Teams, Trust, and AI: Reflections from the MIT Media Lab

By Donna Scarola & Stela Lupushor

Experience over opinions

We spent time at the MIT Media Lab in early October with three people from the Center for Constructive Communication (CCC): Dimitra Dimitrakopoulou, Principal Research Scientist and Head of Translational and Practice; Russell Stevens, Head of Strategy and Development and co-founder of Cortico; and Maximilian Holschneider, a graduate student whose work really inspired us. 

The phrase that stayed with us was simple: experience over opinions. It turns out there is real evidence behind it.

In its Person Over Position research, CCC ran three studies with nearly 2,000 American adults. People who shared a personal experience were seen as more authentic, more trustworthy and more emotionally resonant than people who simply stated an opinion. Even when listeners disagreed with someone's position, hearing the experience behind it created empathy that got past their defensiveness.

The researchers drew a pointed conclusion. Their findings challenge the traditional view that rational debate is the main path to mutual understanding. What works better is creating spaces where people can safely share the lived experiences that shape their views.

Much of how we run organizations assumes the best argument wins. We build business cases, we debate positions, we push for alignment. We rarely ask people what they have seen and lived that leads them to think what they think. That is a design choice, and it may be the wrong one.

Interaction is the work

None of this is entirely new. It fits a long line of research on what makes teams effective.

Sandy Pentland's work at MIT's Human Dynamics Lab equipped teams with wearable sensors and tracked how they interacted. The most important predictor of a team's success was its communication patterns. Those patterns mattered as much as intelligence, personality and talent combined. His team could predict which groups would outperform just by looking at how they communicated, without meeting a single member.

Anita Woolley and colleagues reached a similar conclusion in Science. They found a measurable "collective intelligence" factor in groups. It was not strongly tied to the average or highest individual intelligence in the room. It tracked members' social sensitivity and how evenly people took turns in conversation.

Put those together and the message is clear. Teams are systems of interaction, not collections of smart individuals. Trust and judgment are not soft skills sitting on top of the work. They are the work.

AI cuts both ways

This is where AI gets complicated, and where I think the stakes are highest for us.

The message I heard at MIT was that AI can make human groups stronger when it reinforces connection, trust and judgment. It weakens them when it substitutes for those things. The research shows both edges.

Harvard's field experiment at P&G, with 776 professionals, is the clearest example. Individuals working with AI matched the output of two-person teams working without it. AI also helped people produce more balanced solutions across technical and commercial lines. Those are real gains.

But read the result another way. If one person with AI can do what a team used to do, the easy path is to stop convening the team. The authors themselves note that AI's affirmation can feel productive in the moment while dulling people's evaluative judgment. Fewer interactions. Less productive friction. Weaker judgment. That is how silos form, quietly, one efficient shortcut at a time.

Judgment matters even more because the tools are uneven. CCC research tested leading language models and found they gave less accurate and less truthful answers to users with lower English proficiency, less formal education, or from outside the US. The researchers concluded these models are least reliable for the users who are most vulnerable. For a global, diverse workforce, that is not a footnote. It is a reason to keep human judgment in the loop.

Disagreement has two layers

The most practical idea I saw was an AI note taker the Media Lab team built and Max (Maximilian Holschneider) presented to us.

Before a meeting, it asks each participant essentially one question: what do you think this meeting or decision is actually about? It collects those answers and surfaces them to the group before discussion begins.

The insight behind the solution is rooted in their thesis that discourse has two layers. There is what we substantively disagree on. And there is how differently we interpret the question in the first place. Most meetings burn their time on the second layer. We think we are debating a decision when we are really debating what the decision is. Clear that up front, and the group can spend its energy on the harder stuff that has impact on outcomes.

There is a long lineage here, and Max pointed us to it. Medieval monk-scholars like Lanfranc of Bec and Gratian borrowed Aristotle's logic and applied it to law: set the conflicting authorities side by side, reason through them, then decide. Church courts built that into a written procedure where each side's arguments went on the record before judgment. The French crown noticed. In 1247 Louis IX sent literate friars across the kingdom to hear local grievances case by case, and within a generation French royal courts had replaced trial by combat with proof by witnesses and given one judge the job of writing up both sides before the court deliberated.

What those monks built was a map. Decisions were recorded, precedent accumulated, and a shared sense of what a fair and logical outcome looked like took shape over time. That is the role I want AI to play for us. The monk did not replace the judge or the parties. He captured what each of them knew, laid it out in order, and made the reasoning visible so the humans in the room could decide well. The question MIT is asking is what that practice looks like in the age of AI, with experts and machines in the same room and the goal of sharper reasoning and healthier discourse. I find that a far more interesting question than whether AI can take better notes.

Disagreement is a skill you can practice

CCC is also treating constructive disagreement as a skill, not a trait.

In Squat Rack, an MIT project that used a voice-based system where people spar with a deliberately difficult AI counterpart, then debrief with a coaching agent on specific conversational micro-skills. Think asking open, curious questions, not jumping to blame, and grounding arguments in lived experience. The diagnosis of traditional training is honest: sessions are infrequent and dependent on a facilitator, so people leave with vocabulary but very little practiced skill. The team is now running randomized experiments to see whether that practice transfers to real human conversations.

A related project, Agora, lets people practice building consensus. An AI interviewer spoke with 90 American workers about the minimum wage. Each interviewee became an AI persona that predicts how they would react to a proposed policy, with predictions about 85% accurate to their actual beliefs, and every response tied to verbatim clips from their real interview. Users draft a policy, hear the lived experiences behind each reaction, revise, and watch support shift in real time.

Notice the design choice in both. The AI is the sparring partner, not the replacement. The goal is better human dialogue, not less of it.

Go deep, not wide

Another idea that stuck with me when it comes to understanding people, you cannot go wide. You have to go deep.

The American Conversation Project puts this into practice. It is a national, nonpartisan initiative run by the National Conference on Citizenship in partnership with Cortico, the nonprofit that works alongside CCC. Its goal is to rebuild the civic muscle of listening by bringing communities together for honest, recorded conversations. So far it counts 257 conversations, 102 hosts, 1,102 participants and nearly 13,000 minutes of recorded dialogue. One host's reflection captured the spirit well: participants did not agree on every issue, but what mattered was their willingness to listen and engage respectfully.

The scale is the point. A thousand people is tiny by survey standards. But every one of those minutes is a real voice telling a real story, which CCC's tools can then make sense of. Anthology turns raw conversation audio into interactive maps people can explore. Their Conversation Health research uses language processing and visualization to make the dynamics of a conversation visible, including how participants relate to each other and how facilitators shape the exchange.

For those of us who live in engagement surveys and dashboards, this is a useful provocation. A score tells you that something is wrong. A story tells you why. We have spent a decade getting better at measuring sentiment, but we may need to get much better at listening.

Redesigning how we gather

The work most directly relevant to us is a new initiative called New Social Forms, which CCC is running in collaboration with Google Xi. Dimitra's translational team is the bridge between this research and the organizations that want to test it.

Its premise is that many of the ways we come together were inherited rather than designed, and built to convey information rather than build trust. People have redesigned how they gather before when the moment demanded it. The Athenian agora was built so merchants, philosophers and magistrates would regularly cross paths. Alcoholics Anonymous started as an experiment in radical honesty among strangers.

The initiative's first year focuses specifically on teams. The argument is that meetings, decision processes and workshops were built to move information, not to support the complexity of human collaboration now. They are asking what it would look like to redesign those formats deliberately, and whether technology can amplify the result. 

What this means for us as operators and practitioners

The social systems of work are being reordered by AI, and it is happening fast. We can let that reordering happen to us, or we can design it. A few questions we can all ponder as we consider the future of work and changing dynamics of today:

  • Connection or substitution? Where are our AI tools connecting people, and where are they quietly replacing the conversations that build trust?
  • Scores or stories? Where are we measuring sentiment when we should be listening to experience?
  • Debate or understanding? Do meetings surface how differently people interpret the question before we argue about the answer?
  • Practice or theory? Could we practice hard conversations the way we practice presentations?
  • Design or inheritance? Which of our team rituals were actually designed, and which did we just inherit?

References

Sources

Get the latest episodes directly in your inbox