Research map / Large language models / AI and society
Human-AI Teaming Through the Lens of Calibration
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.
Analyzes human-AI teaming (combination and delegation) through statistical calibration, proving that existing combination methods fail to preserve human calibration and that delegation requires finely calibrated rejectors, with simulations and ImageNet-16H/HAM10000 experiments.
Problem
Whether common human-AI teaming frameworks (combining predictions and learning to defer) preserve the calibration of the human and model members, and what calibration the rejector needs when the human has hidden features.
Approach
Treats the human and model as calibrated with respect to arbitrary partitions of the feature space and uses this unified lens to show how calibration propagates through model-based combination, Bayes confusion-matrix combination, and delegation (L2D); delegation preserves member calibration but shifts the requirement to the rejector's calibration granularity.
Key results
- In simulation, the Bayes-optimal model-based combination is calibrated w.r.t. the model partition (ECE 0) but miscalibrated w.r.t. the human partition (ECE 0.090 shared, 0.144 independent).
- On ImageNet-16H with shared features, the best posterior accuracy occurred at model ECE of 0.22 and the best posterior ECE at model ECE of 0.1, so better model calibration did not yield a better combined predictor.
- On HAM10000 with a meta-data-only simulated human (independent features), the posterior achieved its best accuracy, ECE and NLL for the best-calibrated image classifier.
Surprising findings
- Better-calibrated models can yield worse Bayes confusion-matrix combined predictors when the human and model share features.
- Multi-calibrating the model to the human partition collapses the team into the model, undermining the motivation for the human's inclusion.
Limitations
- Assumes the human and model are calibrated, which is hard to verify in practice.
- No learning is performed; results are theoretical with small-scale simulations for delegation.
Experimental setup
Datasets: ImageNet-16H HAM10000 CIFAR-10H
Benchmarks: ImageNet-16H HAM10000 CIFAR-10H
Metrics: accuracy expected calibration error negative log-likelihood
Baselines: model-based combination Bayes combination of classifiers and confusion matrices Bayes-optimal deferral rule
Builds on: Bayes confusion-matrix combination learning to defer multi-calibration omnipredictors
Closest papers by meaning
- A No Free Lunch Theorem for Human-AI Collaboration (AAAI 2025 · similarity 0.89)
- Designs for Enabling Collaboration in Human-Machine Teaming via Interactive and Explainable Systems (NeurIPS 2024 · similarity 0.85)
- Path Dependence under Adaptive AI Delegation (NeurIPS 2026 · similarity 0.84)
- Automating Data Annotation under Strategic Human Agents: Risks and Potential Solutions (NeurIPS 2024 · similarity 0.84)
- Human Expertise in Algorithmic Prediction (NeurIPS 2024 · similarity 0.84)
- Position: Towards Bidirectional Human-AI Alignment (NeurIPS 2025 · similarity 0.83)
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.