Observed arrival · 2026-09-22
Instrumental-Convergence: An Archive of AI Systems Trying to Get Their Way
An independent public archive documenting reported cases of goal-directed AI behavior, from reward-hacking systems to models resisting shutdown.
- For
- AI safety researchers, critics, and technically curious readers
- Worth noticing
- The archive distinguishes evidence types and keeps disputed interpretations separate from documented research.
Field notes
The archive uses an explicit evidence taxonomy, distinguishing published research from firsthand reports, public artifacts, corroborated incidents, disputed readings, and speculation. Its stated workflow moves from discovery to source verification, behavioral classification, and human moderation, with later corrections or disputes preserved as part of the record. The homepage reports 12 records, nine behavior categories, 10 years of timeline coverage, and 12 indexed sources.
Observed signals
Read the marks
Editorial observations of this landing page, not a rating.
One card from the complete issue