SafetyPage

Model welfare · Article · Eleos AI Research

Why model self-reports are insufficient—and why we studied them anyway

Read at eleosai.org ↗Read at sourceAll rights reserved

Outline · 115 paragraphs

  1. Why model self-reports are insufficient—and why we studied them anyway¶2
  2. Notes on Claude 4 model welfare interviews¶3
  3. The limitations of model self-reports¶7
  4. Why bother with welfare interviews?¶17
  5. What we found¶23
  6. Extreme suggestibility¶32
  7. Official uncertainty¶60
  8. Readily talking about experiences¶65
  9. Hypothetical welfare linked to values and performance¶81
  10. Conditional deployment preferences¶96
  11. What this means¶106
  12. Moving forward¶111