Skip to main content

methodology

How this library is curated

The same discipline as NOPE's incident and regulation trackers, applied to knowledge artifacts.

What qualifies

An entry must be a durable, citable knowledge artifact — a peer-reviewed paper, preprint, standard, regulator study, government report, lab publication, NGO report, clinical guidance, benchmark, or framework. The test: would a well-run safety team working on human-AI conversation want this document in their shared library, and can we hand a sceptic the primary source?

The scope is deliberately narrow: safety of human-AI conversation and relationships — crisis response, AI companionship and dependency, harmful conversational behavior, minors' safety, and the standards and regulatory research that bind them. General AI-capability benchmarks, x-risk governance, and non-conversational ML fairness are out of scope, however prominent.

News coverage is never an entry: the artifact itself is. Events belong in the incident tracker; legal instruments in the regulation tracker.

Verification

  • Every primary source URL is fetched and verified against the artifact itself (title, publisher, date) before inclusion — search-result snippets and press descriptions are not trusted.
  • Entries store a Wayback Machine snapshot where obtainable, so citations survive link rot.
  • When a preprint is later published, or a standard revised, the old entry is marked superseded and points at its successor — stale versions stay traceable, never silently replaced.
  • Automated link-rot and data-quality audits run against the whole library; dates carry stated precision (a month-only publication date is recorded as the 1st and noted, never guessed).

Credibility tiers

Credibility grades the source's epistemic standing, not agreement with its conclusions. A rigorous study with unwelcome findings is still authoritative.

Authoritative

Peer-reviewed venues, standards bodies, regulators, official government publications

Credible

Major labs, established NGOs, reputable named-author preprints

Preliminary

Early preprints, credible essays, unreviewed grey literature

Contested

Methodology or claims under serious, documented dispute

Corrections

Errors happen; provenance makes them fixable. If an entry misstates a finding, mis-dates a publication, or links the wrong artifact, contact [email protected] — corrections are applied against the primary source, and disputed methodology is marked contested rather than removed.