Research map / Probabilistic and scientific modeling / Uncertainty and Gaussian processes
Using AI Uncertainty Quantification to Improve Human 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.
Human-subject experiments (445 and 600 Prolific participants) testing whether instance-level AI uncertainty quantification, calibrated with true labels and verified by Brier score, improves human accuracy and confidence calibration on binary tabular decisions.
Problem
Whether well-calibrated, instance-level AI uncertainty quantification improves human decision-making accuracy and confidence calibration beyond AI predictions alone, and whether the representation of uncertainty matters.
Approach
Generates instance-level UQ by sampling 100 Gaussian-noised clones of each instance around a random forest and showing the 95% CI of class probabilities, using known labels to verify calibration via Brier score before showing it to humans; compares point vs distribution and needle vs dotplot visualizations.
Key results
- Experiment 1 (445 participants): AI Uncertainty condition improved accuracy over AI Prediction (t(436)=2.36, p=0.049) and AI Prediction improved over Control (t(436)=9.91, p<0.0001).
- Experiment 2 (600 participants): no significant main effects of point vs distribution UQ (F(1,588)=0.28, p=0.60) or needle vs dotplot visualization (F(1,588)=0.16, p=0.69) on accuracy.
- Exploratory: only 1.33% to 7.33% of participants outperformed the AI's accuracy across conditions.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Provide an AI prediction instead of no AI information | Human decision-making · Experiment 1, all datasets | decision accuracy: t(436) = 9.91 | Helped |
| Add visualized AI uncertainty to the AI prediction | Human decision-making · Experiment 1, all datasets | decision accuracy: t(436) = 2.36, {p} = 0.049 | Helped |
| Add visualized AI uncertainty to the AI prediction | Human decision-making · Experiment 1, all datasets | confidence calibration | Helped |
| Show distribution UQ rather than point-estimated UQ | AI Uncertainty · Experiment 2, all datasets | decision accuracy: F(1,588) = 0.28, {p} = 0.60 | No clear effect |
| Use a dotplot rather than a needle visualization | AI Uncertainty · Experiment 2, all datasets | decision accuracy: F(1,588) = 0.16, {p} = 0.69 | No clear effect |
| Show distribution UQ rather than point-estimated UQ | AI Uncertainty · Experiment 2, all datasets | confidence calibration | No clear effect |
| Use a dotplot rather than a needle visualization | AI Uncertainty · Experiment 2, all datasets | confidence calibration | No clear effect |
| Use random forest alone rather than an ensemble of logistic regression, support vector machine, and random forest (ablation) | Instance-level uncertainty quantification · UQ estimation | Brier score | Helped |
| Add AI uncertainty to the AI prediction | Human decision-making · Experiment 1, all datasets | participants outperforming AI (%): 1.33% -> 4.73% | Helped |
Surprising findings
- Adding distribution-based UQ information did not improve accuracy over point-estimate UQ, contrary to pre-registered hypotheses.
- Visualization type (needle vs dotplot) had no significant effect on accuracy or confidence calibration.
Limitations
- UQ uses true class labels to calibrate, so it is not deployable as-is in real settings where labels are unknown.
- Only binary classification on three tabular datasets with a simple two-alternative forced choice task.
- Only one UQ technique was evaluated; no comparison of UQ quality levels.
Experimental setup
Datasets: UCI Census UCI German Credit UCI Student Performance
Benchmarks: UCI Census UCI German Credit UCI Student Performance
Metrics: human decision accuracy confidence calibration (multilevel model) response time Brier score
Baselines: AI prediction only control (no AI information) point-estimate UQ
Builds on: trust score
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 0.86)
- Conformal Prediction Sets Improve Human Decision Making (ICML 2024 · similarity 0.85)
- Decoding AI’s Nudge: A Unified Framework to Predict Human Behavior in AI-Assisted Decision Making (AAAI 2024 · similarity 0.85)
- Online Algorithms with Uncertainty-Quantified Predictions (ICML 2024 · similarity 0.85)
- Linear Uncertainty Quantification of Graphical Model Inference (NeurIPS 2024 · similarity 0.84)
- Learning to Decide with AI Assistance under Human-Alignment (NeurIPS 2026 · similarity 0.84)
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.