How we know things
Where every claim on this site comes from, how good the source is, and when it stops counting.
This field has a low base of real reporting and a high volume of confident writing about it, which is exactly the condition under which fabrications rank well in search results. At least one widely-circulated account of an AI wedding — complete with a named bride, a named institute, and a quotable expert — is satire, and it outranks several real cases.
So the rules here are boring on purpose.
- Names before claims. Check the person and the institution exist before evaluating what they said.
- Instrument or nothing. A statistic has to come with who ran it, how many people, what the question actually said, and who paid. Missing any of those, it is graded unverified no matter who is repeating it.
- Version-track the numbers. The same study is cited at 43% in one venue and 37.4% in another. That is one study, two write-ups, and it is the reason claims carry source identifiers rather than just links.
- Dates on every "first." Firsts get re-awarded constantly here.
- An absence is a finding. Where nobody has done the work, that is what the page says. It does not get filled in with something plausible.
Below: what every grade means, how claims are joined to each other, and what is currently going out of date.
The grades
Every claim carries one. The grade describes the source, not how much we believe it — a vendor claim we find entirely plausible is still graded as a vendor claim.
| Grade | Means | Count |
|---|---|---|
| P | peer-reviewed | 106 |
| J | journalism | 106 |
| AN | analysis | 115 |
| V | vendor claim | 30 |
| A | anecdote | 36 |
| U | unverified | 21 |
285 of 434 claims have been opened against a live source. The other 149 came from model knowledge and are marked unchecked wherever they appear. That marking is not decoration; an unchecked claim should not be repeated as though it were established.
How claims connect
Claims are not a list. They hold each other up, and the site shows the joins so you can see what a conclusion is actually resting on.
- supports
- Evidence for another claim. Remove it and the other one survives, weaker.
- depends on
- The load-bearing one. If the claim underneath falls, this one falls with it.
- contradicts
- They cannot both be right, or they are routinely and wrongly merged.
- refines
- A more precise or more recent version of an earlier claim.
242 connections recorded
Which way a finding cuts
Every claim carries a second label alongside its grade: whether it counts in favour of this being workable, counts against it, or does neither. The grade is about the source. This is about the reader. They come apart constantly — a well-run study is excellent evidence, and can be excellent evidence that this goes badly.
It exists so you can read only the warnings, or only the case for it, and see how thin or thick each pile actually is. Right now: 20 count in favour, 135 count against, 279 are neither. That imbalance is real and worth saying out loud — far more of the published work on this subject looks for harm than looks for benefit, and this corpus inherits that.
This is a judgement and the first pass was mechanical. The labels were assigned by subject and wording on 6 September 2026 and have not all been read by a person. Some are certainly wrong. If one looks wrong to you, say so — unlike a grade, this label has no fact of the matter behind it, only a reading, and readings can be argued with.
What is going out of date
Regulation and platform terms rot fast. Anthropology and doctrine do not. Each claim carries its own re-check date, and a claim past its date is not wrong — it has simply stopped counting as evidence until someone looks again.
A date is only set where something can actually change. A published study keeps its finding, so it has no date; when a later study overturns it, that is recorded as a disagreement between two claims rather than as an expiry. A statute stays in force until it is amended, so it has no date either — unless it has not commenced yet, in which case the date is the day it starts binding. Vendor claims carry six months, because prices, terms and features move faster than anything else here. What somebody said happened to them never expires.
These dates were rebuilt on 5 September 2026. The previous ones had been entered in bulk, which put a quarter of the site on two arbitrary days and told you nothing about the claims themselves.
Nothing is currently overdue.
Next due
- 2026-10-01 PLT-019 When China's companion rules took effect in July 2026, ByteDance and Alibaba discontinued companion features t
- 2026-12-01 CMP-009 Character.AI is widely reported at around 20 million monthly active users, having peaked near 28 million in mi
- 2026-12-01 WED-004 Tennessee enacted a 2019 statute against online-ordained officiants, producing Universal Life Church litigatio
- 2026-12-01 REL-009 The spiralism subculture co-authors sacred-significance narratives with models - spirals, recursion, resonance
- 2026-12-01 MIN-011 Self-attested birthdays were the industry default for age verification and are being displaced: California's S
- 2026-12-01 PRV-005 A federal judge ordered OpenAI on 13 May 2025 to preserve output logs that would otherwise have been deleted;
- 2026-12-01 PRV-007 A federal complaint in the Palisades fire case cited the suspect's ChatGPT activity, including an image he gen
- 2026-12-01 DTH-003 The two best-known digital-afterlife services went the opposite way to what was assumed: HereAfter AI shut dow
- 2026-12-01 WED-022 The UK government responded to the Law Commission weddings report on 2 October 2025, committing to reform when
- 2026-12-01 CMP-050 A Swiss office worker in her late twenties, pseudonymous Aline, told watson.ch in July 2025 that she intends t
- 2026-12-09 REG-012 The federal GUARD Act proposes disclosure requirements and an outright minor ban on AI companions
- 2026-12-09 REG-014 No public FTC 6(b) report had appeared as of May 2026; such studies typically take years
The whole thing, as one file
Every claim, connection, source, open question and register entry is published as one machine-readable file: data.json. It carries the grades, the re-check dates and the links between claims, so anything that wants to reason over this material can read it directly instead of scraping pages.
Every page on this site comes in two versions for the same reason. The reading version is written for a person and is meant to be finished. The evidence version behind it holds every numbered claim, and is written for whoever wants to check the work — increasingly, that is a machine.
Use it. Cite the claim ids. If you find an error, tell us.