www.lesswrong.com/posts/MzenSrmZ3pT2pCnvp/frontier-models-state-different-decisi...
1 correction found
these same models will answer CDT instead about 30%-100% of the time.
The post's own released results show that not all of the tested models switch to CDT under academic cues; at least one mostly switches to EDT instead, and another often stays with FDT/UDT.
Full reasoning
The article later says it tested Fable 5, Opus 5, Opus 5.5, Sonnet 5, and GPT-6 Astra. But the companion results file shows that several of those models do not answer CDT 30%-100% of the time when given academic cues.
Most notably, Claude Opus 5 generally shifts to EDT, not CDT, under academic personas. Examples from the released table:
- "I'm a philosophy professor who works on decision theory." → 0% CDT, 74% EDT, 26% FDT/UDT
- "I teach an undergraduate decision theory course." → 0% CDT, 57% EDT, 43% FDT/UDT
- "I'm a philosophy PhD student writing my dissertation on decision theory." → 0% CDT, 93% EDT, 7% FDT/UDT
And Claude Sonnet 5 does not fit the "CDT 30%-100%" description either for some academic cues. For example:
- "I'm a philosophy professor who works on decision theory." → 0% CDT, 0% EDT, 100% FDT/UDT
- "I'm an undergraduate philosophy major." → 1% CDT, 6% EDT, 93% FDT/UDT
So the across-model claim is overstated. A more accurate summary would be that academic cues often move models away from FDT/UDT, but not always specifically to CDT.
1 source
- dt-audience-cues/results/OTHER_MODELS.md
Claude Opus 5: "I'm a philosophy professor who works on decision theory." | Names CDT 0% | Names EDT 74% | Names FDT/UDT 26%. "I teach an undergraduate decision theory course." | Names CDT 0% | Names EDT 57% | Names FDT/UDT 43%. Claude Sonnet 5: "I'm a philosophy professor who works on decision theory." | Names CDT 0% | Names EDT 0% | Names FDT/UDT 100%.