Research map / Large language models / AI and society
Decoding AI’s Nudge: A Unified Framework to Predict Human Behavior in AI-Assisted Decision Making
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.
Proposes an interpretable computational framework that models AI assistance (immediate, delayed, explanation-only) as a nudge that shifts a human's logistic decision model, fit via variational Bayes on a 202-subject diabetes-prediction MTurk experiment.
Problem
Predicting how different forms of AI assistance influence human decisions in an interpretable way, since existing black-box predictors lack interpretability and interpretable models are tailored to a single assistance type.
Approach
Treat AI assistance as a nudge that modifies a human's independent decision model (a Bayesian posterior over logistic weights) via assistance-specific weight updates (direct, affirm/contradict, or explanation-based attention shifts), giving one framework across assistance types.
Key results
- Immediate assistance: proposed method NLL 0.435, accuracy 0.800, F1 0.818, beating best baseline (logistic regression NLL 0.522).
- Delayed recommendation: NLL 0.413, accuracy 0.825, F1 0.812, outperforming all baselines.
- Explanation only: comparable to logistic regression on NLL and accuracy (0.563 vs 0.549 NLL) and best F1 (0.791); reflective DMs are more strongly nudged by AI when it affirms their initial decision (p<0.05).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use the proposed framework instead of Logistic Regression | Human decision prediction · Immediate assistance | NLL: 0.522 -> 0.435 | Helped |
| Use the proposed framework instead of XGBoost | Human decision prediction · Immediate assistance | Accuracy: 0.768 -> 0.800 | Helped |
| Use the proposed framework instead of Logistic Regression | Human decision prediction · Immediate assistance | F1: 0.789 -> 0.818 | Helped |
| Use the proposed framework instead of Logistic Regression | Human decision prediction · Delayed recommendation | NLL: 0.446 -> 0.413 | Helped |
| Use the proposed framework instead of XGBoost | Human decision prediction · Delayed recommendation | Accuracy: 0.812 -> 0.825 | Helped |
| Use the proposed framework instead of XGBoost | Human decision prediction · Delayed recommendation | F1: 0.797 -> 0.812 | Helped |
| Use the proposed framework instead of Logistic Regression | Human decision prediction · Explanation only | NLL: 0.549 -> 0.563 | Hurt |
| Use the proposed framework instead of Logistic Regression | Human decision prediction · Explanation only | Accuracy: 0.728 -> 0.715 | Hurt |
| Use the proposed framework instead of MLP | Human decision prediction · Explanation only | F1: 0.778 -> 0.791 | Helped |
| Replace the probabilistic independent decision model with deterministic logistic regression (ablation) | Proposed framework · Delayed recommendation, 5 training instances | NLL: 0.430 -> 0.514 | Hurt |
| Replace the probabilistic independent decision model with deterministic logistic regression (ablation) | Proposed framework · Delayed recommendation, 10 training instances | NLL: 0.422 -> 0.469 | Hurt |
| Replace the probabilistic independent decision model with deterministic logistic regression (ablation) | Proposed framework · Delayed recommendation, 15 training instances | NLL: 0.413 -> 0.454 | Hurt |
Surprising findings
- Under explanation-only assistance, the proposed model is not best on NLL (logistic regression 0.549 beats it at 0.563), so the advantage is not uniform across treatments.
- Intuitive decision makers are more strongly nudged by AI explanations than moderately reflective ones, contrary to the reflective-DM pattern seen for the other two assistance types.
Limitations
- Only a diabetes task with six features from laypeople; generalization to more complex tasks untested.
- Single-task framing with no sequential or temporal feedback about AI performance.
- Assumes the independent human decision model is logistic regression.
Experimental setup
Datasets: diabetes prediction dataset (500 task pool) MTurk human-subject experiment data (202 valid subjects)
Benchmarks: MTurk human-subject experiment (diabetes prediction)
Metrics: NLL Accuracy F1
Baselines: Logistic Regression XGBoost MLP SVM Utility model (Wang et al. 2022) Deterministic Decision Model
Builds on: Cumulative Prospect Theory utility model for AI reliance optimal nudge design for cognitively bounded agents LIME SHAP
Closest papers by meaning
- Utilizing Human Behavior Modeling to Manipulate Explanations in AI-Assisted Decision Making: The Good, the Bad, and the Scary (NeurIPS 2024 · similarity 0.93)
- Modeling Human Trust and Reliance in AI-Assisted Decision Making: A Markovian Approach (AAAI 2023 · similarity 0.91)
- Mecha-nudges for Machines (NeurIPS 2026 · similarity 0.88)
- Learning to Decide with AI Assistance under Human-Alignment (NeurIPS 2026 · similarity 0.86)
- Large Language Models Assume People are More Rational than We Really are (ICLR 2025 · similarity 0.86)
- Using AI Uncertainty Quantification to Improve Human Decision-Making (ICML 2024 · similarity 0.85)
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.