The construct of collective intelligence has become increasingly popular in a variety of scientific domains, as it provides a theoretical and empirical framework that allows researchers to analyze how groups can perform better than individuals. Many studies have been conducted in a small-group environment where participants collectively generate a product that researchers can compare with an equivalent product resulting from individual actions or answers. Sharing a common methodology, the studies by Woolley et al. (2010), Engel et al. (2014), Meslec et al. (2016), Aggarwal and Woolley (2018), Aggarwal et al. (2019), Woolley and Aggarwal (2020) are a good example of the practical application of the premise of collective intelligence. Although those studies have nevertheless received a certain amount of criticism, researchers who have applied other methods have reached similar conclusions (Navajas et al., 2018; Toyokawa et al., 2019; Pescetelli et al., 2021).

This paradigm of dealing with collaborative group tasks in small groups featuring face-to-face interaction is fundamentally transformed when interactions are transferred to the Internet environment: a “digital space,” as designated by Pescetelli et al. (2021). One of the most important differences between face-to-face and online interaction is that the online environment allows researchers to considerably improve the number of individuals participating and collaborating in tasks. They can thus ascertain whether a large number of jointly interacting people might improve the chance of resolving the open, complex problems that confront humanity today (Lorenz et al., 2011; Hamada et al., 2020). The crowd intelligence construct provides us with a theoretical and empirical framework to approach such problem-solving processes on a collaborative basis in large groups; it is assumed, however, that the participants do not necessarily need to interact with one another to reach the solution (Hamada et al., 2020). One of the major research interests in this field is to ascertain under which conditions group performance can be optimized (Becker et al., 2017; Yahosseini and Moussaïd, 2020; Pescetelli et al., 2021). Key issues in such empirical analyses of the productions of large groups include: (1) ascertaining which is the most adequate method to aggregate potential solutions with the ultimate purpose of confirming that the collective product is better than the individual solution, and (2) studying how such products can be operationalized and quantified. One frequent solution is to devise estimation tasks that are easy to quantify; such tasks, however, seem to be far removed from the more complex, open kind of tasks with which human beings are usually faced in natural settings (Geifman and Raban, 2015).

Another key issue in the study of collaborative tasks is establishing a series of conditions that allow participants to interact with one another to achieve the best solution to a problem. Indeed, if overall performance needs to be improved, such collaborative interaction is a necessary condition (Gürçay et al., 2015; Navajas et al., 2018; Toyokawa and Gaissmaier, 2022), but it also serves as a differentiating factor among diverse types of collective intelligence processes (Hamada et al., 2020). As mentioned above, face-to-face interactions are limited to a reduced number of people, usually four to six individuals; in such contexts, the ideal number of participants in a group is five (Navajas et al., 2018). The number can be increased in some instances, as in particularly demanding tasks that require a more enhanced degree of coordination among members and a more significant amount of incentive (Mao et al., 2016). In the online environment, however, interactive spaces can gather elevated quantities of individuals who can interact with one another to analyze a problem or simply share their experiences regarding a particular product. Their outlook evolves naturally through interactions in a dynamic system (Almaatouq et al., 2020). The task of analyzing and understanding what goes on in such “digital spaces” is thus more complex.

This context thus poses two empirical questions. How does large-group interaction take place in such digital contexts? Is it truly possible to achieve quality solutions to particular problems via online interactions? Both of these issues ultimately lead us to analyze the nature of such interactions (Navajas et al., 2018; Massari et al., 2019) along with further key issues, including social influence processes (Gürçay et al., 2015; Becker et al., 2017; Toyokawa et al., 2019) and the participants’ level of involvement (Hamada et al., 2020). It is also important to define and establish the type of problem/task to be solved (Geifman and Raban, 2015; Sulik et al., 2022). An additional element worthy of study resides in the design of the task itself: the key issue is to ascertain whether the system foresees for the group to produce a consensual solution to the problem or whether the solution should be generated a posteriori through aggregation mechanisms that are external to the group (Yahosseini and Moussaïd, 2020).

A considerable number of complex issues are thus involved. Our study aims to analyze the interactions that arise when a large number of people (in our case, close to 800 individuals) jointly explore a task in an interactive environment on an online platform until they propose solutions. One key issue lies in resolving how the productions of a significant number of people should be quantified or measured, since their interactions can only be quantified as a series of thousands of individual events. We thus propose to study how those interactions evolve in the context of a large group that is dealing with an ethical problem consisting in the analysis of a case of cyberbullying. We chose and designed this case with the purpose of encouraging engagement on the part of the participants, who were asked to emit judgments about “what happened.” We propose to study the situation using big data tools designed to analyze emotions (Aggarwal, 2022; Trisna and Jie, 2022; Chan et al., 2023).

It is also essential to grasp the nature of the interactive environment in which a task is carried out. The IT environments in which large group interactions take place can have many different characteristics and goals, ranging from social media (on which information can be shared) to collaborative environments such as Wikipedia (Holtz et al., 2018). Certain tools are specifically designed to create collaborative environments in which information can be shared, joint solutions can be reached, and consensus mechanisms can be established (Geifman and Raban, 2015). This is also the case on the Collective Intelligence platform we used in this study: its characteristics are described further below and can also be consulted in Orejudo et al. (2022).

This study’s objective was to analyze how information spreads in an online interaction environment designed according to the principles of collective intelligence. We hypothesize that the emotional dimension plays a key role in this process and predicts how information spreads within that environment.

As can be observed in Table 1, participant activity in this experiment was not equally distributed along all its successive phases, nor was the response typology always equal. Thus, for the total of three questions, we found a response rate > 1 per participant in Phase 1 (the individual phase) since we had a total of 794 participants. That response rate decreased gradually and notably in Phases 2, 3, 4, and 5 (the phases of interaction with four neighbors: four real-time colleagues whose answers could be viewed by the participant). Then, in Phases 6 and 7, the response rate increases once again (these are the phases in which each participant is allowed to view the group’s most frequent responses). All three questions followed a similar pattern in this respect, although Questions 2 and 3 elicited more responses than Question 1 (see Table 1).

Those differences according to work phase can also be observed in the type of response emitted by the participants. The response length variable increased significantly in the answers appearing in Phases 6–7, and a similar increase can be observed in the transition from Phase 1 (individual work) to Phases 2, 3, 4, and 5 (work in tandem with one’s “neighbors”). The same pattern can be observed in the number of copies made of each response. Responses still present in Phases 6–7 were much more frequently copied than those that appeared in Phase 1, the individual phase. The latter were copied less than the responses appearing in collaborative phases 2, 3, 4, and 5. In terms of final frequency, we only found differences among responses appearing in Phases 6–7 and the remainder (Phases 1, 2, 3, 4, and 5). In all three questions, we observe considerable size effects according to the Eta-square value, and those size effects are associated with the considerable differences that can be noted between Phase 6–7 and the previous ones.

In Table 1, we can compare phases in terms of the program’s own Confidence in its evaluation of responses and the Polarity assigned to them. In the Confidence variable, we find the same pattern as in the ones mentioned above: Phase 1 (individual) generates the greatest degree of Confidence, followed by less Confidence in the 4-neighbor group phases (2, 3, 4, and 5) and even less Confidence in Phases 6–7. Although the response in Phases 6–7 is the one to which the least degree of Confidence can be ascribed according to the system, the mean still lies around 90 points for each of the three responses. Regarding Polarity, we note the following evolution of responses: a notable increment of the positive value from Phase 1 (0.13) to Phase 6–7 (1.34) and an increment of the negative value on Question 2 (−0.15 in the first phase to −1.62 in the last) and of the negative value on Question 3 (−0.62 in the first phase to −1.63 in the last).

The dichotomous variables of Subjectivity and Agreement also evolve notably along successive phases (Table 2). Differences are statistically significant in all three questions. In Phase 1, responses are more objective and have a greater degree of internal agreement, whereas in Phase 6–7 they are more subjective and with a lesser degree of internal agreement. That change occurs progressively in the intermediate phases (2, 3, 4, and 5), where participants work in tandem with their neighbors. Moreover, in Table 2, we can once more observe significant differences among the three questions in terms of these two variables (Subjectivity and Agreement). Ever since the first, individual phase, Question 1 is the one featuring the greatest degree of internal disagreement in the responses (39.7% for Question 1 vs. 33.3% for Question 2 and 34.2% for Question 3). Answers to Question 1 also display a greater degree of subjectivity (74% for Question 1 vs. 54.9% for Question 2 and 48% for Question 3). These differences among questions are maintained during the collaborative phases (2, 3, 4, and 5). However, by the time we reach Phase 6–7, they have become narrower between Questions 1 and 2, while Question 3 remains an outlier (Phase 6–7, Agreement, 72.9% for Question 1, 73.6 for Question 2, and 65.8% for Question 3; Subjectivity, 89.2% for Question 1, 84.2 for Question 2, and 71.0% for Question 3). In terms of irony, no notable differences can be observed, given that the majority of responses lack any sort of irony.

Table 3 displays the estimations of regression weights for the path models generated by each question. In all three cases, we were able to find models that adequately fit the data (P1, χ² = 16.494, d.f. = 6; p = 0.011; χ²/DF = 2.749; RMSEA = 0.028; CFI = 0.998; P2: χ² = 3.827; d.f. = 4; p = 0.430; χ²/DF = 0.957; RMSEA = 0.000; CFI = 1.00; P3: χ² = 3.968, d.f. = 2; p = 0.138; χ²/DF = 1.984; RMSEA = 0.016; CFI = 1.00), excluding regression weights that were not statistically significant (represented by zeros in this table). In all three cases, we obtained a good estimation of response length (41.7, 32.9, and 31.7%). In all three questions, the Confidence variable was the best predictor (γ_p1 = −0.323, γ_p2 = −0.434, γ_p3 = −0.264), followed by Agreement for Question 1 (γ = −0.263) and Subjectivity for Questions 2 (γ = −0.236) and 3 (γ = −0.244).

The prediction of the phase in which the final response appears is more efficacious for Question 1 (28.2%) than for Questions 2 and 3(10.7 and 6.7%, respectively). Among all predictors, Length is the most relevant one (γ_p1 = 0.473, γ_p2 = 0.246, γ_p3 = 0.173). Polarity also appears with a certain relevant weight in all three questions, but with a different sign in Question 1 than in the two others (γ_p1 = 0.123, γ_p2 = −0.110, γ_p3 = −0.084). The two remaining variables, Final Frequency and Number of Copies, are now much better explained for Question 1; moreover, Length and Phase are the two most relevant variables for all questions, as can be observed once again in Table 3. Nevertheless, exogenous variables are also statistically significant, especially Polarity and Confidence in Questions 2 and 3.

The overall relevance of each variable is shown in Table 4, which gathers the total set of effects of each variable within the model. Many of those effects are indirect, i.e., occurring via mediating variables. Thus, Confidence has an indirect effect on Questions 1 and 2 in the Phase variable (γ_p1 = −0.153, γ_p2 = −0.107, γ_p3 = −0.046), on Question 1 in the Final Frequency variable (γ_p1 = −0.219, γ_p2 = 0.021, γ_p3 = −0.015), and on the Number of Copies (γ_p1 = −0.205, γ_p2 = 0.013, γ_p3 = −0.014). We observe similar results for the remaining emotional variables, i.e., Subjectivity, Agreement, and Polarity, with indirect weights, particularly in Question 1.

This study’s goal was to analyze the evolution of responses generated by 794 participants who were asked to fulfill the task of analyzing a cyberbullying situation on a Collective Intelligence platform designed to encourage joint collaboration in the elaboration of responses aiming for consensus. The platform’s features are designed to maximize interaction elements that play a key role in the emergence of collective intelligence. as described in our Introduction section.

In the course of the process generated on the platform. The responses’ emotional dimension evolved with clear gradual differences between the initial individual phase (Phase 1). the middle phases of work in small groups (Phases 2, 3, 4, and 5) and the final phases where consensus was sought across the board (Phases 6 and 7). Apart from that relationship with the phases, we found a direct relationship with three further indicators of activity on the platform: (1) length of responses, (2) the final frequency with which certain responses appeared, and (3) the number of times they were copied throughout the system. Moreover, responses evolved differently along time according to the question. The role of the main characters in the ethical case – victim and aggressors – determined that evolution to a large extent.

Regarding these results, it is important to bear in mind that participants’ responses depend on the way the activity on the platform is organized. Phase 1 corresponds with individual work in which participants make their first contribution to the task set before them. In the next four phases (Phases 2, 3, 4, and 5), users share their responses with four “neighbors.” In each of those four phases, the neighbors are different, as well as the conditions under which the responses are displayed (the fixed answers at first, then real-time permutation and permutation of users). This design has the goal of ensuring that interaction among users should take place with a sufficient quantity of information that enables the participants in each group to analyze the responses. It also ensures that responses (thus information) can be distributed throughout the system in its entirety. In this process two aspects are key: the number of participants in each interaction “among neighbors” (five) and the system of permutation of users, designed to rotate user positions and contacts, thus enabling users to go on distributing the information that gradually emerges. This model guarantees interaction, which is key in the emergence of collective intelligence (Navajas et al., 2018). Moreover, the choice to reduce all interaction groups to only five users ensures that each user is confronted with a manageable amount of information. The herd effect is reduced, because the limitation to five neighbors encourages all users to participate actively. As we mentioned in the Introduction section, studies in the field of collective intelligence have found that reduced groups work best (Woolley et al., 2010; Navajas et al., 2018; Toyokawa et al., 2019; Pescetelli et al., 2021). Nevertheless. the design of the Collective Learning platform improves upon the model of local work in small groups, given that it manages to spread the information throughout the digital space, ensuring that all participants are allowed to collaborate in seeking the solution for the problem that has been posed: participant neighbor positions rotate and each one of them, in turn, acts as a new node in a series of successive interactions. This strategy optimizes the potential for interaction within the digital space (Almaatouq et al., 2020; Yahosseini and Moussaïd, 2020; Pescetelli et al., 2021).

The results in these small-group phases (Phases 2 to 5) yield similar response patterns for the three questions that made up the task. Compared to Phase 1, the number of responses in Phases 2, 3, 4, and 5 was reduced, thereby implying that the participants were indeed carrying out the task: they were analyzing the other participants’ answers. The fruit of their analysis, in turn, generated new solutions. In the course of Phases 2, 3, 4, and 5, we thus observe that chosen and emergent responses gradually started to lengthen (as shown by the Length variable). Moreover, they gradually tended to be copied more times (i.e., the averages of the number of copies and the final frequency are directly related to the phase in which the responses appear). The emotional polarity of responses also tends to vary, curiously, we ascertain that in Phase 1, polarity is positive; it turns negative in Phases 2 and 3, and also starts to increment. These results might indicate that the platform is achieving the objective we set ourselves: the aim of facilitating an environment that guarantees the spread of information across the entire group. This seems to derive from the direct relationship between the phase in which the response is elaborated and its greater diffusion across the network, observed in the Number of Copies and the Final Frequency, as well as in the fact that the number of neighbors we chose for this experiment (groups of 5) guaranteed a diversity of responses that was sufficient to ensure and encourage response evolution. The conditions we chose also seem to have allowed for an adequate amount of time for the group to develop: the group, as a unit of synchronized analysis, requires such a minimum amount of time to confront ideas and positions through interaction and interchange in order to achieve a maximum degree of cooperation and interdependency in the final consensus. The platform supports and encourages the collective construction of a solution from a perspective of co-responsibility (Palacín and Aiger, 2014; Sierra-Pérez et al., 2021).

Phases 6 and 7 are the phases of consensus and maximum interdependence. To achieve that state, the platform calculates in real time the frequency of each response within the system, and it progressively presents the participants with the “Top 10” answers (the most frequent ones). At the same time, it generates an opposite effect by eliminating the least frequent responses. Users who had chosen less popular answers now have the option of deciding whether they develop a new response of their own pen (Phase 6) or if they decide to adopt one of the Top 10 answers as their own. The mechanism is designed to inspire users to seek consensus: among other methods, it represents an alternative designed to reduce the diversity of responses observed elsewhere and the conditions that lead to such diversity (De Vincenzo et al., 2017; Mann and Helbing, 2017; Massari et al., 2019; Sulik et al., 2022). However, compared to other mechanisms of social influence, this system does not start to function until the responses are elaborated and distributed throughout the network, thereby limiting the initial effect of popularity (Lorenz et al., 2011; Bernstein et al., 2018). The nature of responses is an indication of the importance of structuring the group’s activity by imposing a series of successive phases in time (individual phase, interactive phases, then group phase) in order to encourage the group’s emergent maturity and the synchronicity of its responses, achieved via group consensus. Interaction enhances a user network’s potential for “listening,” for generating discrepancies. For harmonizing points of view, and for strengthening agreements (Palacín and Aiger, 2014; Sierra-Pérez et al., 2021). From an emotional perspective, we observe that the two last phases produce a response pattern that is clearly different from the previous ones. The adopted responses are now even longer and have an elevated degree of final frequency within the system. Of course, the system’s mechanism already tended to encourage such responses; now, however, their emotional dimensions are even more pronounced than in the previous ones; their positive or negative polarity is likewise more extreme than before. In this final phase, the system has fulfilled its purpose: that of reducing the number of responses and proposing a consensus solution. It could be argued. However, that this process has led to a response that is somewhat extreme from an emotional point of view, reflecting a certain “polarization effect” that can easily emerge when small groups interact with the same task (Navajas et al., 2018). However, in a task/situation such as the one proposed herein, the emotional tendency can be viewed as something more positive than negative: it favors a positioning close to the victim’s interests, rejecting those of the aggressors (Machackova and Pfetsch, 2016). It would thus seem that the dynamic generated by the system has exerted a homogenizing effect while producing a response that is formulated with a greater degree of complexity or richness in terms of argumentation. This can be seen in the greater length of favored responses, but also in further indicators stemming from the sentiment analysis: the degree of Confidence in estimations. The responses’ internal agreement among themselves, and the degree of subjectivity.

These indicators, on the whole, present a profile that seems to indicate that a response of greater complexity has been achieved, thereby confirming that a process that aims to generate collective intelligence should produce quality solutions. Moreover, the process itself should also offer a learning experience for those who participate in it (Bernstein et al., 2018; Orejudo et al., 2022). The data resulting from this study gave us a series of clues regarding such complexity from an emotional vantage point. Thus, on the one hand. we have found that the length of responses increases along the successive phases: in the case of all three questions, the path model directly relates that aspect of length with the four emotional dimensions under consideration. Which. taken as a whole, have the capacity to explain a remarkable percentage of responses: 37.9% (for Question 1), 33.1% (for Question 2), and 31.4% (for Question 3).

Moreover, four of the emotional dimensions (except irony) provided by the sentiment analysis tool have significant relationships among themselves and vary in the same manner from one phase to the next. This, in turn, supports our hypothesis that complexity gradually emerges in the course of the system’s successive interaction phases. Emotional valence is increasingly clearly defined as either positive or negative. Two further indicators, Subjectivity and Disagreement, are related, on the one hand, to the responses’ emotional content and, on the other, to the internal structure of the texts that make up those responses. Those three indicators (Emotional Valence, Subjectivity, and Disagreement) are joined by Confidence, which is based on the reliability of the system’s categorization, and which might be associated with less complexity. At any rate, our hypothesis that emotional dimensions can play a relevant role in the interactions that occur on this collective intelligence platform is confirmed by (1) the pattern of responses produced through the experiment, (2) their direct relationship with the Phase variable, (3) the latter’s direct relationship with Length, and (4) the relationship between Length and Final Frequency and Number of Copies. Our data analysis also confirms that the sentiment analysis tool is adequate for the analysis of large quantities of data such as those generated here, or even those generated by interactions on social networks (Karagozoglu and Fabozzi, 2017; Ferchaud et al., 2018; Lozano-Blasco et al., 2021; Zuheros et al., 2023).

The different Polarity valences – positive for Question 1 but negative for Questions 2 and 3 – likewise seem to suggest that these participants have gone through the real process of analyzing the task, as reflected by the successive responses the interaction system has generated. They respond as if they had authentically witnessed the occurrences, achieving a great degree of emotional implication with the actors in the ethical case and also experiencing a certain degree of emotional activation (Caravita et al., 2016). Participants responded by choosing the side of the victim, a choice which generates higher levels of cognitive and emotional empathy with the victim than with the aggressors, as also tends to occur in real-life scenarios (Machackova and Pfetsch, 2016). The group dynamic further accentuated the empathetic response, from a cognitive perspective. Bautista et al. (2022) had already ascertained that the same task and situation tended to improve participants’ moral reasoning capacities. Social neuroscience studies have used group somatic markers (measured by group electrodermal activity) to show that (1) a group needs a certain amount of time to synchronize the act of “listening to one another,” (2) group processes imply improved attention levels in the individual, and (3) emotional activation somewhat attenuates a group’s overall listening capacity (Aiger, 2013). The rate of change tends to accelerate when the group is able to carry out a process of cognitive-emotional revaluation (Vicente et al., 2006) in the phases of cooperation and maximum interdependence (which, in our study, had their equivalent in the final consensus phases. i.e., Phases 6 and 7).

Our results lead us to reflect upon the educational value that might be inherent in contexts of interaction, such as the environment provided by our experiment. Other authors in the field already find a great degree of educational potential in interactive large-group contexts. Large groups allow for their members to construct authentic learning communities; participants can pool and share their resources and they can also create collaborative/cooperative spaces such as the one described herein (Castellanos et al., 2017; Chen, 2018; Ciccone, 2019; Tenorio et al., 2021). Moreover, since such interactive projects take place in “virtual” spaces, the environment is familiar to adolescent users and akin to the type of online space where they are often confronted with threatening risks, such as cyberbullying. Thus, within a virtual space, adolescents can start acquiring the necessary competencies that help them adapt to online environments in general, such as messaging platforms and social networks (Cebollero-Salinas et al., 2022). Such collective intelligence makes it possible to gather a sizeable “crowd” to resolve relevant issues. The system presented in this paper is capable of managing up to 5.000 participants simultaneously in one project. The user can thus feel that they are part of a learning community, or of a collective that takes up joint action when faced with the same problems. This idea of a collective could be a further variable that plays a certain role in the productivity of groups facing a task, such as the group of users featured in this study. An additional advantage found in environments such as this one can be seen in the occasion that it offers for real-time interaction. By adding a fixed time limit, such virtual platforms can provide a consensus mechanism that delivers certain guarantees. We are aware that such group phenomena do not always guarantee success; it is thus essential to describe under which conditions collective intelligence is capable of producing quality solutions (Bigham et al., 2018; Orejudo et al., 2022). The current investigation provides a solid basis for the ongoing study of complexity in collective intelligence. Moreover, certain features of the platform – such as the possibility of modulating the rate of interchange in the small-group work phases, or of choosing the number of responses eliminated in the final consensus rounds – can give rise to different scenarios which, in turn, offer additional alternatives that could be of educational use. If the platform presented herein were programmed to act as a knowledge management system by modifying structures and group processes plus introducing the use of feedback (which, in turn, affects interventions) and by incorporating the aspect of social abilities, it could be reckoned among the leading alternatives in terms of group management designed to efficiently conduct tasks while ensuring that a group can serve as an instrument of change, thanks to the advantages of CI (Collective Intelligence).

However, we should still ask ourselves whether the final consensus phases (Phases 6 and 7) represent the type of interaction that allows users to have a learning experience. From the vantage point of our research, we cannot guarantee it, since social influence seems to be the major factor in these two final stages. But at least the participants are also engaged in social learning, as they are selecting the options they consider best among all the options they had previously been dealing with (Navajas et al., 2018). Other crowd intelligence studies assume that learning is acquired by merely sharing such information, analyzing it and evaluating it (Toyokawa et al., 2014; Hertz et al., 2021); we thus likewise assume that information is being actively processed and that such processing requires a certain amount of previous activity in formulating a response to the question. This, in turn, involves decision-making. Which brings us to the vicinity of active learning theories.

The raw data supporting the conclusions of this article will be made available by the authors, without undue reservation.

The studies involving humans were approved by Research Ethics Committee of the Autonomous Community of Aragon (CEICA). C.P. – C.I. PI22/216. The studies were conducted in accordance with the local legislation and institutional requirements. Written informed consent for participation in this study was provided by the participants’ legal guardians/next of kin.

SO: Conceptualization, Data curation, Formal analysis, Funding acquisition, Investigation, Methodology, Supervision, Validation, Writing – original draft, Writing – review & editing. RL-B: Conceptualization, Data curation, Formal analysis, Investigation, Methodology, Software, Validation, Visualization, Writing – original draft, Writing – review & editing. PB: Conceptualization, Data curation, Formal analysis, Investigation, Methodology, Resources, Software, Validation, Visualization, Writing – original draft, Writing – review & editing. MA: Conceptualization, Formal analysis, Investigation, Methodology, Supervision, Validation, Visualization, Writing – original draft, Writing – review & editing.