When less (communication) is more (collective intelligence)
Team work, complex problems, and preserving diversity in the ecosystem of ideas
Communication is key to teamwork, but what if, when faced with complex problems, the most effective teams might need to speak a little less? That’s the surprising conclusion from research I’m looking at in today’s newsletter.
There are trade-offs in working together: a group can do more than one person, but as more people join a group there are also added costs. There are transaction costs, for example: larger groups require more overhead in terms of coordination. Another cost is “social loafing”, where individuals put in less effort when they are in a group.
For simple physical tasks - like digging a ditch maybe, or pushing a car - these costs of the group are typically less than the gains from having more people on the job. For intellectual tasks, the cost-benefit equation shifts.
One element of this is that for some, but not all, intellectual tasks the measure of success is the best solution anyone comes up with, not the sheer number of solutions. Problem solving tasks, from puzzles to some of the grand challenges in science and maths, have this nature. More people working on the task gives you more chances of success, but ultimately the success of the group is no more than the success of the single best solution. Unlike pushing a car, it isn’t clear that every little bit helps
As well as altering the payoff, the costs of group working for intellectual tasks are also altered. The ordinary costs of group work may still apply - transaction costs and social loafing - but there are now different dynamics which arise from having a group working on the task.
More people mean more exploration of the space of possible solutions, with people quickly exploring in parallel options which might take a single individual a long time to cycle through on their own. People with different solutions can inspire each other, and promising angles propagate through the group.
Social influence isn’t all good, though. Groups are famous for herding behaviour, conformity or “groupthink”, in which a single dominant view suppresses other options.
Experimenting with more and less communication
A neat study from 2018 nicely illustrates one way a group can benefit from bringing more people to work on problem, while avoiding the perils of too much groupthink.
Bernstein, Shore & Lazer (2018) use a neat puzzle called the Travelling Salesman Problem (TSP) to study exploration and copying in problem solving. In the TSP you have to draw a line between dots - representing the stops of the salesman - to make the shortest route connecting through each dot and arriving back at the beginning.
There’s a whole mathematics of this class of problem, which is interestingly complex. Two features are that 1) better answers are hard to find, but easy to verify - you can see immediately if you’ve found a shorter route, and that 2) you can’t always improve an answer by making a small change. The rule about not revisiting any node means that if you change one connection you have to change at least one other to provide a complete route. This second feature means that you can’t find the best solution by incremental hill-climbing - tweaking your existing answer to slightly improve it won’t work, since the best solution might look nothing like the second best solution, and anyone with the second best solution would need a radical overhaul of their approach to get to the best solution.
Participants in the 2018 experiment were given 17 attempts, each lasting 50 seconds, to find the best solution they could to particular TSP (a particular arrangement of dots in the 2D space).
In our own work we’ve looked at how small groups can use the exchange of arguments to generate insights into the best solutions to complex problems. Bernstein and colleagues simplify things by limiting communication to the exchange of solutions - in fact, the only communication is when participants are shown the current solutions of other group members (3 participants per group). No arguing, just information transmission.
This mechanism let the researchers compare fully connected groups (CT, “constant ties”) with a control condition where people are nominally in groups, but they didn’t see each others’ solutions (NT, “no ties”). A final, most interesting, case they tested was intermittently connected groups (IT, “intermittent ties”), where participants only saw other group members’ solutions every third round.
Their results tell us something about collective intelligence, and also suggest how it can be enhanced.
In the no communication, NT, groups someone found the optimal solution in 44.1% of trials, compared to the 33.3% in the full communication, CT, groups. That’s right, the group which didn’t communicate did best. The cost to this was the average performance. Having group members who independently explored the space of possible solutions meant that these groups had a better chance of containing someone who hit upon the optimal solution. The downside though, is that group members get no benefit from each other - the average performance stays low even if one person finds the right answer.
The full communication group had the mirror image problem - they were less likely to find the optimum solution, although when they did they were able to share it. However, because the group members were aware of each other’s answers they tended to under-explore the solution space, so didn’t get lucky as often. When they did get lucky though, all members could benefit and so the average across the group was better.
The insight from this experiment comes from the intermittent communication, IT, group, where group members saw each others’ solutions every third round. This arrangement seemed to be a sweet spot - providing enough independence that group members did exploration on their own, but also copied from each other when a good strategy was found.
Diversity as a resource for collective intelligence
Key to understanding this is a reversal of the usual story that a group is pulled up by its “leaders”. Rather, the success of the group member with the best solution can be understood as a function of their ability to draw upon the rest of the group as a resource. When the group intermittently communicates it allows the group to explore effectively, and also to share among itself the results of that exploration. The success of the best performing member is supported because they have a large range of solutions to draw upon among the rest of the group. The constant communication groups are too successful at copying each other, meaning that they don’t explore and so the solution resource they provide for each other is diminished (and, conversely, the no communication groups have the richest resource for potential solutions among group members, but their complete lack of communication means they can’t learn from each other).
You can see this in Figure 3 from the paper. For each problem the researchers looked at the proportion of each answer matching the optimal solution. Even bad answers might have contained sections which were optimal, waiting to be spotted by someone in the group and incorporated into their current best answer. The horizontal x-axis shows the proportion of the optimal solution in the best answer from the group. The vertical y-axis shows the proportion of the optimal solution that exists in the solutions by group members who *don’t* currently have the best answer, across the 17 rounds in the trial.
You can see that through successive rounds the constant communication (CT) groups see a rapid and constant decline in the proportion of the optimal solution that is preserved in the other group members’ answers (the red line drops down and stays down). The leading group member gets closer to the full solution, but they become hampered because - on average - as they approach the best solution they don’t have good ideas to copy from. The no communication, NT, groups have a high distribution of good ideas, which is preserved across the 17 rounds (the blue line stays high), and the intermittent communication group manage to combine improvement over successive rounds while maintaining a diversity of ideas to copy (the green line stays higher than the red, even as the group’s best performing member improves their solution). For the intermittent communication group this means that eventually, on average, the best solution in this group is better than in the constant communication groups.
400 years of diversity and exploration in the Go community
There’s a sympathy to this result in a new analysis from Bret Beheim, who looked at 400 years of moves in the strategy game Go. This allowed him to draw some interesting conclusions about how the community of players has explored the space of possible moves, and how the cultural evolution of Go strategy has been affected in the AI era.
By tracking who played who, and what moves were played in those games, Beheim can plot the number and size of different Go playing communities against the diversity of opening moves tried in those games:
There’s a rise and fall of diversity in historical time. At first, as Go became more popular and spread to more countries the diversity of opening moves increased. More players meant more exploration of the possible playing space. But then, with globalisation, then online Go, and then superhuman Go playing AI engines, that diversity collapsed. Although there are more players than ever, they play each other in larger communities and play a smaller diversity of opening moves.
Quoting the paper:
We can also see how the population structure of Go communities sustains behavioural diversity. …Networks that are completely disconnected prevent the spread of ideas altogether, but as people become more connected, moves can more quickly become extinct as the population too quickly converges on consensus or ‘canonical’ strategies. Partially fragmented, small world networks thus exist in a Goldilocks zone between overly connected and overly disconnected
Throttling consensus as a general principle of collective intelligence
From empirical results like these, Paul Smaldino and colleagues attempt to extract the general lesson: Maintaining transient diversity is a general principle for improving collective problem solving. They say that mechanisms which delay consensus forming can help collective intelligence, because - as we saw for the TSP problem - they allow greater diversity in the solutions being considered by the group members, and this diversity of solutions is an important shared resource for eventually coming to better solutions. Whether it is something in the structure of communication, or psychological mechanisms (like independence, or pure stubbornness), mechanisms which delay consensus can actually improve outcomes, albeit at the cost of generally taking up more time.
The paper as a table of different mechanisms, showing that everything from starting from more diverse individual positions, to sticking to your guns (“behavioural inertia”), to slow or intermittent interactions can all achieve the same result, delaying consensus forming and so improve collective intelligence.
There’s an old adage: If you want to go fast, go alone. If you want to go far, go together. With these results, we can now see that analogy holds for cognitive work as well as travel. When we are travelling in physical space we need each other for mutual support. When we travel in intellectual spaces the diversity of our ideas is a resource which the group diminishes too quickly to its own cost.
This newsletter is free for everyone to read and always will be. To support this you can upgrade to a paid subscription (more on why here)
Keep reading for references and further reading on inoculation theory, using AI intelligently, and epistemic infrastructure.
References:
Bernstein, E., Shore, J., & Lazer, D. (2018). How intermittent breaks in interaction improve collective intelligence. Proceedings of the National Academy of Sciences, 115(35), 8734-8739. https://doi.org/10.1073/pnas.1802407115
Beheim, B. (2025) Opening strategies in the Game of Go from feudalism to superhuman AI Evolutionary Human Sciences 7(e28) https://doi.org/10.1017/ehs.2025.10016
Smaldino, P. E., Moser, C., Perez Velilla, A., & Werling, M. (2024). Maintaining transient diversity is a general principle for improving collective problem solving. Perspectives on Psychological Science, 19(2), 454-464. https://doi.org/10.1177/17456916231180100
PAPER: Promoting engagement with social fact-checks online: Investigating the roles of social connection and shared partisanship
Bots which engaged with users before offering a correction (following them and liking three tweets of theirs) got more engagement with the fact-check.
Here’s figure from the paper showing what the engagement and correction looked like (anti and pro Trump conditions left and right):
The experiment reminds me of Munger’s classic ‘Tweetment Effects on the Tweeted: Experimentally Reducing Racist Harassment’. I think the wrong lesson to take from both is that we should get the bots to pretend to like people before correcting them. Rather, we need to find a way to use technology to support fact-checking by those who are already socially connected to spreaders of misinformation.
Reference: Martel, C., Mosleh, M., Eckles, D., & Rand, D. G. (2025). Promoting engagement with social fact-checks online: Investigating the roles of social connection and shared partisanship. Plos one, 20(3), e0319336. https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0319336
PAPER: Community Notes are Vulnerable to Rater Bias and Manipulation
Previously we wrote about Threats to the sustainability of Community Notes on X (my newsletter on this here). One of the issues is that most Notes written are never published. Now, a new analysis shows that a small number of bad actors in the Community Notes system can effectively veto note publication, exacerbating the problem of getting Notes shown on misleading tweets - not just generally suppressing the appearance of Notes, but also introducing a bias (because we assume they coordinate their bad faith actions in a partisan direction). Doesn’t prove it does happen, of course, but the thing about bad actors is that if a vulnerability exists they will tend to find and exploit it
These findings suggest that while community-driven moderation may offer scalability, its vulnerability to bias and manipulation raises concerns about reliability and trustworthiness, highlighting the need for improved mechanisms to safeguard the integrity of crowdsourced fact-checking.
Link: Community Notes are Vulnerable to Rater Bias and Manipulation https://arxiv.org/abs/2511.02615
Our work on Community Notes: Arjmandi-Lari, Z., Mantzarlis, A., Stafford, T. (2025). Threats to the sustainability of Community Notes on X. arXiv preprint arXiv:2510.00650
PAPER: What did Elon change? A comprehensive analysis of Grokipedia
This report uses automated tools to do a side-by-side comparison of Wikipedia and Elon Musk’s derivative, and machine generated, “politically alternative” Grokipedia. Overall the analysis shows how derivative of Wikipedia Grokipedia is. There is higher similarity for non-controversial article topics (e.g. “Mejia Thermal Power Station”) and the selection of sources cited is very similar. For controversial topics (e.g. “Racism in the United States”) the similarity to Wikipedia is lower, and the source analysis shows that Grokipedia cites more social media sources, including both a large number of Elon Musk’s tweets and the output of Grok on X (which strikes me as incestuous).
What did Elon change? A comprehensive analysis of Grokipedia
TIP for using LLMs - see how they misunderstand you
I like this from Mike Caulfield, get an LLM to expand on your bullet points and see how they can be misunderstood, then use those errors to refine the thing you’re trying to unambiguously express.
Link: Oh, Lord, Don’t Let Me Be Misunderstood
…And finally
By Graham Annable, aka Grickle
END
Comments? Feedback? But only if you are sure you aren’t jeapardising optimal levels of transient diversity. I am tom@idiolect.org.uk and on Mastodon at @tomstafford@mastodon.online