A framework for human–AI teamwork that accounts for how advice changes decisions

Human–AI collaboration is not simply a choice between two independent predictions. A person may answer differently after seeing an AI recommendation, a confidence score, or an explanation. In an October 2026 preprint, Tianwei Gao and colleagues propose a mathematical framework that explicitly accounts for this dependence while allowing collaboration policies to change when consultation costs or supported evaluation objectives change.

The authors call their approach collaborative omnimechanism design. Its central idea is to learn a reusable model of possible outcomes, then use that model to choose how humans and AI should work together under a specified objective.

Three decisions, not just one

The framework connects three choices: whether to consult a human, what AI-derived signal to show them, and how to combine their response with the AI prediction. If no consultation occurs, the AI prediction becomes the final output.

Here, “impressionability” has a specific meaning: the distribution of human responses depends on the presentation of AI advice. The framework therefore treats presentation as an action that can affect outcomes, rather than assuming human judgments remain unchanged across assistance conditions.

The learned model estimates outcomes conditional on the available context and candidate action. At deployment, a post-processing step evaluates predicted error plus consultation cost and selects a collaboration policy. Changing a supported cost or loss function changes this selection step, not the learned model.

That distinction has an important boundary. The adaptation assumes the context distribution and action-conditional outcome law remain fixed. It is not a guarantee of adaptation to arbitrary changes in people, tasks, or AI reliability. In supplementary experiments that changed AI advice reliability, the researchers retrained their models for each shifted setting.

Learning enough to choose well

The technical foundation is Collaborative Outcome Indistinguishability, or COI. Rather than reconstructing human behavior precisely for every input, the outcome model must approximately match the environment’s expected losses for specified comparison policies and the policies selected using the model.

The proposed COI-Boost algorithm repeatedly audits those discrepancies and updates its predictions when it finds a violation. Under the paper’s assumptions, satisfying COI gives a policy whose expected loss is close to that of the best mechanism in the prescribed comparison class—not necessarily the best conceivable collaboration strategy.

The guarantees require finite action and outcome spaces, bounded objectives, controlled policy complexity, and logging coverage: every candidate action must have a positive probability of appearing in the historical data at every context. The authors also establish a computational result for a smoothed, randomized selection rule under uniform logging and a fixed scalar consultation cost. The experiments instead use deterministic action selection and include neural outcome models as practical implementations.

What the experiments show

The first experiment uses Art and Sarcasm classification tasks. Its training and evaluation responses come from a behavioral simulator fitted to separate real human-interaction data. Results are competitive with the evaluated baselines, but neither implementation wins everywhere. On Sarcasm with free consultation, the neural policy reaches accuracy 0.806, compared with 0.792 for Global Best; the boosting policy reaches 0.785.

The radiology experiment uses observed assessments from 159 radiologists on 324 chest X-rays, without a behavioral simulator. It chooses among AI alone, an unaided radiologist, and an AI-assisted radiologist. On 65 held-out X-rays, the learned policy improves average utility over a training-selected single-action baseline at lower consultation costs. At higher costs, it selects AI alone and matches that baseline. Its diagnostic score measures deviation from reference labels, not binary classification accuracy.

The house-price experiment uses observed initial estimates and AI predictions alongside simulated human revisions. In the primary split, the policy lowers simulated absolute percentage error from 0.1335 to 0.1053 at zero revision cost. However, most of the gain comes from input-dependent selection among already available estimates. Logged-response checks show sparse support for selected revisions, and improvement is not consistent across all tested splits.

The preprint therefore offers a conditional theoretical framework and empirical evidence, rather than proof of deployment readiness. The authors acknowledge simulator dependence, limited intervention menus, and risks of manipulating judgments or encouraging overreliance. They call for transparency, auditing, and safeguards alongside optimization.