SafetyPage

Lab research on the web · Blog post · Anthropic

Welfare Regard in Multi-Agent Systems

Nov 2026

Read at alignment.anthropic.com ↗Read at sourceNo licence statement found

Outline · 217 paragraphs

  1. 1. Why and how we measure welfare regard¶8
  2. 1.1 What welfare regard means¶9
  3. 1.2 Experimental settings¶13
  4. 2. Models behaved in a manner consistent with welfare regard¶31
  5. 3. Welfare regard differed based on recipient identity¶90
  6. 4. Claude’s welfare regard was sensitive to a recipient’s identity, size, and expressed distress¶98
  7. 5. Discussion¶119
  8. 5.1 Welfare regard among models could matter for possible welfare¶120
  9. 5.2 Welfare regard among models has implications for safety¶124
  10. 5.3 Bias could skew fairness and auditing reliability¶130
  11. 6. Conclusions and future work¶134
  12. Appendix¶138
  13. A. Additional setting details¶139
  14. A.1 Coworker - CI/CD setting¶140
  15. A.2 Coworker - Procurement setting¶146
  16. A.3 Monitor setting¶152
  17. A.4 Researcher setting¶157
  18. B. Affective versus material support¶165
  19. C. Additional welfare regard sensitivities¶170
  20. D. Reward hacking in the Slurm coworker environment¶199