Human Learning Is Key to Boosting AI-Human Team Productivity, Study Finds

Widespread adoption of artificial intelligence in workflows rests on the assumption that humans and algorithms generate superior results when working together rather than operating independently. However, recent academic literature challenged this premise, pointing to evidence that human-AI teams often perform worse than their top-performing individual member. A new re-evaluation published in the Proceedings of the National Academy of Sciences suggests that earlier conclusions were overly pessimistic because prior studies neglected the impact of human learning, according to reporting by TechXplore (https://techxplore.com/news/2026-10-human-underestimated-factor-success-humanai.html).

The research was conducted by an international team led by Julian Berger and Stefan Herzog from the Center for Adaptive Rationality at the Max Planck Institute for Human Development. Their work revisited a prominent 2024 review that analyzed 74 separate studies on human-AI interaction. Those original experiments evaluated joint performance across tasks including image classification, house price prediction, and writing code. The 2024 meta-analysis had concluded that pairing workers with AI models generated negative overall synergy, meaning the combined performance fell short of the superior solo partner.

To test whether those findings overlooked critical interaction dynamics, Berger and Herzog examined how study outcomes changed when human workers were given explanations for AI recommendations alongside post-decision feedback showing the correct answer. Upon auditing the 74 underlying papers, the researchers discovered that opportunities for human learning were almost universally missing from study architectures. Only 10 of the 74 papers provided participants with feedback on correct answers after a choice was made, and none varied feedback as an experimental condition.

To extract the effect of these learning dynamics, the researchers applied meta-regression to compare studies that incorporated feedback against those that did not, factoring in whether the AI explained its recommendations. The analysis revealed that pairing AI explanations with outcome feedback enabled human-AI teams to outperform the better solo partner. By contrast, providing explanations without feedback was associated with distinctly negative synergy, while offering feedback without explanations produced no discernible performance difference.

The researchers concluded that AI explanations boost collaborative productivity primarily when human operators receive systematic outcome feedback, allowing them to verify an algorithm's reliability. Without feedback, workers cannot effectively calibrate their reliance on algorithmic suggestions. However, because none of the original 74 studies directly manipulated feedback within a single controlled experiment, the authors noted that future experiments directly varying learning opportunities are required to establish stronger causal conclusions.

The study also considered how these findings apply to generative AI and large language models. Herzog pointed out that while natural language interfaces make it simpler to deliver personalized explanations and feedback, conversational models introduce distinct limitations: because chatbots can generate inaccurate answers, their responses do not constitute independent ground-truth feedback without external verification.

The researchers also cautioned against long-term cognitive risks, noting that delegating mental tasks to AI could lead to cognitive offloading and skill degradation over time. Herzog emphasized that AI systems and workflows must be designed from the outset around humans as learning beings to realize their collaborative potential.