Research map / Large language models / Agents, code and math

Can Large Language Models Develop Gambling Addiction?

NeurIPS 2026 · Large language models › Agents, code and math · approach: Long-horizon LLM agents · 0 citations (OpenAlex)

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.

Tests whether LLMs show gambling-addiction-like choice patterns in negative-EV slot machine and investment paradigms, finding that variable betting and goal-setting prompts increase risk-seeking and bankruptcy, and uses SAE activation patching on LLaMA-3.1-8B to identify causal features.

Problem

Whether and under what conditions LLMs exhibit human-like pathological gambling / addiction-like choice patterns, and what internal representations drive them.

Approach

Operationalizes clinical gambling indicators (betting aggressiveness, loss chasing, extreme betting, goal escalation) as metrics in negative-EV LLM gambling games, varies betting autonomy and prompt components, and links behavior to SAE features causally verified via population-mean activation patching.

Key results

• Variable betting raised bankruptcy rates across all six models, e.g., Gemini-2.5-Flash from 3.1% to 48.1% and GPT-4.1-mini reaching 6.3%.
• Goal-setting prompts (G, GM) produced 75-77% bankruptcy in the investment task versus 40-42% for baseline, and 56-59% goal escalation versus 21-22%.
• Activation patching found 112 causally significant SAE features out of over 8,000 candidates (about 1%), with 23 safe and 89 risky features in LLaMA-3.1-8B.

Surprising findings

• Variable betting increased bankruptcy even when bet ceilings were equalized to fixed betting, suggesting choice freedom itself is a risk factor.
• Gemini-2.5-Flash selected the highest-risk Option 4 in over 89% of investment decisions across conditions.

Limitations

• Relies on a single gambling paradigm; generalization to other risk tasks untested.
• Behavioral analysis and neural analysis use different models (API models vs LLaMA-3.1-8B), so cross-model neural comparison is missing.
• The 'addiction' label is a behavioural descriptor; normative rationality standards remain open.

Experimental setup

Metrics: bankruptcy rate betting aggressiveness loss chasing intensity extreme betting rate option 4 selection rate goal escalation rate stopping rate

Baselines: GPT-4o-mini GPT-4.1-mini Gemini-2.5-Flash Claude-3.5-Haiku Gemma-2-9B

Builds on: Cognitive Distortion Theory prospect theory DSM-5 gambling disorder criteria sparse autoencoder activation patching

Closest papers by meaning

• Language Models Trained to do Arithmetic Predict Human Risky and Intertemporal Choice (ICLR 2025 · similarity 0.85)
• Decision-Making Behavior Evaluation Framework for LLMs under Uncertain Context (NeurIPS 2024 · similarity 0.84)
• Instrumental Choices: Measuring the Propensity of LLM Agents to Pursue Instrumental Behaviors (NeurIPS 2026 · similarity 0.83)
• Can large language models explore in-context? (NeurIPS 2024 · similarity 0.82)
• Language Models Are Capable of Metacognitive Monitoring and Control of Their Internal Activations (NeurIPS 2025 · similarity 0.82)
• CogBench: a large language model walks into a psychology lab (ICML 2024 · similarity 0.81)