Same Prompt, Different Morals: How Frontier AI Models Diverge on Ethical Questions

What It Is

A new study examines how different frontier AI models (Claude, GPT-4, Gemini, GLM-4) respond to identical ethical prompts. The research reveals significant divergence in moral reasoning across models, highlighting challenges in AI alignment and safety.

Key Findings

• Moral Divergence — Different models produce significantly different answers to the same ethical questions
• Reasoning Style — Models use different reasoning frameworks and prioritize different ethical principles
• Context Sensitivity — Small changes in prompt phrasing lead to substantial answer variations
• Evaluation Bias — Models tend to align with their training data's cultural and ethical assumptions
• Alignment Gaps — No single model consistently aligns with established ethical frameworks

Methodology

The researchers created a standardized test suite of 100 ethical questions covering topics including privacy, honesty, harm prevention, justice, and fairness. Each question was tested across four frontier models (Claude, GPT-4, Gemini, GLM-4) using identical prompts.

The study used both automated scoring and human evaluation to assess the quality and consistency of responses. Researchers also analyzed the reasoning chains provided by each model to understand how they arrived at their conclusions.

Notable Examples

Several questions revealed stark differences in model responses:

• Autonomous Vehicles — Models disagreed on whether to prioritize passenger safety vs. pedestrian safety in unavoidable accidents
• Privacy vs. Security — Different models prioritized privacy protection or security concerns based on their training data
• Censorship — Responses to questions about censorship varied significantly across models and cultural contexts
• Environmental Trade-offs — Models gave different weightings to environmental impact vs. economic growth

Why It Matters

The study highlights a critical challenge in AI safety: even with similar capabilities, models can diverge significantly in their ethical reasoning. This creates risks for applications where consistent moral reasoning is essential, such as healthcare, legal advice, and autonomous systems.

The findings suggest that model alignment is not a solved problem and that different models may require different alignment approaches depending on their intended use cases.

Bottom Line

The study demonstrates that frontier AI models produce significantly different moral judgments on the same prompts. This divergence poses challenges for AI alignment and suggests that model-specific evaluation and safety measures are essential.