[Introduction to Unconventional Collective Intelligence] Part 1: Powering Up by 'Not' Thinking Separately!

Thinking separately is certainly important.

When we hear the term 'collective intelligence,' the first scene that might come to mind is one where 'many people think separately from one another.'

Each person thinks with their own mind.
They are not swayed by the answers of others.
On top of that, the judgments of the majority are averaged or a majority vote is taken.

As a result, individual errors cancel each other out, and the group as a whole arrives at a fairly accurate judgment. This is the basic image of what is known as collective intelligence.

In research terms, what is important here is independence.
The more independent people's judgments are from one another, the easier it is for their respective errors to offset each other. Conversely, if everyone looks at others' opinions and leans in the same direction, the risk increases that the group as a whole will make the same mistake.

The intuition of 'separateness' that also resonates with Rousseau and Hiroki Azuma

This intuition that 'it is important to think separately' is not just about estimation tasks. As I will discuss again later, there is a similar awareness of the issue in Rousseau's theory of the general will and in Hiroki Azuma's 'General Will 2.0'.

For example, 'General Will 2.0' can be read as a book that re-examines the premise that democracy requires deliberation, and envisions whether the 'atmosphere' (kuki) can be technically visualized and used as a basis for consensus building.

In other words, when people read the atmosphere, form factions, and are swayed by special interests, the judgment of the group as a whole becomes distorted. Therefore, it is necessary to first extract individual judgments and desires as separately as possible. It is quite natural to want to think this way.

Research 1: Social influence can destroy collective intelligence

In fact, there is research that supports this intuition.

Lorenz et al. (2011) investigated how seeing others' estimates affects collective intelligence in a quantitative estimation task. Participants first make estimates on their own. Then, they repeat the process of answering the same question again after seeing the estimates of other participants.

As a result, when people saw others' estimates, their answers gradually converged. At first glance, it looks like they are learning by referring to others' information. However, the problem is that as the answers become similar, the diversity within the group is lost.

Collective intelligence is strong because everyone makes mistakes in slightly different directions.
However, when you see others' answers, even the way you make mistakes becomes similar.
As a result, errors become harder to cancel out.

Therefore, the claim that 'collective intelligence requires independence' is correct.

However, 'separateness alone' is not the answer

But if we stop here, another misunderstanding arises.

Unless people think separately, collective intelligence cannot be born.

This is only half true.
Independence is important. However, simply collecting the opinions of people who thought independently is not always the best approach. What recent research shows is actually beyond that.

A well-designed group's wisdom can sometimes produce better judgments than an independent crowd.

Research 2: Creating a Two-Layer Group Structure

A representative example of this is the research by Navajas et al. (2018).

What kind of structure is it?

The key is to make the group a two-layer structure.

In the first layer, participants are divided into several small subgroups. Within each subgroup, members discuss and create a single consensus answer.

In the second layer, the consensus answers produced by each subgroup are aggregated. Here, the subgroups do not discuss with each other directly; instead, the conclusions of each group are further averaged.

In short, the flow is as follows.

Individual independent judgment -> Discussion within subgroups -> Consensus per subgroup -> Aggregation of consensus answers

What is interesting here is that they did not put everyone in one large conference room to discuss. First, individuals think independently. Then, they discuss in small subgroups. And finally, the consensus of each subgroup is further aggregated.

In other words, they did not completely abandon independence; they
hierarchized the group.

Why a two-layer structure works

This two-layer structure has advantages.

Within a subgroup, an individual's extreme estimates or misunderstandings can be corrected. If someone provides a value that is clearly too high, others can say, 'Isn't that a bit too high?' Conversely, another person might provide a clue that one person alone would not have thought of.

However, if you rely on only one subgroup, the biases inherent to that group remain. Therefore, you further average the consensus of multiple subgroups. This allows you to reduce individual errors through group discussion while canceling out errors through differences between groups.

This is a quite important concept.

What is necessary for collective intelligence is not simply 'not discussing.' Rather, it is
when to think independently, when to discuss, in what units to create consensus, and how to aggregate them.

Research 3: Whose opinion reaches whom? ~Reconfiguring the network~

The structure of the network connecting the group is also important.

Becker et al. (2017) investigated a situation where people update their judgments while observing the estimates of others on a network. The issue here is whose opinion reaches whom.

If influence is concentrated in a few central figures (centralized), the entire group might move in a good direction if that person is correct. However, if that person is wrong, the entire group will be pulled along by that error.

On the other hand, in a distributed (decentralized) network, influence is less likely to be concentrated in a few people. Information is exchanged locally, extreme opinions are smoothed out, and as a whole, estimates can be improved.

The conclusion here is the same.

Interaction is not the problem. The structure of bad interaction destroys collective intelligence.

Study 4: Appropriate social information corrects bias

Research by Giles et al. (2017) reinforces this point.
Humans have systematic biases, such as overestimation or underestimation. In such cases, when 'appropriate social information is introduced,' that bias can sometimes be corrected.

In the experiment, 'virtual participants' who made estimates close to the correct answer were mixed in without the participants' knowledge.
The results showed that individual estimates improved and the accuracy of the group as a whole could also be enhanced.

Of course, if you introduce too much social information, conformity occurs.
However, if you don't introduce any at all, the errors held by individuals remain as they are.

What is important is the quantity, quality, and timing of the social information.
Just like medicine, whether it works depends on the dosage and how it is used.

Even with collective intelligence of preferences, how it is presented is an issue

This discussion also connects to the topic of 'collective intelligence of preferences,' such as reviews and rankings.

In Salganik et al.'s (2006) music market experiment, when the number of song downloads was visible, popularity begat popularity, increasing inequality in hits (a large gap between popular and unpopular songs) and unpredictability (it was difficult to predict in advance which songs would become popular).
Research by Muchnik, Aral, and Taylor (2013) also shows that just having an initial positive rating on an online comment can cause subsequent ratings to be excessively inflated.

This result can be understood as a case where seeing others' evaluations—in other words, a type of interaction—distorted collective intelligence.

However, even here, we should not end with 'don't show others' evaluations.'

Study 5: Changing the algorithm that displays group opinions

Rahman et al. (2014) demonstrated the possibility of reducing distortion caused by popularity through the design of display order.

When sorted by popularity, items that are already evaluated are seen more, making them even easier to evaluate. In other words, 'everyone's opinion' is easily distorted.

Therefore, instead of putting items with the highest total number of evaluations at the top, they considered a method of displaying the most recently evaluated items at the top.

This is an idea of designing the display so that popularity does not become too fixed, rather than erasing popularity information.
Even with the same social influence, if you change how it is shown, the way the group explores changes.

Conclusion: The problem is not 'whether or not to think separately'

Ultimately, the problem is not 'whether to think separately or not.'

Thinking separately is important.
However, there is information that cannot be captured just by thinking separately.
There are clues that one cannot notice alone.
Information from others can also correct individual biases.

Therefore, the new development of collective intelligence can be stated as follows.

An independent crowd is strong. However, well-designed collective intelligence can sometimes surpass it.

Collective intelligence does not necessarily mean that keeping everyone separate is the best approach.
On the other hand, it is not just about letting them talk, either.

The implications obtained from the discussion so far are as follows.
First, think independently.
Next, interact in small groups.
Structure the group into two layers.
Consider the shape of the network.
Design which information to display.
And finally, aggregate effectively.

In other words, better collective intelligence is neither mere silence nor unprincipled chatter.

It is designing the group itself.

References

• Lorenz, J., Rauhut, H., Schweitzer, F., & Helbing, D. (2011). How social influence can undermine the wisdom of crowd effect. Proceedings of the National Academy of Sciences, 108, 9020–9025.

• Navajas, J., Niella, T., Garbulsky, G., Bahrami, B., & Sigman, M. (2018). Aggregated knowledge from a small number of debates outperforms the wisdom of large crowds. Nature Human Behaviour, 2, 126–132.

• Becker, J., Brackbill, D., & Centola, D. (2017). Network dynamics of social influence in the wisdom of crowds. Proceedings of the National Academy of Sciences, 114, E5070–E5076.

• Jayles, B., et al. (2017). How social information can improve estimation accuracy in human groups. Proceedings of the National Academy of Sciences, 114, 12620–12625.

• Salganik, M. J., Dodds, P. S., & Watts, D. J. (2006). Experimental study of inequality and unpredictability in an artificial cultural market. Science, 311, 854–856.

• Muchnik, L., Aral, S., & Taylor, S. J. (2013). Social influence bias: A randomized experiment. Science, 341, 647–651.

• Lerman, K., & Hogg, T. (2014). Leveraging position bias to improve peer recommendation. PLOS ONE, 9, e98914.

• Hiroki Azuma (2011/2015) 'General Will 2.0: Rousseau, Freud, Google', Kodansha / Kodansha Bunko.

Latest Research

• Almaatouq, A., Noriega-Campero, A., Alotaibi, A., Krafft, P. M., Moussaïd, M., & Pentland, A. (2020). Adaptive social networks promote the wisdom of crowds. Proceedings of the National Academy of Sciences, 117, 11379–11386.
  A study dealing with networks where people can adaptively change their connections rather than fixed networks. It shows that with feedback and network plasticity, groups can adapt to their information environment and produce collective estimates more accurate than the best individual.

• Almaatouq, A., Rahimian, M. A., Burton, J. W., & Alhajri, A. (2022). The distribution of initial estimates moderates the effect of social influence on the wisdom of the crowd. Scientific Reports, 12, 16546.
  A study showing that whether social influence helps or hinders collective intelligence is not actually determined solely by network structure. It reports that the distribution of people's initial estimates is what is important.

• Burton, J. W., Almaatouq, A., Rahimian, M. A., & Hahn, U. (2024). Algorithmically mediating communication to enhance collective decision-making in online social networks. Collective Intelligence, 3(2).
  A study on algorithmically rearranging who communicates with whom on online social networks. This research moves in the direction of designing communication structures rather than leaving interactions to occur spontaneously.