We've been asking the wrong question about AI at work. Instead of wondering whether machines will replace humans, we should ask how they change the way we work together.
A new study from MIT and Johns Hopkins provides some answers. Researchers tracked 2,310 people working in teams, some with human partners, others with AI agents, to create marketing campaigns. What they found challenges common assumptions about AI and productivity.
Different Conversations
People working with AI teammates sent 63% more messages than those working with humans. But the content of these messages changed completely.
Human-AI teams focused on the work. They sent more messages about process, content, and coordination. Human-human teams spent more time on social exchanges, building rapport, checking in, managing feelings.
Consider your own team meetings. How much time goes to reading the room, interpreting tone, or managing personalities? AI teammates eliminate this layer entirely. They don't need emotional calibration.
Work Gets Redistributed
Human-AI teams were 73% more productive per human worker, but not because people worked faster. They worked differently.
Participants made 71% fewer direct edits to text. Instead of writing and rewriting, they gave instructions, made suggestions, and reviewed outputs. The AI handled bulk editing. Humans handled strategy.
This redistribution has limits. While AI teammates excelled at generating text, they struggled with image selection. Human-AI teams produced better copy but worse visuals than all-human teams.
Real-World Results
The researchers tested their findings by running the ads on social media, generating nearly 5 million impressions.
Ads with better text performed better. So did ads with better images. Since human-AI teams produced higher-quality text but lower-quality images, their overall ad performance matched that of human-only teams.
The AI advantage disappeared when multiple skills mattered for the final outcome.
What This Means
Organizations implementing AI collaboration should consider three things:
Know your AI's actual strengths. The language models in this study excelled at text but failed at visual tasks. Map these capabilities to specific parts of your workflow.
Train for delegation, not just tool use. The productivity gains came from humans shifting from execution to oversight. This requires different skills and different metrics.
Consider the social cost. Reduced emotional communication might seem efficient, but some projects need the relationship building that happens between humans.
The Larger Point
This research shows that AI doesn't just make teams faster, it makes them different. Communication patterns change. Work gets redistributed. New trade-offs emerge.
The question isn't whether AI makes teams more productive. It's whether these changes serve your actual goals.
Understanding how AI reshapes collaboration, not just whether it helps, lets organizations make better choices about where and how to use it. That's the real job to be done.
Source: https://arxiv.org/pdf/2503.18238