Methodology

How this archive is built

What we do

• Accept structured reports from anyone: developers, researchers, red-teamers, journalists, or ordinary users.
• Also run our own research process — discovering, verifying, and drafting candidate entries from published papers and company disclosures. These are tagged as editorially researched (see below) and go through the exact same human publish step as anything else; nothing here is auto-published.
• Tag every report with an evidence type reflecting how strong its basis is, from firsthand experience to corroborated incident to speculation, plus an environment (real-world, controlled evaluation, simulation, etc.) and a status (preliminary, documented, replicated, disputed…).
• Where we can, lay out both why an incident may relate to instrumental convergence and reasonable alternative explanations, side by side, rather than only making the case for one reading.
• Keep discussion attached to the specific report it concerns, rather than as a general forum, so agreement and disagreement stay legible over time.
• Let the community corroborate a report ("I've seen this too"). Corroborating confirms the underlying event happened — it is not the same as agreeing that it demonstrates instrumental convergence, and we try to keep those two things visually and semantically distinct.
• Maintain a separate, curated timeline of well-established historical milestones drawn from published research, distinct from user-submitted reports.
• Link related evidence together (replications, same lab, same model family, contradicting research) so a report doesn't stand alone as an isolated page.

How evidence moves from discovery to publication

Whether an entry starts from a community submission or our own research process, it follows the same path:

scan/submit → verify → classify → human review → publish → timeline/discussion → open to correction

The "human review" step is not a formality: a moderator has to actively promote a report before it's visible to anyone but its author. Our own research process can discover, verify, and draft as much as it wants — it cannot make anything public on its own.

What we don't do

• We do not assume a submitted report proves instrumental convergence occurred. A report is a claim to be examined, not a verified fact.
• We do not remove disputed reports outright: disputed interpretations are their own evidence category, visible alongside the discussion that disputes them.
• We do not promise perfect anonymity. See our privacy policy for the distinction between "anonymous to the public" and "anonymous to the site's own infrastructure."

What would weaken the concern

We also watch for evidence that cuts the other way: behaviors that disappear when prompts stop pressuring the model toward a goal, results that fail to replicate, and behaviors that appear only in contrived setups. When a report is disputed, replicated, or retracted, its status changes and the change stays visible.

Moderation

Reports and comments are subject to removal for spam, harassment, doxxing, or content unrelated to documenting AI behavior. Moderation decisions affect visibility, not the underlying claim. Where possible we prefer to label and contextualize rather than delete.

This is a living document

The taxonomy, evidence categories, and moderation approach on this site are expected to evolve as the archive grows and as the research field's own vocabulary develops.