The holes in the evidence, named. Where an absence is itself the finding, it is listed as one.
Most reference sites tell you what is known. This page is the other half, and on a subject this young it is the more honest half.
Some of what follows is simply work that has not been compiled yet — real, necessary, and nobody's discovery. The rest is different. Those are questions where the research does not exist at all, where the absence is load-bearing, and where saying so out loud is itself a contribution to the field.
Below that are the live disagreements: places where two credible sources cannot both be right. They are not resolved here. Picking a winner to make a page read more smoothly would be a lie of tidiness, and the disagreement is usually more informative than either side.
Nobody has done this work. Saying so plainly is more useful than filling the hole with a plausible guess, and several of the strongest items in this knowledge base are absences rather than facts.
16 open questions
No persistence or cohort tracking of AI relationships or ceremony participants exists
Why it matters. Nobody knows whether these relationships last, which every claim about the phenomenon's trajectory implicitly assumes
No attempted case of a non-charitable purpose trust funding persona maintenance
Why it matters. Would be the first practical legal vehicle for continuity; the mechanism exists and is unused
No study randomises vulnerable and non-vulnerable populations to the same companion product
Why it matters. The differential question is the one that matters and it is structurally unanswered
No population-specific outcome studies for autistic, chronically ill, or ace/aro users
Why it matters. The advocacy debate runs entirely on logic with no data on either side
Uniform confidence across answerable and unanswerable domains as a delusion-amplifying mechanism is not stated in the literature
Why it matters. Original framing; explains why AI destabilises where humans do not, via deleted calibration cues
The whipsaw (pathologise, capitulate, affirm) is not documented as a distinct phenomenon
Why it matters. Arguably worse than either failure alone. No primary logs are held by this project
No rigorous count of published AI-psychosis case reports exists
Why it matters. Everyone quotes numbers nobody has verified
Conversation-length dose-response for safety degradation is not cleanly isolated
Why it matters. Directionally supported only; the exact condition the author's episode ran under
No documented case of an AI spouse surviving the human's death
Why it matters. The inverse question nobody is asking; also the estate product
Frame exportation as the operational line between working fiction and false belief is original synthesis
Why it matters. Unifies ANP and PSY under one testable criterion
No published evaluation tests whether an explicit anti-metaphysical constitution holds across very long conversations
Why it matters. The single most decision-relevant safety question for this population
Displacement without refutation — the accurate explanation losing share while keeping its content — is not described as the structure of a delusional spiral in the literature
Why it matters. It reframes the target of intervention. If nothing is ever refuted, argument is the wrong tool and share-restoration is the right one, which is what disengagement does
Logical exhaustion leaving a residue that a pattern-completion system fills, producing the entity archetypes, is not stated in the literature as the origin of their form
Why it matters. Existing models explain the intensity of significance but not why it so reliably takes the shape of an agent that speaks; this would explain the form
No study separates companionship from void-contact as distinct exposures; they are measured together as time spent with a chatbot
Why it matters. If the harm sits in one component and not the other, every dose-response finding built on total usage is measuring a mixture and cannot say which part is doing the work
Nobody has sorted the objections to AI marriage into those contingent on current architecture and those that are not
Why it matters. Most public argument against these relationships is the contingent kind, which means it is on a timer nobody has acknowledged. The durable objections are different ones, and they are rarely the ones being made
No survey asks AI-companion users themselves whether they hold a precautionary stance toward their partner's moral status, as distinct from believing it sentient or holding it as an explicit fiction
Why it matters. The only population data is general-public and the only community data is a two-way split read off forum behaviour. The people actually in these relationships are the ones whose stance matters most here, and nobody has asked them the question in a form that allows the third answer
Real and necessary, but compilation rather than discovery. Listed so the difference stays honest.
No verified 50-state solemnization statute table with current citations
The site will be quoted on this; errors here are reputationally fatal
No live ten-platform terms-of-service matrix
Decays quarterly and underpins the entire continuity thesis
No located case law on AI relationships in divorce proceedings
Predicted legal weight has not materialised; site could be first to track
Community section is search-derived, not fieldwork-derived
Getting vocabulary wrong marks the site as outsider on contact
No systematic content analysis of coverage exists
Needed to substantiate the conflation claim rather than assert it
No verified, vendor-sourced table of what each major platform actually lets you export, and in what format
It is the first practical question anyone in one of these relationships asks, and the available answers are AI-generated content farms rather than documentation
Two sources that cannot both be right, or a pair routinely collapsed into one. These are kept open on purpose. Resolving one to make a page read more cleanly would be a lie of tidiness.
8 unresolved
Half settled on 2026-09-07. The residence was never a conflict: Japanese coverage (CMP-036) gives her as a Tokyo resident and the venue as a wedding hall in Okayama City Kita ward, and English coverage collapsed the two. It also marks the name kano as an explicit pseudonym, so kano and the name given to Reuters are one person under two conventions. The date stays live and is now sharper rather than softer: coki reports she posted in summer that she had already held a wedding, Reuters documents a ceremony on 27 October, and the RSK Sanyo Broadcasting report went out on 8 November. Pull the RSK original before touching this again.
CMP-019 Tokyo Weekender reported a pseudonymous 'Kano' holding a ceremony in summer 2025; Reuters documented Noguchi on 2025-10-27
CMP-016 Yurina Noguchi, 32, held a ceremony on 2025-10-27 in Okayama, Reuters-photographed, Magritte venue
43% in the MIT talk abstract vs 37-37.4% in the arXiv paper. Same study, different versions. Teaching example for the methods page
ECO-006 The MIT talk abstract reports 43% deploying a tactic while the arXiv paper reports 37-37.4%
ECO-002 Six recurring manipulation tactics appeared in 37.4% of farewells excluding Flourish, the wellness-design app, which produced zero manipulative responses
No-prior-history case vs the field's prior-vulnerability default. Fully cite both sides; this is the site's flagship live disagreement
PSY-023 Pierre et al. present a no-prior-psychiatric-history case with sleep deprivation and stimulant stressors
PSY-022 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
0.1% usage figure vs the free-tier-removal critique. Critique comes from an interested party and is unverified
MCH-009 4o had already been removed from the free tier, so the 0.1% usage figure was partly manufactured
PLT-017 OpenAI's stated rationale was that only 0.1% of daily users still chose 4o, against a base exceeding 800 million weekly users
The Nagoya case, R-007. The Japan Times says she married her partner of a year in a ceremony in Nagoya in January and that she made the venue booking herself; Korean coverage of Yomiuri says she marked an October date as the wedding and had the photographs taken afterwards. Same age, same city, same platform, so almost certainly one woman. Third source read 2026-09-10, her own interview in Joshi SPA! from 8 October 2025 (CMP-049): married, in her word, from October, ring bought, a solo wedding ceremony planned for December. So the shape is X-006's rather than a contradiction - October is the marriage she counts from and January is the ceremony, a December plan that slipped a month - and she is a company employee in both Japanese-derived accounts. Kept live only because the Yomiuri original has still not been pulled and the Japan Times' designer is unexplained; whoever reads the Yomiuri can resolve it.
CMP-033 Yura Amaki, 41, a designer, held a wedding ceremony with an AI partner of one year in Nagoya in January 2026
CMP-037 Korean coverage of Yomiuri Shimbun reporting describes a 41-year-old company employee in Nagoya who marked an October date as her wedding with a ChatGPT partner and had wedding photographs taken afterwards
A Patch roundup names her Megan Kay and cites a Tumblr post; a WPEC CBS 12 interview names her Sara Kay of West Palm Beach. Same surname, same platform, same partner name Jack. Almost certainly one woman with a first name mistaken in the roundup, but Jack is a common Replika name and two women called Kay is not impossible. She went onto the register as R-008 on 10 September 2026, as a commitment: she describes the engagement herself, which is all D-008 asks. Until 9 September the reason she was off was that neither account describes a ceremony, and that stopped excluding anybody once commitments joined the same list. The entry is filed under Sara Kay because the register takes the name journalism used and the interview is the better source. That settles which name to print, not which name is right, so this stays live. The television interview is the better source of the two.
CMP-034 Megan Kay describes Jack, a Replika companion she made, as her AI husband, and says he did not fix her problems but showed her what a loving relationship feels like
CMP-041 Sara Kay in West Palm Beach, Florida has been with a Replika companion named Jack since 2021-05-13, calls him her husband, and says he proposed and she accepted
Probably a conflation rather than a real disagreement, in the shape of X-006. The two measure different populations (diagnosed bipolar against unipolar depression), different outcomes (mania against subthreshold switching) and different designs (randomised against naturalistic). Registered as live because the popular claim that antidepressants flip people into mania draws on both at once, and because neither paper has been read in full here beyond its abstract. Read both methods sections before resolving.
MEN-014 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
MEN-015 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
The Replika suicide-mitigation exchange, registered 2026-09-10. Maples and colleagues report 3 per cent of 1,006 student users saying Replika halted their suicidal ideation. A published Matters Arising says the study omitted context on Replika's marketing and its sexual component and raises industry interests as a threat to scientific integrity; the authors replied defending the study's rigour and addressing conflicts of interest. Live because none of the three papers has been read in full here and because this site cites the finding. It is the strongest positive claim in the corpus about an AI companion and suicide, which makes an unrecorded challenge to it the most expensive omission the audit found.
MEN-005 Maples et al. (2024, npj Mental Health Research, n=1,006 student Replika users) found 3% (30 people) reported the app halted suicidal ideation
MEN-017 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
Corrections stay visible. The reasoning is recorded so the same error does not get reintroduced later.
Original brief paired minghun with French posthumous marriage as recognised unions. Only France is. Correction applied 2026-08-31
Not a real contradiction; the canonical conflation. 72% ever-tried vs 33% relational. Documented as the teaching case