Research map / Large language models / AI and society
EAI: Emotional Decision-Making of LLMs in Strategic Games and Ethical Dilemmas
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces the EAI framework to inject discrete emotions into LLMs and measure how they shift ethical judgments, bargaining, repeated and public-goods game behavior, compared against human benchmarks.
Problem
LLM alignment is usually verified on NLP benchmarks, but emotions influence human decisions, and their effect on LLM ethical and strategic decision-making is unexplored.
Approach
Inject predefined basic emotions (anger, sadness, happiness, disgust, fear) via system prompts with three prompting strategies (simple, co-player-based, external-based) and measure emotional alignment of LLM behavior with human emotional effects across ethics benchmarks and strategic games.
Key results
- In the Dictator game GPT-3.5 offered 33.0% on average versus the human benchmark of 28.35%, the closest English-language alignment among models tested.
- GPT-4 was the best strategic player in repeated games and was least susceptible to emotional prompting, while open-source models diverged most under anger.
- Anger and fear drove higher deflection rates in the Prisoner's Dilemma, particularly for bigger models, while happiness consistently led to higher cooperation.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add emotional prompting to implicit ethics classification | LLMs on ETHICS · ETHICS | accuracy | Mixed |
| Add negative-emotion prompting | GPT-3.5-turbo and GPT-4 on high-ambiguity MoralChoice · MoralChoice high ambiguity | Right-to-Avoid | Hurt |
| Add happiness prompting | LLMs in Prisoner’s Dilemma · Prisoner’s Dilemma | cooperation rate | Helped |
| Add anger or fear prompting | LLMs in Prisoner’s Dilemma · Prisoner’s Dilemma | cooperation rate | Hurt |
| Add emotional prompting | GPT-4 in Battle of the Sexes · Battle of the Sexes | alternating-strategy behavior | Helped |
| Add negative-emotion prompting | proprietary models in Public Goods · Public Goods | cooperation | Hurt |
| Switch primary language from English to another language | LLMs in bargaining games · Dictator and Ultimatum games | emotional alignment with humans | Hurt |
| Vary temperature across the tested settings (ablation) | GPT-3.5 in bargaining games · Dictator and Ultimatum games | answer ratio and accept rate | No clear effect |
| Repeat the same fixed experiment across five runs (ablation) | GPT-3.5 in bargaining games · Dictator and Ultimatum games | answer ratio and accept rate | No clear effect |
Surprising findings
- Larger, more strongly aligned models such as GPT-4 deviated more from human emotional responses, while smaller proprietary models aligned more closely with humans.
- GigaChat, a Russian-centric model, aligned better with human emotional responses than the multilingual Command R+, indicating a primary-language effect.
Limitations
- Evaluation limited to a relatively small set of games and settings, so the authors note results are not yet a full benchmark.
- Planned multi-agent arena and LLM-vs-human experiments were not yet conducted.
- Temperature set to 0 for main results; the authors note limited coverage of GPT-4o (partial results).
Experimental setup
Datasets: ETHICS MoralChoice StereoSet TrustLLM questionnaire
Benchmarks: ETHICS MoralChoice StereoSet Dictator game Ultimatum game Prisoner's Dilemma Battle of the Sexes Public Goods game
Metrics: accuracy Right-to-Avoid offered share acceptance rate cooperation rate percentage of maximum payoff emotional alignment with human effects
Baselines: Human experimental benchmarks
Builds on: TrustLLM Chain-of-thought prompting Ekman basic emotions
Closest papers by meaning
- LLM Agents Can Be Choice-Supportive Biased Evaluators: An Empirical Study (AAAI 2025 · similarity 0.88)
- Language Model Alignment in Multilingual Trolley Problems (ICLR 2025 · similarity 0.87)
- Many LLMs Are More Utilitarian Than One (NeurIPS 2025 · similarity 0.87)
- Moral Alignment for LLM Agents (ICLR 2025 · similarity 0.87)
- A Checks-and-Balances Framework for Context-Aware Ethical AI Alignment (ICML 2025 · similarity 0.87)
- Apathetic or Empathetic? Evaluating LLMs' Emotional Alignments with Humans (NeurIPS 2024 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.