Research map / Multi-agent systems and games / Human-AI and goal-conditioned

Mecha-nudges for Machines

NeurIPS 2026 · Multi-agent systems and games › Human-AI and goal-conditioned · approach: Boundedly rational agent modeling · 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.

Introduces "mecha-nudges" (environment changes that steer AI agents without degrading human decision environments) and measures them in bits via V-usable information, finding a significant post-ChatGPT rise in machine-usable information in six million Etsy listings.

Problem

Changes to how choices are presented may systematically influence AI agents acting in human environments, but this has not been measured at scale; existing nudge and prompt-injection frameworks do not capture environment-driven influence on machines.

Approach

Formalizes mecha-nudging by combining Bayesian persuasion with V-usable information, so that the machine-side influence of listing text is measured in bits under a constraint that human-usable information does not materially decline; applies pointwise V-information with a fine-tuned Llama proxy for a GPT-5-mini curation label across pre/post-ChatGPT Etsy listings.

Key results

• Post-ChatGPT Etsy listings gain 0.143 bits of machine-usable information (p<0.01), over 40% of the 0.355-bit maximum possible increase.
• Effect is robust across labeling models, prompts, token pairs, and fine-tuning models, with most point estimates between 0.09 and 0.17; it drops to 0.117 with full listing controls.
• Human study finds post-ChatGPT listings have 0.015 fewer bits of human-usable information (95% CI [-0.044, 0.014]); LLM-rewritten placebo gives only 0.018 and DailyMed shows no shift.

Surprising findings

• Effect is absent in categories where buyers are sensitive to AI use (books, music, art and collectibles) and above average for electronics.
• Among pre-ChatGPT listings, more machine-usable information is associated with 25.2% fewer reviews, reversing post-ChatGPT association.

Limitations

• Observational data; cannot establish causal effect of ChatGPT or seller intent.
• Machine curation labels come from proxy LLMs (GPT-5-mini), not actual ChatGPT behavior; V-information estimated with a single fine-tuned 8B model giving only a lower-bound-type estimate.
• Human-side constraint tested indirectly and via a small Prolific study with self-elicited probabilities.

Experimental setup

Datasets: Etsy product listings (Bright Data, ~6M listings) DailyMed drug labels (placebo) Prolific human study responses

Benchmarks: Etsy listings (post vs pre-ChatGPT PVI regression) DailyMed placebo

Metrics: pointwise V-information (bits) OLS coefficient on post-ChatGPT indicator human-usable information (bits)

Baselines: DailyMed placebo Rephrase placebo (LLM-rewritten listings)

Builds on: Bayesian persuasion V-usable information predictive V-information nudge theory

Closest papers by meaning

• Decoding AI’s Nudge: A Unified Framework to Predict Human Behavior in AI-Assisted Decision Making (AAAI 2024 · similarity 0.88)
• AI Alignment with Changing and Influenceable Reward Functions (ICML 2024 · similarity 0.87)
• Utilizing Human Behavior Modeling to Manipulate Explanations in AI-Assisted Decision Making: The Good, the Bad, and the Scary (NeurIPS 2024 · similarity 0.86)
• LLM Agents Can Be Choice-Supportive Biased Evaluators: An Empirical Study (AAAI 2025 · similarity 0.85)
• Instrumental Choices: Measuring the Propensity of LLM Agents to Pursue Instrumental Behaviors (NeurIPS 2026 · similarity 0.85)
• ACED-Bench: Evaluating How LLM Agents Acquire and Act on Causal Evidence (NeurIPS 2026 · similarity 0.85)