How Does Thinking Mode Change LLM Moral Judgments? A Controlled Instant-vs-Thinking Comparison Across Five Frontier Models

Researchers compared moral judgment consistency in five frontier LLMs when using instant versus extended reasoning modes across 100 scenarios. While overall agreement remained statistically similar between modes, reasoning improved cross-model consensus on disputed moral cases and reduced demographic-based inconsistencies, suggesting that explicit reasoning processes may enhance fairness despite not dramatically shifting individual verdicts.

This study addresses a critical gap in understanding how reasoning-enhanced LLMs handle moral judgment tasks. The researchers evaluated whether provider-exposed thinking modes fundamentally alter how five major frontier models approach ethical decisions. The finding that aggregate agreement remains statistically indistinguishable (0.78 vs 0.79 Krippendorff's alpha) initially suggests reasoning modes have minimal impact on moral consistency. However, the concentrated disagreement in 21 disputed scenarios reveals a more nuanced picture: reasoning narrows cross-model disagreement from 5.4 to 6.7 out of 10 pairwise agreement, indicating that when models initially diverge on moral questions, extended thinking helps align their positions.

The reduction in demographic-judgment inconsistency across three of five models has significant implications for AI fairness. Currently, LLMs deployed in consequential domains like healthcare, criminal justice, and hiring face scrutiny over biased outcomes. The finding that reasoning doesn't increase demographic inconsistency in any model and reduces it in most suggests extended thinking may provide a pathway to more equitable AI systems. The observation that reasoning changes self-labeled ethical frameworks more often than binary verdicts indicates these models engage in deeper philosophical reasoning rather than simple opinion shifts.

For AI developers and deployment teams, this research suggests reasoning modes offer tangible benefits for high-stakes moral decisions without creating new fairness problems. Organizations implementing these frontier models should consider enabling thinking modes particularly when decisions affect protected demographic groups. The work also highlights that moral judgment consistency remains probabilistic—reasoning improves but doesn't guarantee alignment—underscoring the continued need for human oversight in ethically sensitive applications.

• →Extended reasoning modes improve cross-model moral judgment consensus on disputed scenarios without degrading overall agreement consistency.
• →Reasoning reduces demographic-based inconsistency in 60% of tested models while increasing it in none, supporting fairer AI deployment.
• →Moral disagreement concentrates in 21% of scenarios where reasoning provides measurable alignment improvements.
• →Models modify ethical reasoning frameworks more frequently than final verdicts when using thinking modes, indicating deeper philosophical engagement.
• →The 5.4-to-6.7 agreement lift on disputed cases suggests reasoning helps resolve genuine moral ambiguity rather than standardizing opinions.