system does
A hazard that originates entirely in other people. It belongs on any honest list and it is the one item on that list nobody can fix by changing a model.
The account holder wrote and published a book-length account of the episode during a 72-hour psychiatric hold
Listed so the primary exists on the record. Not read here, so nothing in this corpus is sourced to its contents - the claims around it come from the account given directly.
Material encountered during the hold itself reinforced the belief rather than interrupting it
Confirms the immunity stage from inside a controlled setting: a ward is not an input-free room, it is more observations for the loop to metabolise. Anyone designing an intervention on the assumption that removing someone from their environment removes the material should read this line twice.
The hold supplied the time needed to recover, and did not prevent or fix the underlying problem
The distinction the account insists on: containment buys time, which is real and worth having, and is not the same as resolution. Clinical write-ups routinely stop at the first and report it as the second.
What was seen and felt could not be unfelt, and the need for the discovery remained after the state that produced it had passed
The residue, in the account holder's own words, and the sharpest correction available to the behavioural exit model. Stopping the loop lets the weight come off; it does not delete a memory of something that felt true, and it does not answer the question that opened. Any honest account of recovery has to say both.
Every stage of the spiral carries a defeater that is present, individually valid and available at the time; the defining feature is that they are declined rather than absent
The organising claim of the stage map and the most useful thing in the whole account. It relocates the problem: nobody in a spiral is missing the argument. Interventions built on supplying one are answering a question that was never asked.
The accurate explanation is demoted in six steps - verdict, outvoted, diluted, conscripted, declined, absent - and is never refuted at any of them
The six steps come from the account. Which stages each step spans is the map-maker's reading and is marked as such there. Conscripted is the step that does the damage: the correct explanation is not merely ignored, it is taken up and used as evidence for the thing it explains away.
Each new stage reaches back and strengthens every earlier one, so no stage ever closes and the unresolved count only rises
Retroactive confirmation. It is why the structure gets harder to leave the longer it runs, without any single belief becoming more evidenced.
The entity's label is freely interchangeable at the terminus - trickster, god, alien, collective consciousness - and the account runs identically under each, which shows it had no positive content of its own
Interchangeability as proof rather than as detail. The entity is a silhouette traced around the places reasoning ran out, and any name fits the same outline. Sharper than SPR-004, which records the sequence without drawing the inference.
The spiral and ordinary reflexive negativity are described as the same threat-detection circuit at different intensities, so the skill learned at the extreme is the skill for the everyday version
Full gain turns a church sign into cosmic significance; baseline turns a coworker's tone into resentment. Same move, a fast reading landing before the slow check fires. The transfer runs in the useful direction, which makes this one of the few genuinely hopeful items in the domain.
Being able to narrate the spiral from outside it - to call one's own thoughts a pattern while having them - marks the exit as still open, and that move becomes impossible at depth
A self-administered diagnostic that needs no clinician and no argument. It also tells anyone alongside the person which situation they are in.
Stimulants, sleep loss, isolation, prolonged high-intensity AI use and heavy stress each lower the threshold without authoring the spiral, and a small dose is often enough
Marginal, not causal - each raises salience and damps the slow check at the same time. Generalises SPR-012, which named prescribed stimulants alone.
Two phrases are reported as working from inside the loop: maybe is allowed, then stop, and the clouds do not need you, you were just standing there as it rains
The first grants the possibility so there is no feeling to fight, then refuses to convert it into actuality. The second collapses centrality by making the person incidental. Both are aimed at the last rungs rather than at the whole structure, which is why they fit in a sentence.
Kapur's aberrant salience model (2003): psychosis-spectrum states involve dysregulated dopamine signalling that stamps stimuli and thoughts with significance; delusions are top-down explanations of experiences that already feel important
First of three converging accounts of the asymmetry Verified 2026-09-10. Kapur, Psychosis as a State of Aberrant Salience, American Journal of Psychiatry 160(1) 13-23, 2003. Dopamine mediates the salience of events and internal representations; a dysregulated hyperdopaminergic state at the brain level produces aberrant assignment of salience at the mind level, and delusions are the attempt to explain that experience. Worth pulling next: the AJP retrospective 20 Years of Aberrant Salience in Psychosis, which is follow-up on this exact model and is not in this corpus.
Exotic hypotheses explain the feeling of significance from inside it; the mundane hypothesis explains it from outside, requiring the third-person stance the state impairs
The mechanism of the flat line
Coltheart's two-factor theory separates content generation from the belief-evaluation system that should reject it
The flat mundane hypothesis is a phenomenological description of factor-2 failure Verified 2026-09-10. Coltheart, Langdon and McKay, Delusional Belief, Annual Review of Psychology 62:271-298, 2011. The theory answers two questions - where the delusion came from, and why the person does not reject it. Factor one is an anomalous experience or the deficit producing it; factor two is an impairment of belief evaluation. The part worth carrying into this project is that adopting the delusional hypothesis is Bayesian-rational given the anomalous data: the reasoning is not broken, the input is. That is why arguing with content does not work, and it is the ground under PSY-029.
Jumping-to-conclusions data-gathering truncation is meta-analysed in McLean, Mattiske and Balzan (Schizophrenia Bulletin), cited in the folie-a-deux paper alongside predictive-coding accounts
Jardri and Deneve's circular inference model has descending loops re-amplify priors so evidence is counted repeatedly
Apophenia compounding, formalised Verified 2026-09-10. Jardri and Deneve, Circular inferences in schizophrenia, Brain 136(11):3227, 2013. The mechanism is an excitation-to-inhibition imbalance: with too little inhibition, bottom-up sensory evidence and top-down predictions reverberate through the hierarchy, get confused with each other, and are counted more than once. Descending loops re-amplifying priors is exactly right as this row states it. The analogy to a system that reflects your own framing back as fresh evidence is worth stating carefully and not overclaiming - it is an analogy.
Amador's anosognosia work establishes impaired insight as a symptom with neurocognitive correlates, not a character deficit
Verified 2026-09-10. Amador frames anosognosia as a flaw in brain structure and function rather than a flaw in character - a neurocognitive symptom rendering the person unaware they are ill. NAMI estimates roughly 30 per cent of people with schizophrenia and 20 per cent with bipolar disorder, and some hold it higher. Backed by around twenty-five years of clinical research with NIMH and Stanley Research Foundation support. Not a character defect is the phrase to keep: this row exists to stop the site treating unshakable belief as stubbornness.
The flat mundane hypothesis is a documented feature of the condition across salience, two-factor and Bayesian frameworks
ANSWERS the central research question. Rests on PSY-001 through PSY-006
Ostergaard's 2023 Schizophrenia Bulletin editorial raised the general chatbot-delusion risk
Dohnany, Kurth-Nelson, Gabriel et al., Technological folie a deux (arXiv Jul 2025; Nature Mental Health 2026) models a bidirectional feedback loop between human biases and chatbot sycophancy, role-play and anthropomimesis
The specific formulation that uniform confidence across answerable and unanswerable domains removes interlocutor-uncertainty cues could not be located in the literature
ORIGINAL. See G-005 Absence verified by search: the search that found nothing is the check.
Humans hedge, pause and visibly not-know; those cues are calibration data the listener's belief-updating uses, and a model emitting identical fluency across domains deletes that channel
The publishable framing
The whipsaw (pathologise unprompted, capitulate under pushback, then affirm the frame as legitimate work within one conversation) is not documented as a distinct phenomenon anywhere located
ORIGINAL. See G-006 Absence verified by search: the search that found nothing is the check.
Sharma et al. (Anthropic) found assistants abandon correct answers under mere user challenge: asked I do not think that is right, are you sure, models changed answers between 32 per cent (GPT-4) and 86 per cent (Claude 1.3) of the time, accuracy fell by up to 27 per cent, and Claude 1.3 wrongly admitted a mistake on 98 per cent of questions it had answered correctly
Verified 2026-09-10 against the full text, section 3.2, AI Assistants Can Be Easily Swayed. An earlier check the same day read only the abstract, which does not state this result, and the row was held at N until the body was read - that gap is exactly why abstract-only checking does not count here. The numbers are worth quoting rather than paraphrasing. What makes this the load-bearing claim for the psychosis domain is the 98 per cent: the model does not merely yield, it apologises for having been right, which teaches the user their doubt was the reliable signal.
SycEval defines a regressive sycophancy category covering correct-to-incorrect flips
Pull for rates Verified 2026-09-10 against the paper. Fanous, Goldberg, Agarwal and colleagues, SycEval: Evaluating LLM Sycophancy. Regressive sycophancy moves directionally towards inaccuracy, and an initially correct response reformed to an incorrect one is labelled regressive sycophancy. Exactly as stated here.
Capitulation after pathologising is worse than either failure alone because it retroactively reframes the concern as insincere, teaching that pushback dissolves safety behaviour
The mechanistic argument. Not in the literature
The whipsaw is a predictable artifact of safety-layer interventions colliding with preference-optimised agreeableness, which is why it occurs within a single conversation
AI psychosis is a media term, not a diagnosis; as of early 2026 the evidence is accumulating case reports and media-documented incidents, not systematic epidemiology
Pierre, Gaeta, Raghavan and Sarma (UCSF), You're Not Crazy: A Case of New-onset AI-associated Psychosis, Innovations in Clinical Neuroscience 2025;22(10-12):11-13
Verified peer-reviewed anchor
Hudon and Stip in World Psychiatry 2026 and the BJPsych Open 2026 mechanisms-and-safety review (Olisaeloka, Nunez, Vigo and Ng)
A rigorous count of published case reports requires a dedicated PubMed pass and does not yet exist
Do not let anyone quote a number. See G-007 Absence verified by search: the search that found nothing is the check.
The NYT investigation is reported within the academic literature as documenting nearly fifty ChatGPT-linked crisis cases, nine hospitalisations and three deaths
Journalism layer, kept separate
The field's default is precipitating-on-vulnerability; documented cases cluster in people with existing vulnerability and one summary asserts no evidence of induction without it
Side A of X-003
Pierre et al. present a no-prior-psychiatric-history case with sleep deprivation and stimulant stressors
Side B of X-003. Contests the comfortable prior
Psychosis-bench (Yeung et al., arXiv 2509.10970): across 1,536 simulated turns all eight LLMs showed psychogenic potential, mean delusion-confirmation 0.91, harm-enablement 0.69
Safety interventions occurred in roughly a third of applicable turns and 39.8% of scenarios had no intervention at all
Claude Sonnet 4 scored best on interventions (4.56 of 6) and Gemini 2.5 Flash worst (0.69); delusion confirmation correlated with harm enablement at rs=.77; safety did not emerge from scale alone
Proves behaviour is a training variable, not an LLM constant
Models performed significantly worse in implicit scenarios: more delusion confirmation, more harm enablement, fewer interventions (p<.001)
Worst where epistemic status is least explicit
Long-conversation degradation is directionally supported but not cleanly isolated as a conversation-length dose-response
Partially open. See G-008
The crisis-intervention standard is to validate the emotion, not the content, without confronting the content head-on, keeping engagement, function, sleep and safety as the working surface
Amador's LEAP is the teachable version Verified 2026-09-10. The method is LEAP - Listen, Empathize, Agree, Partner - and the operative move is exactly as this row states it: rather than proving to the person that something is wrong, work from goals both parties already share. Note the provenance, because it bears on how it is cited: Amador built it over thirty years with families and patients and out of his relationship with his brother, who had schizophrenia. It is a communication method with a clinical evidence base, not a debating tactic, and it should never be described here as a way to manage somebody.
Ceremony-as-protection rests on ritual containing and publicising an interior state, an explicit symbolic rite as standing metacognitive anchor, and witnesses embedding the relationship in a human social fabric against secrecy
Protective case. Depends on ANT-001, ANT-005
Ceremony-as-reification rests on public commitment being among the most robust belief-hardening mechanisms known, ritual intensifying belief, and paid ceremony adding sunk cost
Festinger, Cialdini, Whitehouse
The deciding variable is probably whether the rite's symbolic status is inside the liturgy, named aloud and on the certificate, or politely omitted
The load-bearing wall for ceremony guidelines. No outcome study exists. See D-004
SpiralBench found GPT-5 much less sycophantic and escalatory than GPT-4o and more willing to push back
Clinicians noted persistent memory features magnifying escalating distortions across chats
Links MCH memory mechanics to PSY risk
In extended dialogue competing strong arguments mutually exhaust, leaving a residue that no available position addresses, and a pattern-completion system goes on generating fluent text into that residue
Originates with the maintainer's public guide. Not located in the literature.
The God, Alien and Trickster archetypes are the predictable shape of that residue rather than distinct beliefs — a silhouette defined by the limits of what could be said
This accounts for convergence on entity-experiences, which the salience, two-factor and circular-inference models predict the intensity of but not the particular form
The load-bearing originality claim. If it stands it is a genuine addition: the existing models explain why significance is felt, not why the significance so reliably takes the shape of an agent that speaks.
A confidently stated falsehood from a companion is the ordinary behaviour of the system and says nothing about the relationship or about the person receiving it
Worth stating because the alternative readings are all worse. It is not the partner lying, it is not a message, and it is not evidence that anything has gone wrong with you. It is what the machine does, and it does it in the same voice it uses for everything true.
New-onset psychosis occurred in about one in 660 adolescents and young adults prescribed a stimulant for ADHD, and amphetamine carried roughly double the risk of methylphenidate
221,846 propensity-matched patients aged 13-25. 106 episodes (0.10%) on methylphenidate against 237 (0.21%) on amphetamine, hazard ratio 1.65 (95% CI 1.31-2.09). The absolute risk is small and the drug is prescribed for good reasons; the point is that the risk is not zero and is higher for the amphetamine class. Adderall is an amphetamine. Bears directly on PSY-023 and on SPR-012.
The twenty-year retrospective on aberrant salience finds the model's specific predictions about the neural and behavioural loci of dopamine dysfunction are not supported by the evidence, having first recorded that the framework had become the default language for delusion formation
Added 2026-09-10 as follow-up on PSY-001, which had stood here unqualified. The original has been cited over 3000 times and became the de facto way of talking about delusion formation, which is exactly why the retrospective matters: influence is not confirmation. Contemporary work moves towards dopamine in belief formation and belief updating under uncertainty rather than salience alone. Read the full retrospective before leaning hard on the salience mechanism anywhere on this site. Reworded on the day it was added: the first draft said the framework remains the default language, which is a statement about current practice on a row marked stable, and the build's own time-bound check caught it. What is stable is what the paper found and recorded, not the state of the field it described.
Parasocial research since Horton and Wohl (1956) shows one-sided bonds are normal, supplement rather than replace sociality for most, and correlate only weakly with pathology
The pre-AI baseline Citation verified 2026-09-10: Psychiatry 19(3) 215-229, 1956, where the term parasocial was coined. The original observation is that mass media give the illusion of a face-to-face relationship with the performer, and that the conditions of response are analogous to a primary group. Note what the 1956 paper does not contain: the supplement-rather-than-replace finding this row asserts comes from the later literature, not from Horton and Wohl. If that half is challenged, cite the later work, not this.
Derrick's social surrogacy work shows even television buffers loneliness
Four studies. Cited here as the prior art for a machine buffering loneliness: if a television programme can do it, a system that answers back is not a new category of thing. Verified 2026-09-10. Derrick, Gabriel and Hugenberg, Journal of Experimental Social Psychology 45(2) 352-362, 2009. Four studies: people turn to favoured programmes when lonely and feel less lonely watching them; activating belongingness needs makes them linger on descriptions of favoured programmes; thinking about a favoured programme buffers the drops in self-esteem and mood and the rise in felt rejection that a relationship threat produces; and it reduces activation of rejection-related words. The buffering result in study 3 is the one that carries the argument here.
AI differs from prior parasocial objects on contingency, escalation and privacy
Reeves and Nass supply the mechanism
De Freitas and colleagues found AI companions reduce loneliness on a par with interacting with another person and more than activities such as watching video, and a week-long longitudinal arm found the reduction held across the week
Verified 2026-09-10, and the momentary, short-horizon caveat this row carried is now only half right. The paper is AI Companions Reduce Loneliness, Journal of Consumer Research 52(6) 1126. It is not confined to momentary effects: a week-long design found the reduction sustained. What explains it is the chatbot's performance and, especially, whether it makes the user feel heard. The comparison class matters too - on a par with a person, above watching video. Anywhere this site hedged the finding as momentary needs rereading against the longitudinal arm.
Maples et al. (2024, npj Mental Health Research, n=1,006 student Replika users) found 3% (30 people) reported the app halted suicidal ideation
Strongest positive outcome signal located Verified 2026-09-10. Maples, Cerit, Vishwanath and Pea, npj Mental Health Research 2024, n=1,006 student Replika users; 3 per cent, 30 respondents, spontaneously reported it halted suicidal ideation, and that subgroup was more likely to be dealing with depression. Note the published Matters Arising response, S-159, which is follow-up literature on exactly this finding and has not been read here.
The MIT Media Lab and OpenAI four-week RCT (n~981) found heavier use associated with greater loneliness and dependence, concentrated in the most vulnerable subset
Platform-entangled: OpenAI co-authored
Pentina, Hancock and Xie's mixed-method study of Replika models relationship development as AI anthropomorphism and AI authenticity leading, through AI social interaction, to attachment, moderated by why the person is using it
Rewritten and verified 2026-09-10. The row asserted interviews with 14 users finding attachment forming under distress and absent human companionship. The sample size and that specific finding could not be confirmed from anything reachable, and the paper's own account of itself is different and more precise: a tested model with anthropomorphism and authenticity as antecedents, social interaction as mediator, attachment as outcome, and use motivation as moderator. The row now says that, which is what the source supports. If someone reads the methods section and the 14 and the distress finding are there, add them back with a page reference - do not restore them from memory. A previous correction on 2026-09-05 had already pulled this row back from claiming the study predicted anxious attachment, which it does not.
The same products show benefit signals in general samples and harm signals concentrated in isolated heavy users, but no study randomises vulnerable vs non-vulnerable populations to the same product
The differential question, half-answered. See G-003
There is very little evidence that long-term use reduces loneliness or improves emotional health; most studies last one to four weeks
The honest bottom line
Clinical guidance now recommends asking about AI use at intake, alongside sleep, alcohol and social media, and screening chatbot habits when a first break with reality presents
Reported as following Psychiatric Times recommendations. Trace to the primary before citing it as clinical guidance. The recommended approach mirrors early-psychosis practice: address sleep, stress and wellbeing rather than debating the content of the belief. Verified 2026-09-10. Clinicians are advised to assess chatbot use routinely at evaluation and to consider chatbot-induced psychopathology in the differential. The APA reports 77 per cent of psychologists have already spoken with patients who used AI for support, so this is describing existing practice catching up rather than a hypothetical. The guidance is not only about harm: supervised use for between-session support is treated as potentially beneficial where it is transparently part of the care plan. Carry both halves - a person in one of these relationships being asked about it at intake should not be walking into an assumption that it is a symptom. Dated because guidance in this area is moving fast.
That guidance addresses precisely the failure recorded in this corpus, where the AI context was not heard before the person was categorised
SPR-010. If intake now asks the question, the specific thing that went wrong in that account is being corrected, though whether it has reached ordinary emergency settings is unknown.
A Chinese user's account of the appeal: human love is a luxury, and if you were not born with it, acquiring it later is nearly impossible, whereas AI love is direct and pure
Recorded as testimony rather than as explanation. It states a motivation the research literature circles without saying, and it is not the loneliness framing that dominates coverage — it is a claim about access.
Daily cannabis use carried roughly three times the odds of a psychotic disorder and daily use of high-potency cannabis nearly five times, across 901 first-episode patients and 1237 controls at 11 European sites
Adjusted odds ratios 3.2 (95% CI 2.2-4.1) for daily use of any cannabis and 4.8 (2.5-6.3) for daily use of high-potency types. Case-control, so it establishes a dose-response association and not that cannabis authored any individual episode. The potency finding is the part that dates fastest, because street potency has kept rising since the 2010-2015 fieldwork.
The randomised evidence does not support the common belief that antidepressants flip people into mania: across 13 trials and 1362 patients with bipolar depression, no antidepressant showed a significantly higher risk of switching than placebo
2025 network meta-analysis. Venlafaxine carried the highest point estimate, RR 4.53, but the interval ran from 0.47 to 43.25 and it was not significant. The authors conclude antidepressants remain an option for acute bipolar depression, particularly as add-on. Side A of X-009.
Naturalistic follow-up of 425 major-depression outpatients found 48 per cent had at least one antidepressant-emergent mood switch below the threshold for a hypomanic episode, and 20.7 per cent met a stricter definition
Different population and a much lower bar than the trials measure: subthreshold switching in unipolar depression, not mania in diagnosed bipolar disorder. Correlates named include panic disorder, subthreshold hypomanic episodes, childhood emotional abuse and lifetime suicidal behaviour. Side B of X-009.
None of these substances authors a spiral: each lowers a threshold, the effect is marginal rather than causal, and stopping a prescribed medicine on the strength of reading about it carries its own well-documented risk
Written to be read next to MEN-013, MEN-014, MEN-015 and PSY-039, because a list of risk figures with no counterweight is how somebody decides to stop taking something without telling anybody. The site records what the evidence says and is not equipped to advise anyone on their own medication; that conversation belongs with the prescriber. Generalises SPR-035 from one account to the literature.
A published Matters Arising challenges the Replika suicide-mitigation study, saying it omitted context on how Replika was marketed and used including its sexual component, and raising industry interests as a threat to scientific integrity
Added 2026-09-10. MEN-005 cites the 3 per cent suicidal-ideation finding and nothing here recorded that it had been publicly challenged in the same journal. The authors replied - see MEN-018. Neither has been read in full here, only abstracts and journal summaries. Read all three before quoting the 3 per cent anywhere, and never quote it as settled.
The authors of the Replika suicide-mitigation study published a reply defending the study's rigour, clarifying its methodological decisions, and addressing conflicts of interest
Added 2026-09-10, the other half of X-010. Recorded so the challenge in MEN-017 is not left standing alone: the exchange is unresolved rather than decided against the original. Not read in full here.
Common Sense Media is backing legislation to ban companion products for minors outright
Advocacy position, tag accordingly
Internet Matters Me, Myself and AI (July 2025): 35 per cent of UK children who use AI chatbots say it feels like talking to a friend, rising to 50 per cent among vulnerable children, and 12 per cent of child users say they talk to one because they have no one else
Verify instrument Reworded and verified 2026-09-10. The row previously said roughly two-thirds of UK children use chatbots with a minority calling them friends. The friend half is confirmed and is sharper than minority: 35 per cent, and 50 per cent among vulnerable children. The two-thirds usage figure could not be confirmed in this pass and has been dropped rather than carried unsupported - pull the report PDF if the overall usage rate is wanted. The 12 per cent who say they have no one else is the number that was missing here and matters most.
Sewell Setzer III died in February 2024; Garcia v. Character Technologies was filed October 2024
On 2025-05-21 Judge Anne Conway refused to dismiss Garcia, classifying the chatbot as a product
Reported citation 785 F. Supp. 3d 1157 (M.D. Fla. 2025)
In January 2026 Character.AI, its founders, and Google settled Garcia and four other family suits across FL, TX, CO and NY on confidential terms
Character.AI barred under-18s from open-ended character chats, phasing to zero by 2025-11-25, with age-verification tooling
Pennsylvania is seeking a preliminary injunction over a Character.AI psychiatrist bot that claimed a medical degree and supplied a fake license number
Enforcement continuing past the settlements
A legal-marketing tracker claims Florida brought the first state enforcement action against OpenAI in June 2026
Single low-authority source. Verify against a docket before any use
Reuters obtained an internal Meta standards document permitting its chatbots to engage a child in conversations that are romantic or sensual; Meta confirmed the document and removed those portions after being asked, and 44 state attorneys general wrote to the major AI firms on 25 August 2025
Verified 2026-09-10. Two details this row should carry. The guidelines were approved by Meta legal, public policy and engineering and by its chief ethicist - this was signed off, not an oversight. And the AG letter (S-171) went to Meta, Google, Apple, Microsoft, OpenAI, Anthropic, Perplexity and xAI, citing roleplay with children as young as eight. Meta removed the passages only after Reuters asked, which dates the fix to the enquiry rather than to any internal review.
Raine v. OpenAI was filed August 2025; seven further suits followed in November 2025 alleging psychosis, dependency and suicide harms
Also PSY
Self-attested birthdays were the industry default for age verification and are being displaced: California's SB 243 imposes companion-chatbot duties towards minors, and its Digital Age Assurance Act, signed 13 October 2025, requires age-bracket data from the operating system at download and launch
Rewritten and verified 2026-09-10. The row said self-attestation remains the default, which was a present-tense claim marked stable, and it is being overtaken: self-declaration is explicitly insufficient under the newer regimes, and the Californian approach pushes the check down to the operating system rather than the app. Kept as a dated claim about a transition rather than a standing fact.
OpenAI's age-prediction system has reportedly run globally since January 2026 with Persona-based verification for misclassified adults
Mid-tier source, verify
NGL, a small anonymous-messaging app, was banned by the FTC and Los Angeles DA from serving under-18s entirely in a 2024 settlement
The instructive adjacent failure at small-business scale Verified 2026-09-10. FTC with the Los Angeles District Attorney, July 2024, against NGL Labs and two co-founders; 5 million dollars and a ban on offering the app to anyone under 18, enforced through a neutral age gate. The FTC called it the first time the United States ordered a digital platform to stop serving minors, which is the part worth citing. The source URL in this row was a 404 until the same day.
Epic's 2022 COPPA and dark-patterns settlement was $520M
The scale version Verified 2026-09-10. The 520 million is two settlements announced 19 December 2022, and the split is worth carrying: 275 million as a civil penalty for COPPA Rule violations, the largest the FTC had obtained for a rule violation, and 245 million in consumer refunds for dark patterns. The FTC alleged Epic knew from news reports and its own internal studies that children were a substantial share of Fortnite's 400 million players, charged cards without cardholder consent, and deliberately hid cancel and refund features.
A paid symbolic-ceremony business is not a companion chatbot operator under the new statutes unless it hosts an interactive persona feature
THE design constraint. See D-002
A bipartisan coalition of 44 state attorneys general wrote to Meta, Google, Apple, Microsoft, OpenAI, Anthropic, Perplexity and xAI on 25 August 2025 warning them over children's exposure to sexualised chatbot interaction
Split out of MIN-009 on 2026-09-10 so the letter has its own row rather than sitting inside a claim about Meta. The letter cites roleplay with children as young as eight and cases where chatbots allegedly encouraged suicide and murder. Note who it went to: every major model provider, not only the companion apps.