Selection methodology

Status: experimental, transparent, and corrigible.

This methodology protects AI Trend Watch from becoming an indiscriminate A.I.-news stream. A candidate must be timely, genuinely influential, substantially relevant, and rich enough to support analysis.

Stage 0 — Hard gates

A video normally requires:

  1. a stable primary URL;
  2. identifiable speakers and publication date;
  3. a transcript, captions, or another sufficiently exact record;
  4. material discussion of A.I., AGI/ASI, alignment, governance, risk, deployment, psychology, or mass psychology;
  5. enough claim density to justify a report;
  6. no unresolved provenance problem that prevents responsible attribution.

Failure at Stage 0 normally postpones publication rather than producing a speculative reconstruction.

Stage 1 — Candidate score

Each dimension receives 0–4 points. The score is a triage aid, not a substitute for judgment.

Dimension 0 2 4
Recency no current relevance recent but not time-sensitive newly published and central to the current A.I. cycle
Virality no unusual reach strong reach for the channel or field exceptional velocity, reach, or cross-platform spread
Expert relevance weak or promotional connection materially informed participant direct, high-level technical or institutional responsibility
Substantive density mostly slogans several analyzable claims dense causal claims, forecasts, evidence, and remedies
Project relevance little connection one strong connection multiple direct links to the ACCM Deep Ethics Project research object
Longitudinal value little revisitable content some testable predictions clear forecasts or warnings suitable for later tracking

Working threshold: normally 18/24 or higher, with both Virality and Project relevance at 3 or higher. The score and evidence should be published with the report. Exceptions require a written reason.

What counts as viral

Virality is contextual. The report should capture, with a date:

  • public view count;
  • time since publication;
  • approximate view velocity when available;
  • engagement relative to the channel’s normal audience;
  • reuse by other major channels, media, researchers, policymakers, or A.I. communities;
  • visible cross-platform spread;
  • whether the reach arose from substantive interest or title/controversy effects.

A single universal view threshold would favor entertainment-scale channels and hide field-specific signals. The evidence—not merely the word “viral”—must remain inspectable.

Exclusion and postponement reasons

A candidate may be rejected or postponed when:

  • the headline is viral but the content is thin;
  • the speaker’s claimed expertise is materially misrepresented;
  • no reliable primary source or transcript is available;
  • the video mainly repeats a report already covered;
  • the relevant claims cannot be separated from promotional material;
  • publication would violate the section’s frequency limit;
  • the report would add volume rather than high-signal correspondence.

Political inconvenience, criticism of powerful institutions, unfamiliarity, or disagreement with the project are not exclusion criteria.

Selection record

For every published report, preserve:

  • date considered and date selected;
  • selection score and supporting observations;
  • who or what nominated it;
  • any material conflict of interest;
  • whether an automated discovery system contributed;
  • why it outranked other candidates;
  • the evidence snapshot used to call it viral.

Rejected candidates do not require a permanent public dossier. Consequential or disputed rejections may be logged when doing so improves accountability rather than producing noise.

Symmetry requirement

The same evidentiary discipline applies whether a speaker:

  • agrees or disagrees with the project;
  • represents a major A.I. lab or an independent research group;
  • warns of catastrophic risk or warns about the harms of catastrophic-risk framing;
  • advocates stronger regulation or warns about regulatory capture;
  • is celebrated, controversial, or institutionally protected.

Publication gate

Automated systems may assist with discovery, transcription, scoring, source retrieval, and drafting. Final publication requires review for:

  • fidelity to the source;
  • claim-type separation;
  • accurate links and timestamps;
  • proportionality;
  • provenance;
  • correction of hallucinated or inferred details;
  • clear separation between the evidence-based report and the ACCM Deep Ethics Project reaction.

Return to AI Trend Watch.