The Human-AI Complementary Performance Gap

50 concepts30 papers1 proposal

When we design artificial intelligence to assist human experts, we operate under an intuitive assumption: a human combined with an AI should perform better than either could alone. This synergy is known as complementary performance. Mathematically, we can express it as:

$$\text{Performance}(\text{Human} + \text{AI}) > \max(\text{Performance}(\text{Human}), \text{Performance}(\text{AI}))$$

If a team fails to clear this bar, the human intervention is either redundant or actively harmful.

Despite the theoretical promise of human-in-the-loop systems, a growing body of empirical research reveals a sobering reality: human-AI teams frequently fail to achieve complementarity. Instead, they encounter a complementary performance gap, where the combined team underperforms compared to the AI operating entirely in isolation.

The Collaboration Paradox

Why does adding human oversight degrade high-capacity models? In a systematic review of clinical applications, Peng Fei Liu in Human-AI teaming in healthcare: 1 + 1 > 2? documents a "collaboration paradox." While AI assistance significantly elevates the performance of junior practitioners, senior experts face diminishing returns. As an autonomous system's standalone accuracy increases, the human's ability to effectively integrate its outputs decreases, leading to a lower ratio of captured potential.

This integration failure is not unique to medicine. Guoyong Wang et al. in Human–large language model collaboration in clinical medicine: a systematic review and meta-analysis show that human intervention frequently introduces new cognitive biases or errors that override a model's correct baseline judgments. Similarly, Maia Jacobs et al. in How machine-learning recommendations influence clinician treatment selections: the example of antidepressant selection demonstrate that incorrect algorithmic recommendations paired with simple explanations lead experts to abandon their own accurate insights, actively degrading team performance.

The root of the problem lies in mutual error detection. As Gagan Bansal et al. observe in Does the Whole Exceed its Parts? The Effect of AI Explanations on Complementary Team Performance, achieving complementarity is exceptionally difficult when humans and AI systems have comparable skill levels. To improve the joint score, the human must accurately identify and correct the AI's subtle errors while simultaneously recognizing when their own intuition is flawed. In practice, humans struggle to navigate this boundary, oscillating between blind overreliance and stubborn rejection.

The Illusion of Explainability

To bridge this trust gap, the research community has heavily invested in Explainable AI (XAI). The intuition seems straightforward: if the AI explains why it made a prediction, the human can better evaluate its reliability.

However, controlled experiments show that current explanation paradigms do not solve the complementarity gap. Laura Marusich et al. in Does Explainable Artificial Intelligence Improve Human Decision-Making? found that while AI predictions help humans, adding explanations provides no significant boost to team accuracy. Instead, explanations often act as a justification engine, inducing a false sense of security that causes humans to follow the AI even when it is wrong.

This failure stems from a lack of verifiability. The authors of In Search of Verifiability: Explanations Rarely Enable Complementary Performance in AI-Advised Decision Making argue that an explanation only adds value if it provides a "shortcut to the truth"—allowing the human to quickly and independently confirm the recommendation. Most feature-attribution or counterfactual explanations do not offer this verifiability; they merely present a plausible-sounding narrative.

Furthermore, Han Liu et al. in Understanding the Effect of Out-of-distribution Examples and Interactive Explanations on Human-AI Decision Making demonstrate that interactive explanations can inadvertently reinforce existing human biases, failing to help users catch model failures even on out-of-distribution data. From a theoretical standpoint, Chacha Chen et al. prove in Machine Explanations and Human Understanding using causal diagrams that model-derived explanations cannot improve a human's task comprehension unless they explicitly interact with pre-existing, accurate human intuition. If the human lacks domain expertise, explanations alone are mathematically insufficient to achieve complementarity.

Information Asymmetry and Side Information

To understand how to actually achieve complementarity, we must look at human-AI collaboration through an information-theoretic lens.

Complementarity requires information asymmetry. If the AI and the human utilize identical data features to arrive at their conclusions, their knowledge is redundant. Ziyang Guo et al. formalize this in The Value of Information in Human-AI Decision-making by evaluating teams against an idealized decision-maker. They show that team performance is optimized only when the AI provides complementary information—insights the human has not already exploited—and when the interface successfully signals when to rely on that information.

Redundant Information:
[Human Knowledge: Features A, B, C] <---> [AI Model: Features A, B, C] 
Result: High agreement, low complementarity. Human cannot correct novel AI errors.

Complementary Information:
[Human Knowledge: Features A, B + Side Info] <---> [AI Model: Features A, B, C]
Result: Information asymmetry. Potential for team performance to exceed individual baselines.

This asymmetry is precisely where human expertise can shine. Rohan Alur et al. introduce the concept of algorithmic indistinguishability in Human Expertise in Algorithmic Prediction. While an AI model might achieve high average accuracy across a dataset, there exist specific subsets of data where the algorithm is effectively "blind"—predicting no better than a random guess because it lacks access to alternative data modalities. If a human expert possesses unique "side information" (such as a patient's physical demeanor or unquantified contextual clues), they can systematically correct the model in these blind spots, pushing team performance past the AI baseline.

Designing for Joint Performance

To close the complementarity gap, researchers are shifting away from optimizing models in isolation and moving toward optimizing the interaction loop.

1. Embracing Strategic Miscalibration

Conventional machine learning wisdom dictates that models should be perfectly calibrated (i.e., a confidence score of 80% should mean the model is correct exactly 80% of the time). However, Kailas Vodrahalli et al. challenge this in Uncalibrated Models Can Improve Human-AI Collaboration. Because humans do not process probabilistic advice linearly, intentionally making a model uncalibrated—often by exaggerating its confidence—can alter human trust dynamics in ways that maximize the final joint team score.

2. Granular, Instance-Level Evaluation

Evaluating systems using aggregate metrics (like an overall accuracy of 92%) obscures the asymmetric failure modes that break human-AI teams. As Ryan Burnell et al. argue in Rethink reporting of evaluation results in AI, high aggregate scores can hide instances where an AI fails on a task that is trivial for a human. Moving to instance-by-instance reporting allows designers to see exactly where human and machine capabilities interlock.

3. Structural Workflow Interventions

Instead of a simple sequential setup where a human reviews and rubber-stamps an AI's final output, alternative interaction architectures can preserve collective intelligence. For example:

Ultimately, overcoming the complementary performance gap requires treating the human and the AI not as a supervisor and a tool, but as a joint system whose combined information dynamics must be deliberately engineered.

Key Foundational Work

How machine-learning recommendations influence clinician treatment selections: the example of antidepressant selection (2021) — contributed 3 concept elements to this direction.
Understanding the Effect of Out-of-distribution Examples and Interactive Explanations on Human-AI Decision Making (2021) — contributed 2 concept elements to this direction.
Does Explainable Artificial Intelligence Improve Human Decision-Making? (2021) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_complementary_performance_gap_6e70e3c7,
  title = {complementary performance gap},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/complementary-performance-gap/rdr_6e70e3c7f6624916aef1b6f5e00f7944},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-22},
}

This exact version (version 1, frozen text):

@misc{lacuna_complementary_performance_gap_6e70e3c7_v1,
  title = {complementary performance gap},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/complementary-performance-gap/rdr_6e70e3c7f6624916aef1b6f5e00f7944/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-22},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake