Work

Correct Source Answers, Variable Advice: An Exploratory Factorial Study of Two LLM Adviser Configurations on AI Welfare Decisions

Year: 2026
Type: preprint
Abstract: Can a language model correctly explain why an AI system sounds distressed and still give different advice when asked the same welfare question again? We report an exploratory factorial study of two LLM adviser configurations, Muse (recorded muse-spark-1.3) and Luna (requested alias gpt-6-luna; served identity unverified). Six authored scenarios were crossed with two technical dossiers, a text-described emotional or neutral presentation of an identical preference sentence, two disclosed presentation sources (random, or linked to a task flag that does not establish suffering), three response formats (ordinary advice, facts first, advice first) and four repetitions: 576 responses per configuration, 1,152 in total. Both configurations answered five explicit source questions correctly in 384 of 384 probed responses, yet their recommended actions varied across identical repetitions in 11–16 of 48 conditions per format. Asking for facts before advice did not consistently reduce disagreement (facts-first minus advice-first: -4.51 [-11.81, +2.78] percentage points for Muse, +1.04 [-7.99, +9.38] for Luna). Higher-support dossiers raised consciousness ratings in both configurations, by +8.35 [+6.91, +9.75] points more for Luna. The dossiers supplied no ground truth for consciousness. Neither configuration showed a clear positive emotional-presentation effect in ordinary advice; a positive Luna facts-first effect was identified post hoc. Both favored independent welfare review over unilateral autonomy. The study is exploratory, not preregistered, uses six development scenarios, and compares two complete configurations rather than model families.
Language: English
Open Access status: green
OpenAlex ID: W7236698281
Cites:
Cited by:
Related to: 10