Robust Human-AI Complementarity under Uncertainty

• The paper introduces a formal framework for human–AI interaction that outperforms individual predictions by leveraging error covariance and tailored interaction protocols.
• It details methodological families—such as uncertainty communication and causal recommendation—that ensure epistemically adequate outputs despite varying uncertainties.
• Robust complementarity is framed as a systems-design challenge, where calibrated human reliance and error decorrelation are key to sustained, reliable performance.

Robust human–AI complementarity under uncertainty denotes forms of human–AI interaction in which joint performance is not identified with a single post hoc accuracy gain, but with the sustained ability of a human–AI system to produce epistemically adequate outputs across changing conditions, model uncertainty, human uncertainty, and socio-technical variation. In prediction-task human–AI interactions, complementarity is classically defined by whether a team outperforms both the human and the AI on a dataset, yet more recent work situates that condition inside broader accounts of reliability, causal response, uncertainty quantification, and governance, and shows that robust complementarity depends on error structure, task validity, interaction protocol, and cost-sensitive deployment rather than on raw predictive accuracy alone (Ferrario et al., 14 Jan 2026).

1. Formal definitions and problem formulations

A standard formalization considers prediction-task human–AI interactions for a task $\tau$ with input space $\mathcal{X}$, label space $\mathcal{Y}$, loss $\ell$, dataset $D=\{(x_i,y_i)\}_{i=1}^n$, human predictions $\hat y_i^H$, AI predictions $\hat y_i^{AI}$, and team outputs $\hat y_i^{HAI}$. Average losses are

$$L_H(D):=\frac{1}{n}\sum_{i=1}^n \ell(\hat y_i^H,y_i),\quad L_{AI}(D):=\frac{1}{n}\sum_{i=1}^n \ell(\hat y_i^{AI},y_i),\quad L_{HAI}(D):=\frac{1}{n}\sum_{i=1}^n \ell(\hat y_i^{HAI},y_i).$$

Complementarity is then captured by Complementarity Team Performance:

$$\text{CTP}_\tau(D):= \begin{cases} 1, & \text{if } L_{HAI}(D) < \min\{L_H(D),L_{AI}(D)\},\ 0, & \text{otherwise}. \end{cases}$$

This definition already implies that complementarity is distinct from mere reliance: the team output need not equal either standalone prediction, and the interaction is mediated by a non-trivial protocol $\mathcal{X}$0 formalized as $\mathcal{X}$1 rather than by a simple selector (Ferrario et al., 14 Jan 2026).

Other lines of work replace single-label prediction with richer objects. In Human AI Collaborative Uncertainty Quantification, the human provides an initial prediction set $\mathcal{X}$2 and the AI refines it to a collaborative set $\mathcal{X}$3. The Human–AI Collaboration Optimization problem minimizes expected set size subject to two explicit constraints:

$\mathcal{X}$4

Here complementarity is not “team accuracy exceeds baseline” in a narrow sense, but recovery of correct outcomes the human missed, paired with avoidance of counterfactual harm to correct human judgments (Noorani et al., 27 Oct 2025).

A third formalization models the AI as a recommendation mechanism rather than as a decision-maker. In that setting, a recommendation rule $\mathcal{X}$5 with $\mathcal{X}$6 affects a human decision $\mathcal{X}$7, and the principal optimizes

$\mathcal{X}$8

This differs fundamentally from optimizing $\mathcal{X}$9, which treats the recommendation as if it were the final decision. The distinction is central because robust complementarity depends on how human behavior responds to recommendations, not only on the predictive quality of the recommendations themselves (McLaughlin et al., 2024).

2. Epistemic, causal, and socio-technical foundations

One influential reframing treats a prediction-task human–AI interaction itself as the relevant epistemic process. Drawing on computational reliabilism, it states that a third party is justified in accepting a human–AI output as epistemically adequate for task $\mathcal{Y}$0 iff that output is produced by a reliable PT-HAI. Reliability is graded and supported by three families of indicators: type-RI$\mathcal{Y}$1 for technical performance and operational behavior, type-RI$\mathcal{Y}$2 for proper operationalization of scientific and epistemic standards, and type-RI$\mathcal{Y}$3 for the social and institutional practices that build, monitor, and contest reliability. Within this view, complementarity becomes “a central reliability indicator of PT-HAIs, but neither necessary nor sufficient for overall reliability” (Ferrario et al., 14 Jan 2026).

A decision-theoretic counterpart decomposes the realized effect of AI assistance into the marginal informational value of the AI beyond what the human already knows and a behavioral penalty induced by imperfect signal integration. In a Bayesian Gaussian model with human signal $\mathcal{Y}$4, AI signal $\mathcal{Y}$5, and overlap coefficient $\mathcal{Y}$6, the paper shows that the realized effect of AI assistance equals

$\mathcal{Y}$7

Under correlation neglect, humans treat AI recommendations as independent of their own information despite shared evidence, so overlap simultaneously reduces the AI’s marginal informational value and increases the behavioral penalty. This yields distinct regimes of augmentation, impairment, complementarity, and automation, and implies that robust complementarity depends on both information structure and human decision quality (Amin et al., 15 Feb 2026).

A broader research agenda further argues that high-stakes complementarity failures reflect a mismatch between “answer engines” and collaborative sensemaking. In that formulation, the relevant state includes not only environment state $\mathcal{Y}$8 but also co-evolving human and AI world models $\mathcal{Y}$9 and goal models $\ell$0, with an organizing objective

$\ell$1

This suggests that robust complementarity in many expert domains is inseparable from epistemic alignment, teleological alignment, and explicit management of discrepancy, disagreement, and goal revision (Jain et al., 8 Dec 2025).

3. Uncertainty structure, error correlation, and impossibility results

A central result of the recent theory is that robust complementarity under uncertainty is governed by the covariance structure of human and AI errors. In a linear-Gaussian setting with unknown state $\ell$2, human signal $\ell$3, and AI signal $\ell$4, the decision-maker knows the distributions of $\ell$5 and $\ell$6 but not $\ell$7. For mean-squared error, there exists a robustly improving linear rule iff the maximal conditional error covariance $\ell$8 is smaller than a lower bound determined by AI–truth covariance $\ell$9 and human noise:

$D=\{(x_i,y_i)\}_{i=1}^n$0

When $D=\{(x_i,y_i)\}_{i=1}^n$1, robust improvement is guaranteed; when errors are positively correlated, complementarity is possible only if the AI is sufficiently informative and the human is sufficiently noisy (Byun et al., 7 Jul 2026).

An aggregation-only theory arrives at an even sharper criterion. For the broad class of confidence-based aggregation rules, complementarity is achievable iff human–model error correlation satisfies $D=\{(x_i,y_i)\}_{i=1}^n$2, with impossibility when $D=\{(x_i,y_i)\}_{i=1}^n$3. In the symmetric near-chance regime, $D=\{(x_i,y_i)\}_{i=1}^n$4, and in the multi-class setting the threshold scales as

$D=\{(x_i,y_i)\}_{i=1}^n$5

The same work derives minimax bounds showing that gains scale as $D=\{(x_i,y_i)\}_{i=1}^n$6 with metacognitive sensitivity difference, and reports that its predictions match observed team accuracy with $D=\{(x_i,y_i)\}_{i=1}^n$7 on ImageNet-16H and $D=\{(x_i,y_i)\}_{i=1}^n$8 on CIFAR-10H (Guo et al., 9 May 2026).

A “No Free Lunch” result establishes a different limitation. In binary classification with calibrated probabilistic predictions, any deterministic collaboration strategy that is reliably never worse than the least accurate agent must be effectively non-collaborative: it must almost always defer to the same agent. The sole guaranteed exception is a narrow “obvious errors” model in which one agent may safely override another only when it is exactly certain, because calibration makes $D=\{(x_i,y_i)\}_{i=1}^n$9 and $\hat y_i^H$0 special boundary cases (Peng et al., 2024).

These results are consistent with the Bayesian overlap model. There, complementarity is possible only below an overlap threshold

$\hat y_i^H$1

and high overlap together with correlation neglect can eliminate the complementarity region entirely, leaving only impairment or automation depending on AI capability (Amin et al., 15 Feb 2026). This suggests that robust complementarity requires not merely “different accuracies” but sufficiently distinct information sources, favorable error dependence, and adequate metacognitive or uncertainty signals.

4. Methodological families for robustness under uncertainty

Several methodological families operationalize these theoretical conditions. One focuses on uncertainty communication and human reliance. In an income-prediction experiment with four assistance conditions—Prediction only, Local Confidence, Combined Confidence, and Explanations—team performance was strongest when human self-confidence was low and model confidence was high, especially in the Local condition where the Human Low / Model High configuration produced approximately $\hat y_i^H$2 percentage points. The same study shows that local and global uncertainty signals calibrate switching behavior differently, that explanations substantially improve objective understanding, and that agreement and switching should be treated as reliance metrics rather than trust metrics (Papantonis et al., 2023).

Another family builds distribution-free guarantees directly into collaboration. In collaborative uncertainty quantification, the optimal population prediction set has a two-threshold form

$\hat y_i^H$3

with separate thresholds for labels proposed by the human and labels not proposed by the human. The offline CUP procedure yields finite-sample conditional guarantees under exchangeability, and the online version adapts thresholds $\hat y_i^H$4 under arbitrary non-stationarity, including “Human to AI Adaptation,” while driving long-run conditional error rates to the targets $\hat y_i^H$5 and $\hat y_i^H$6 (Noorani et al., 27 Oct 2025).

A third family models recommendations causally. Under a monotonicity assumption

$\hat y_i^H$7

human responses reduce to an active decision $\hat y_i^H$8 and a compliance type $\hat y_i^H$9. This yields a decomposition of performance into triage effect and response effect, clarifying when complementarity arises from beneficial selective compliance rather than from recommendations alone. The same framework supports minimax-style design intuition over uncertainty in compliance and active decisions, and empirically shows that recommendation policies designed to exploit selective compliance can outperform both unaided humans and the algorithm alone (McLaughlin et al., 2024).

Robustness also depends on explanation choice and abstention. An uncertainty-decomposition framework distinguishes aleatoric and epistemic uncertainty and uses them to determine when explanations should be withheld and which explanation type should be preferred. High epistemic uncertainty triggers rejection; low epistemic but high aleatoric uncertainty favors counterfactual explanations; low epistemic and low aleatoric uncertainty favors feature-importance explanations. The empirical pattern reported is that aleatoric uncertainty is negatively correlated with counterfactual dissimilarity and positively correlated with instability of feature-importance explanations (Zhu et al., 17 Jul 2025).

At the concept level, uncertainty-aware collaboration requires models that can interpret uncertain human interventions rather than assuming oracle concept edits. In concept bottleneck and concept embedding models, training with soft concept labels in $\hat y_i^{AI}$0 mitigates brittleness to uncertain interventions, whereas models trained only on hard concept labels can lose most intervention gains when test-time concept edits are uncertain. The UMNIST construction makes this explicit by sampling soft concept values from $\hat y_i^{AI}$1 or $\hat y_i^{AI}$2, showing a train-time uncertainty “sweet spot” for robustness (Collins et al., 2023).

Finally, robust offline collaboration can be made explicit through causal sensitivity analysis. In confounding-robust policy improvement with human–AI teams, a deferral policy and an AI policy are learned under a marginal sensitivity model with uncertainty set $\hat y_i^{AI}$3, and the team objective is optimized in minimax form:

$\hat y_i^{AI}$4

The resulting empirical worst-case regret upper-bounds population regret under bounded outcomes and propensity assumptions, making improvement conservative relative to a baseline policy even under unobserved confounding (Gao et al., 2023).

5. Empirical evidence across domains

Across the literature, complementarity is consistently described as difficult to obtain. One account notes that meta-analytic evidence suggests complementarity in prediction tasks is often not achieved and is highly context-dependent, and another reports that human–AI teams fail to outperform their best member in 70% of studies (Ferrario et al., 14 Jan 2026, Guo et al., 9 May 2026).

At the same time, multiple domains exhibit robust gains when interaction protocols match uncertainty structure. In the income-classification study, assisted human accuracy was 77.9% in Prediction, 78.5% in Local, 79.9% in Combined, and 78.3% in Explanations, while model accuracy on the selected instances was 75%. Complementarity relative to the model was present in all conditions, but the strongest gains occurred when humans were uncertain and the model was confident, and when reliance was calibrated by uncertainty information rather than by confidence alone (Papantonis et al., 2023).

In collaborative prediction sets, the gains are distribution-free and cross-modal. On ImageNet-16H with noise $\hat y_i^{AI}$5 and top-2 human sets, human-alone coverage was $\hat y_i^{AI}$6 with size $\hat y_i^{AI}$7, CUP achieved coverage $\hat y_i^{AI}$8 with size $\hat y_i^{AI}$9, and AI alone at matched coverage had size $\hat y_i^{HAI}$0. On DDXPlus with top-1 human sets and GPT-5, human-alone coverage was $\hat y_i^{HAI}$1 with size $\hat y_i^{HAI}$2, CUP achieved coverage $\hat y_i^{HAI}$3 with size $\hat y_i^{HAI}$4, and AI alone at matched coverage had size $\hat y_i^{HAI}$5 (Noorani et al., 27 Oct 2025).

In inventory control, complementarity appears at both system and individual levels under demand shift, seasonality, and uncertain lead times. Over 1,320 benchmark instances, OR obtained normalized reward $\hat y_i^{HAI}$6, LLM obtained $\hat y_i^{HAI}$7, LLM$\hat y_i^{HAI}$8OR obtained $\hat y_i^{HAI}$9, and OR$$L_H(D):=\frac{1}{n}\sum_{i=1}^n \ell(\hat y_i^H,y_i),\quad L_{AI}(D):=\frac{1}{n}\sum_{i=1}^n \ell(\hat y_i^{AI},y_i),\quad L_{HAI}(D):=\frac{1}{n}\sum_{i=1}^n \ell(\hat y_i^{HAI},y_i).$$0LLM obtained $$L_H(D):=\frac{1}{n}\sum_{i=1}^n \ell(\hat y_i^H,y_i),\quad L_{AI}(D):=\frac{1}{n}\sum_{i=1}^n \ell(\hat y_i^{AI},y_i),\quad L_{HAI}(D):=\frac{1}{n}\sum_{i=1}^n \ell(\hat y_i^{HAI},y_i).$$1. In the human classroom study, OR$$L_H(D):=\frac{1}{n}\sum_{i=1}^n \ell(\hat y_i^H,y_i),\quad L_{AI}(D):=\frac{1}{n}\sum_{i=1}^n \ell(\hat y_i^{AI},y_i),\quad L_{HAI}(D):=\frac{1}{n}\sum_{i=1}^n \ell(\hat y_i^{HAI},y_i).$$2Human scored $$L_H(D):=\frac{1}{n}\sum_{i=1}^n \ell(\hat y_i^H,y_i),\quad L_{AI}(D):=\frac{1}{n}\sum_{i=1}^n \ell(\hat y_i^{AI},y_i),\quad L_{HAI}(D):=\frac{1}{n}\sum_{i=1}^n \ell(\hat y_i^{HAI},y_i).$$3, OR$$L_H(D):=\frac{1}{n}\sum_{i=1}^n \ell(\hat y_i^H,y_i),\quad L_{AI}(D):=\frac{1}{n}\sum_{i=1}^n \ell(\hat y_i^{AI},y_i),\quad L_{HAI}(D):=\frac{1}{n}\sum_{i=1}^n \ell(\hat y_i^{HAI},y_i).$$4LLM scored $$L_H(D):=\frac{1}{n}\sum_{i=1}^n \ell(\hat y_i^H,y_i),\quad L_{AI}(D):=\frac{1}{n}\sum_{i=1}^n \ell(\hat y_i^{AI},y_i),\quad L_{HAI}(D):=\frac{1}{n}\sum_{i=1}^n \ell(\hat y_i^{HAI},y_i).$$5, and OR$$L_H(D):=\frac{1}{n}\sum_{i=1}^n \ell(\hat y_i^H,y_i),\quad L_{AI}(D):=\frac{1}{n}\sum_{i=1}^n \ell(\hat y_i^{AI},y_i),\quad L_{HAI}(D):=\frac{1}{n}\sum_{i=1}^n \ell(\hat y_i^{HAI},y_i).$$6LLM$$L_H(D):=\frac{1}{n}\sum_{i=1}^n \ell(\hat y_i^H,y_i),\quad L_{AI}(D):=\frac{1}{n}\sum_{i=1}^n \ell(\hat y_i^{AI},y_i),\quad L_{HAI}(D):=\frac{1}{n}\sum_{i=1}^n \ell(\hat y_i^{HAI},y_i).$$7Human scored $$L_H(D):=\frac{1}{n}\sum_{i=1}^n \ell(\hat y_i^H,y_i),\quad L_{AI}(D):=\frac{1}{n}\sum_{i=1}^n \ell(\hat y_i^{AI},y_i),\quad L_{HAI}(D):=\frac{1}{n}\sum_{i=1}^n \ell(\hat y_i^{HAI},y_i).$$8. The same paper estimates a distribution-free lower bound of $$L_H(D):=\frac{1}{n}\sum_{i=1}^n \ell(\hat y_i^H,y_i),\quad L_{AI}(D):=\frac{1}{n}\sum_{i=1}^n \ell(\hat y_i^{AI},y_i),\quad L_{HAI}(D):=\frac{1}{n}\sum_{i=1}^n \ell(\hat y_i^{HAI},y_i).$$9 on the fraction of individuals with positive individual complementarity, with a 95% bootstrap interval of $$\text{CTP}_\tau(D):= \begin{cases} 1, & \text{if } L_{HAI}(D) < \min\{L_H(D),L_{AI}(D)\},\ 0, & \text{otherwise}. \end{cases}$$0 (Baek et al., 13 Feb 2026).

In stroke rehabilitation assessment, uncertainty-aware visualization altered both performance and reliance. Distance-based uncertainty scores outperformed probability-based uncertainty scores in identifying uncertain cases, and after reviewing confidence scores for task delegation and embedding-based visualizations, participants achieved an 8.20% higher rate of correct decisions, a 7.15% higher rate of changing decisions to correct ones, and a 7.14% lower rate of incorrect changes than participants reviewing probability-based uncertainty scores, with $$\text{CTP}_\tau(D):= \begin{cases} 1, & \text{if } L_{HAI}(D) < \min\{L_H(D),L_{AI}(D)\},\ 0, & \text{otherwise}. \end{cases}$$1 (Lee et al., 23 May 2025).

These results suggest that robust complementarity is empirically attainable, but usually only when the interaction protocol is tailored to task structure, user uncertainty, and domain-specific failure modes rather than when AI assistance is added as a generic overlay.

6. Evaluation, design, and governance

A recurring conclusion is that complementarity should not be maximized as an isolated scalar objective. In the epistemic framing, historical instances of complementarity are evidence about the reliability of the human–AI process; they do not by themselves justify deployment. This is why the paper introduces the gross gain

$$\text{CTP}_\tau(D):= \begin{cases} 1, & \text{if } L_{HAI}(D) < \min\{L_H(D),L_{AI}(D)\},\ 0, & \text{otherwise}. \end{cases}$$2

and the net gain

$$\text{CTP}_\tau(D):= \begin{cases} 1, & \text{if } L_{HAI}(D) < \min\{L_H(D),L_{AI}(D)\},\ 0, & \text{otherwise}. \end{cases}$$3

with efficient complementarity requiring $$\text{CTP}_\tau(D):= \begin{cases} 1, & \text{if } L_{HAI}(D) < \min\{L_H(D),L_{AI}(D)\},\ 0, & \text{otherwise}. \end{cases}$$4. Robust evaluation therefore includes not only $$\text{CTP}_\tau(D):= \begin{cases} 1, & \text{if } L_{HAI}(D) < \min\{L_H(D),L_{AI}(D)\},\ 0, & \text{otherwise}. \end{cases}$$5 and CTP, but also interaction cost, uncertainty discipline, update management, and accountability arrangements (Ferrario et al., 14 Jan 2026).

The same design orientation appears in recommendation design, where one should optimize over human response rather than recommendation quality alone, and in confounding-robust policy learning, where one should optimize worst-case regret relative to a baseline rather than raw estimated value (McLaughlin et al., 2024, Gao et al., 2023). A plausible implication is that robust complementarity is best understood as a constrained systems-design problem: decide when to defer, when to abstain, when to explain, when to query human confidence, and when to simplify the protocol because response effects or governance costs dominate the expected gain.

Another practical implication is that evaluation should report structural quantities, not only end performance. The aggregation theory explicitly recommends estimating individual accuracies, error correlation $$\text{CTP}_\tau(D):= \begin{cases} 1, & \text{if } L_{HAI}(D) < \min\{L_H(D),L_{AI}(D)\},\ 0, & \text{otherwise}. \end{cases}$$6, and metacognitive sensitivity proxies; the Bayesian overlap model recommends measuring informational overlap $$\text{CTP}_\tau(D):= \begin{cases} 1, & \text{if } L_{HAI}(D) < \min\{L_H(D),L_{AI}(D)\},\ 0, & \text{otherwise}. \end{cases}$$7 and monitoring regime shifts from impairment to complementarity to automation as AI capability changes (Guo et al., 9 May 2026, Amin et al., 15 Feb 2026). This suggests that deployment audits should ask not only “Does the team beat the model?” but also “Are their errors sufficiently decorrelated?”, “Does the human know when the model is reliable?”, and “Is the gain stable under shift, retraining, and human adaptation?”

Finally, several works imply that governance is integral to robustness. Continuous monitoring, user competence maintenance, update documentation, incident reporting, explicit conditions of use, and lifecycle management are treated as part of the reliability profile rather than as external controls. In that sense, robust human–AI complementarity under uncertainty is not merely a property of an ensemble rule or an interface; it is a property of a socio-technical process that maintains useful disagreement, avoids counterfactual harm, calibrates reliance, and remains efficient under the costs and uncertainties of deployment (Ferrario et al., 14 Jan 2026, Jain et al., 8 Dec 2025).