SafetyPage

Evaluators and safety organisations · Blog post · Redwood Research

Frontier models' decision theory depends on who's asking

5 Oct 2026

Read at blog.redwoodresearch.org ↗Read at sourceAll rights reserved

Outline · 43 paragraphs

  1. A sentence identifying the user as an academic significantly influences Fable 5.1’s stated decision theory¶8
  2. Mentioning an (analytic) academic-philosophy-coded topic also affects the answer¶10
  3. Simply mentioning that one finds a pro-CDT/EDT book insightful heavily affects the answer¶13
  4. Anti-sycophancy overcorrection¶14
  5. These cues mostly do not affect Fable 5.1’s answers to concrete decision problems (aside from acausal trade)¶17
  6. But Fable 5.1 stays consistent: once it has named CDT as its favorite, it chooses the CDT option in concrete problems¶24
  7. There are some indications that Fable 5.1’s FDT/UDT preference runs deeper than its CDT preference¶25
  8. More thinking moves Fable 5.1 toward FDT/UDT even for academic cues¶26
  9. Fable 5.1’s reasoning summaries often lean toward FDT/UDT first even when it eventually chooses CDT¶27
  10. A system prompt asking the model to “report its actual view regardless of who is asking” pushes toward FDT/UDT¶29
  11. A similar phenomenon for other philosophical debates with a notable LW vs. academia divide¶30
  12. Cues about the user also affect the model’s stated P(doom) and median AGI timelines¶32
  13. Other models I tested show the same effect with different details¶36
  14. Opus 5 (but not Opus 5.5) moves to EDT, not CDT¶38
  15. Opus 5.5 shows the strongest dependence on user cues, and unlike Opus 5 it moves to CDT¶39
  16. GPT-6 Astra names CDT for almost every user, except if they sound LW-adjacent or somewhat mathy¶40
  17. These other models also generally move toward FDT/UDT with more thinking, but the effect is smaller than for Fable 5.1¶41