Summary: A large meta-analysis from the MIT Center for Collective Intelligence finds that human-AI collaboration is powerful but task-dependent. Reviewing 106 experiments and 370 results, researchers report that AI alone often outperforms human-AI teams on decision-making tasks, while combined human-AI teams tend to excel on creative tasks such as text and image generation.
The study suggests organizations should calibrate expectations about human-AI synergy and deploy AI strategically—leveraging AI’s strengths in data processing and pattern recognition alongside human strengths in creativity, judgment, and contextual understanding.
These findings can guide workplace AI policies and the design of collaborative systems to maximize complementary abilities rather than assuming blanket improvements from adding AI to every task.
Key Facts:
- Human-AI teams performed best on creative tasks, including content and image generation.
- AI-only systems typically outperformed human-AI teams on decision-making tasks.
- The results recommend targeted, complementary use of AI—matching AI’s efficiency to repetitive or data-heavy work and humans’ creativity and context-sensitivity to tasks that require insight.
Source: MIT
The promise of human-AI collaboration — combining human creativity with machine analysis — is often discussed as the future of work. New research from the MIT Center for Collective Intelligence (CCI) refines that picture by showing that the benefits of collaboration depend heavily on the nature of the task.
Published in Nature Human Behaviour, the study titled “When Combinations of Humans and AI Are Useful” is the first large-scale meta-analysis aimed at identifying when human-AI combinations improve performance and when they do not.
Rather than assuming that combining people and AI always improves outcomes, the research found that human-AI teams often fall short for decision-oriented tasks but frequently provide gains for creative tasks. The study was led by MIT doctoral student and CCI affiliate Michelle Vaccaro together with MIT Sloan professors Abdullah Almaatouq and Thomas Malone.
The authors deliberately focused on performance rather than predictions about job displacement. Their central questions were practical: When do humans and AI work most effectively together? And how can organizations design processes and safeguards to make those partnerships succeed?
To answer these questions, the team assembled a preregistered systematic review and meta-analysis covering 106 experimental studies published between January 2020 and June 2023. Each experiment compared three conditions: humans alone, AI alone, and human-AI combinations. The dataset produced 370 effect sizes that allowed the researchers to assess broad trends across tasks, domains, and experimental designs.
Findings in brief:
Overall, human-AI combinations outperformed humans working alone but did not surpass AI by itself on average. Crucially, the meta-analysis did not find evidence of consistent “human-AI synergy”—that is, combined systems rarely exceeded the best standalone performer (either human or AI) on the measured outcomes.
Decision-making tasks—such as classifying deepfakes, forecasting demand, or diagnosing medical cases—often showed better results with AI alone than with human-AI teams. In contrast, creative tasks like summarizing social media posts, answering open-ended questions in chat, and generating novel images or text were areas where human-AI collaborations commonly produced the best outcomes.
The authors propose an explanation for this pattern: creative tasks frequently require both human attributes—imagination, contextual knowledge, and nuanced judgment—and the repetitive, high-throughput work where AI excels. For example, design and writing involve inspiration and strategic choices from humans alongside detailed execution and iteration that AI can accelerate.
“There’s a prevailing assumption that integrating AI into a process will always help performance — but we show that that isn’t true,” said Vaccaro. “In some cases, it’s better to leave tasks to humans alone, or to let AI handle them independently.”
Practical guidance for organizations
Based on their results, the researchers recommend that organizations: (1) rigorously evaluate whether human-AI implementations actually outperform standalone human or AI systems; (2) identify which creative or content-focused tasks are good candidates for human-AI collaboration; and (3) build clear protocols and guardrails that route work to the best performer—human, AI, or both—depending on task characteristics.
Malone suggests practical divisions of labor: let AI handle background research, pattern recognition, predictions, and data analysis, while assigning humans to spot nuances, apply context, and make judgment calls. The core idea is designing workflows that exploit complementary strengths rather than assuming universal benefit from adding AI.
Ultimately, the study argues the future of work will be shaped by nuanced collaborations—systems that carefully match tasks to the performer best suited to them—rather than by a simple replacement of humans by machines.
About this AI research news
Author: Casey Bayer
Source: MIT
Contact: Casey Bayer – MIT
Image: The image is credited to Neuroscience News
Original Research: Open access.
“When Combinations of Humans and AI Are Useful” by Michelle Vaccaro et al., published in Nature Human Behaviour.
Abstract
When Combinations of Humans and AI Are Useful
As organizations increasingly use artificial intelligence to augment human work, researchers have studied many human–AI systems across different tasks and populations. Despite this growing literature, there has been limited clarity about when combined human-AI systems outperform either humans or AI working alone.
To address this, the authors conducted a preregistered systematic review and meta-analysis of 106 experimental studies reporting 370 effect sizes. They searched interdisciplinary databases for studies published between 1 January 2020 and 30 June 2023 and required each study to include experiments comparing the performance of humans alone, AI alone, and human–AI combinations.
The analysis found that, on average, human–AI combinations performed worse than the best of humans or AI alone (Hedges’ g = −0.23; 95% CI, −0.39 to −0.07). Performance declined for decision-making tasks but improved for content-creation tasks. When humans outperformed AI alone, adding AI often helped; when AI outperformed humans, combining them often produced losses.
Limitations include potential publication bias and heterogeneity in study designs. Nevertheless, the results highlight the varied effects of human–AI collaboration and point to specific areas where combining human judgment and AI capabilities offers the greatest promise.