Collective intelligence and swarms at work: concepts and evaluation
Image generated by the Inferama Editorial Team
01

Four different approaches that should not be confused

The phrase “swarm intelligence” can refer to very different activities. A group of people coordinating their responses in real time does not work like an algorithm that moves candidate solutions through a search space. Nor is it equivalent to a generative AI system that produces text, images, or other content. The mechanism and objective determine which approach is being described.

Collective intelligence studies how people, computers, or both can act together to achieve a goal. In a business, it may involve pooling knowledge, coordinating work, and using tools to gather or process contributions. It does not require every participant to respond at the same time, nor does it require an animal-inspired algorithm.

A human swarm is a connected group that contributes responses or decisions in a coordinated way, often simultaneously. Unanimous AI described a service called UNU in 2015 as an application for forming such groups online and in real time. That historical announcement illustrates one form of interaction; it does not confirm that the service is currently available or, by itself, prove that the method is effective.

Particle swarm optimization, or PSO, is a computational search technique. Each “particle” represents a candidate solution and moves based on information about its own path and those of other solutions. The metaphor of a group searching for treasure can help explain the mechanism, but the particles are values or configurations within a calculation, not people.

Generative AI is another category: models that emulate the structure and characteristics of input data to produce derived synthetic content. It can be incorporated into a collaborative process to suggest ideas, but generating options is not the same as deliberating, reaching consensus, verifying facts, or finding an optimal solution. Generating content and coordinating collective intelligence are distinct functions, even when a product combines them.

What each approach does

ApproachWhat is coordinated or transformedIntended result
Collective intelligenceContributions from people, computers, or bothJoint action directed toward a goal
Human swarmResponses or decisions from connected participantsA group response or decision through coordination
PSOCandidate solutions within a search spaceA solution that optimizes a defined function
Generative AIPatterns learned from input dataSynthetic content, such as text or images
02

Collective intelligence: organizing contributions toward a goal

A practical definition of intelligence is the ability to achieve a goal. Applied to groups, collective intelligence refers to people, computers, or both acting together in a way that seems intelligent. The unit of analysis is the system formed by the participants and the rules, tools, and channels that organize their contributions.

In an organization, that system might be as simple as a team combining perspectives to solve a problem, or as structured as a platform that distributes tasks and aggregates results. **Technology can facilitate coordination, but it does not automatically make every interaction effective collective intelligence.** The quality of contributions, the task, the way responses are combined, and the incentives all matter.

A group process can also fail in ways that an individual response does not. Participants may share the same incomplete information, give too much weight to the majority view, or stop contributing their own information once the group appears to have decided. The number of participants alone is therefore not enough to evaluate a method: what matters is whether their contributions add relevant information and how they affect one another.

Projects at the MIT Center for Collective Intelligence span scientific models, empirical studies, and the design of collectively intelligent systems. That agenda shows that the concept covers both studying group behavior and creating tools. It does not mean that a particular solution has demonstrated a general improvement across all businesses.

When assessing a business application, define the objective before choosing a technique. If the problem is to gather distributed knowledge, a process for obtaining and combining contributions may be worth testing. If it requires searching for parameters under constraints, an optimization method may be appropriate. If initial proposals are needed, generative AI can supply options for people to evaluate.

From objective to method

  1. 01Describe the decision or task to be resolved, not just the tool you want to test.
  2. 02Identify what information people have, what data the system supplies, and what is missing.
  3. 03Define how contributions will be combined and who will review the result.
  4. 04Choose a measure that matches the objective, such as accuracy against a reference value, time, cost, or the quality of the options.
  5. 05Compare the process with a relevant alternative before attributing an improvement to it.
03

Human swarms and simultaneous coordination

In a human swarm, several connected people can contribute responses to a question while a system coordinates the interaction. Rather than simply collecting independent votes and adding them up at the end, a swarm design may let the group adjust its response during the process. The details depend on the particular system; the label alone does not specify how opinions are weighted or what safeguards are in place.

One historical example is UNU, an application Unanimous AI described in a 2015 post. The company presented online, real-time sessions in which groups could answer questions, resolve disagreements, or make predictions. The post referred to a limited beta and invitations. That describes the offering announced at the time, not current availability.

The provider’s description helps explain the type of interaction it proposed, but it is not a substitute for independent evaluation. A claim that a group can “tap into its collective wisdom” is a product description, not a comparative measurement of accuracy, quality, or business impact.

Simultaneous coordination may be useful when an organization needs a group response quickly and participants have relevant information or judgment. It may also be unsuitable when the task requires extended deliberation, an individual audit of every argument, or a detailed explanation of why a decision was made.

Before a trial, specify who participates, what information each person sees, whether responses are anonymous, how consensus is communicated, and what happens if disagreement persists. A fast interaction does not ensure that the group has sufficient expertise, that participants are independent, or that the result is correct.

04

Particle swarm optimization

Particle swarm optimization (PSO) is an algorithmic method for searching for solutions to a problem defined by an objective function. The user specifies what to minimize or maximize and, where relevant, the constraints that solutions must satisfy. The algorithm initializes a set of candidates and updates their positions based on information associated with each candidate’s experience and information shared by the group or its neighbors.

PySwarms documentation describes each particle as a candidate moving through a search space; its movement depends on an individual tendency and collective influence. In this context, “swarm” is a computational metaphor. No people exchange judgments unless a separate system incorporates people into the process.

PSO can be of interest when searching for a solution to a problem whose function does not need to be differentiable. However, the search does not guarantee that it will find the global optimum. The result depends on how the problem was formulated, its constraints, the parameters, and the search behavior; a plausible result still needs to be checked.

In a hypothetical business application, a company could define a resource allocation plan as a combination of variables and evaluate each candidate with a function that penalizes delays and costs. Before using the result, it would need to check that the variables represent real decisions, that the constraints reflect operations, and that the function does not reward undesirable solutions.

The objective function is a modeling choice, not a neutral measure of “what is best.” If costs, labor limits, or service levels are omitted, the algorithm may optimize an incomplete representation of the problem. Human review and testing against known cases can help detect formulation errors, but they do not turn the method into a guarantee.

Questions for formulating an optimization problem

ElementEvaluation questionRisk if omitted
VariablesWhat decisions can the solution change?Candidates that do not correspond to possible actions
ObjectiveWhat value is being minimized or maximized, and why?Optimizing a measure that does not represent the actual need
ConstraintsWhat conditions must every acceptable solution satisfy?Proposals that are infeasible or violate operational limits
ValidationHow will the solution be checked outside the search?Confusing a model score with an operational result
05

Generative AI: support for creation, not an arbiter of truth

Generative AI includes models that emulate the structure and characteristics of their input data to produce derived synthetic content. That content may include text, images, audio, video, or other digital formats. The definition describes a class of models and their generative function; it does not establish that their outputs are true, legally original, or suitable for a decision.

In a collective activity, a generative tool can suggest alternatives that people then discuss, revise, or reject. This changes the process: the group is no longer simply sharing its own knowledge but is also responding to proposals from the model. It is worth recording what content the system contributes and how people review it.

The MIT Center for Collective Intelligence describes Supermind Ideator as a system that combines a large language model with instructions, fine-tuning, and an interface designed around creative problem-solving techniques. The center says it conducted a preliminary evaluation of early user experiences and was continuing broader empirical evaluations. That scope is not equivalent to a demonstration of general commercial effectiveness.

Another example is DesignAID, described as a tool that generates visual concepts from ideas expressed in words. The center’s page summarizes a study with participating designers and compares automatic image generation with an image-search condition. The reported result concerns the dimensions and conditions evaluated, not the quality of every AI-generated design or every creative team.

Generative AI can also introduce biases present in its data or in the way a prompt is written. In business settings, a convincing output may receive more trust than it deserves. Review should consider both the content and its intended use: a list of ideas does not need the same controls as a recommendation affecting people, safety, or resources.

06

What human–AI studies show—and what they do not establish

The results of collaboration between people and AI depend on the task and how the work is organized. A meta-analysis published in 2024 and summarized by the MIT Center for Collective Intelligence found that, on average, human–AI combinations performed worse than the better-performing component alone. The summary also reports differences by task type: combinations performed significantly better on creation tasks than on decision tasks.

That result challenges the assumption that adding an AI tool to a team will always improve performance. It does not show that every human–AI collaboration will perform worse, nor does it determine in advance what a particular business will achieve. The analysis average combines different studies and conditions; an organization needs to evaluate its own task and setup.

The MIT page also presents projects examining how generative AI can support human groups and mixed systems. For example, the Supermind Ideator project focuses on creative ideation and reports a preliminary user-experience evaluation alongside broader evaluations in progress. Claims about that system should remain within this context.

A useful evaluation distinguishes the performance of a team using AI from that of people without AI and the system without human intervention, where those comparisons are appropriate. Criteria might include quality assessed with a rubric, accuracy against a reference, time spent, or total cost. Measuring speed alone can conceal errors or correction work.

Who conducts the evaluation and what counts as success also matter. If participants know the expected answers, the set of cases favors one method, or only one dimension is measured, the result may not represent everyday work. A small pilot can inform a local decision, but it does not by itself establish a sustained or general effect.

Comparisons that help interpret a pilot

ComparisonWhat it can showWhat it does not establish on its own
People with AI versus the same people without AIChange associated with using the tool in that setupThat the change will persist across other tasks or teams
AI with people versus AI aloneThe contribution or cost of human intervention in the evaluated caseThat either component is superior in every context
New process versus usual practiceDifference from a real operational alternativeThat the difference came only from AI if the process also changed
Result versus a quality criterionWhether the output meets requirements defined for the taskThat all costs, risks, or downstream effects were measured
07

How to evaluate a pilot using checkable criteria

A pilot begins with a clearly defined task and a fair comparison, not a promise of return. Before activating a tool, define the population or cases to be evaluated, the expected outcome, the observation period, and the criteria for stopping or expanding the trial. If the goal is to improve a decision, specify how its quality will be measured and who can assess it.

Establish a baseline using the current process and preserve the conditions needed for a reasonable comparison. If staffing, workflow, and the system all change at once, it will be difficult to attribute the result to a single component. When a controlled comparison is not possible, describe clearly what changed and limit the conclusions.

Measure more than one dimension when the use of a tool could shift costs or risks. A tool that saves production time may increase review time; a system that suggests more ideas may increase the selection burden. Record quality, total time, operating and review costs, relevant errors, and the consequences of decisions.

Avoid choosing an indicator only because it is easy to obtain. For a creative task, the number of proposals does not demonstrate their usefulness. For an algorithmic search, a high function score does not prove that the solution is feasible. For a coordination system, agreement does not replace comparing the response with an independent criterion.

At the end, separate what was observed from what is inferred. If a team improves a measure in a brief test, it may be stated that the measure changed under the pilot conditions, provided the design supports that interpretation. That should not become a promise of future performance, a guarantee of return, or a general conclusion about the whole organization.

Evaluation sequence

  1. 01Define the problem and state the result that would be considered valuable.
  2. 02Record usual practice as a baseline and choose an appropriate comparison.
  3. 03Determine which cases enter the trial and who will assess the results.
  4. 04Measure quality and side effects alongside time and review costs.
  5. 05Document errors, unresolved cases, and conditions that limited the trial.
  6. 06Decide whether to expand, modify, or stop the pilot using criteria set before interpreting the results.
08

Data, bias, privacy, and adoption costs

Choosing a method does not replace reviewing the data it uses. A human swarm may expose personal information if responses are identifiable or shared without appropriate controls. A generative system may receive data in its prompts, and an optimization algorithm may use operational records containing sensitive information. Before a trial, determine what data is needed and who can access it.

Anonymity may reduce direct identification, but it does not prove that an activity is completely private or that identity cannot be inferred from other details. In 2015, Unanimous AI described UNU as anonymous in its beta announcement; that statement about the product at the time should not be interpreted as a current guarantee for another setup or date.

Bias can enter through participants, examples used to develop a system, operational data, the objective function, or evaluation criteria. A practical review asks who is represented, which groups may be left out, and whether the system produces different results under comparable conditions. A small test may detect problems, but it does not certify the absence of bias.

Adoption costs include more than a possible software fee: preparation, integration, training, human review, error management, and maintenance. The sources described here do not provide a general estimate of business costs or a return on investment applicable to a particular organization. Total cost needs to be calculated using the organization’s own data and workflow.

Oversight should match the impact of the task. For brainstorming, someone may filter suggestions before sharing them. For decisions about resources or people, more rigorous review of criteria, data, and consequences may be necessary. Oversight does not automatically make a result correct; it is part of the process of control and accountability.

Preliminary review by approach

ApproachWhat to reviewUseful pilot control
Human swarmParticipation, stated anonymity, influence dynamics, and handling of disagreementExplain how responses are combined and retain an evaluable record
PSOData, objective function, constraints, and definition of an acceptable solutionTest candidates against known cases and verify feasibility outside the algorithm
Generative AIInput data, output quality, bias, and intended useReview outputs against explicit criteria before using them
Human–computer systemInteraction, responsibilities, costs, and comparison with usual practiceRecord which component produced each contribution and how the final decision was made
09

Practical questions before choosing or deploying a method

An informed decision begins by specifying the problem and the evidence that would be enough to change current practice. If the goal is to generate more ideas, the criterion might focus on quality and useful diversity, not just volume. If the goal is a group estimate, a reference value is needed. If the goal is to optimize an operation, constraints and the cost of an infeasible solution should be explicit.

Ask whether the method requires connected participants, candidate solutions, or a model that generates content. Find out what data it receives, what result it retains, who reviews its outputs, and what happens when there is no agreement or a solution violates a constraint. Answers should describe the specific product and setup, not just the technology category.

Request a comparison that matches the use case and measures relevant outcomes. An average from research can inform a hypothesis, but it does not replace a local pilot. Conversely, a single demonstration does not prove that the result will recur under other conditions.

Define in advance how errors, uncertain cases, sensitive information, and harmful outcomes will be handled. Clarify who can stop the process and who is responsible for the final decision. A system that facilitates ideas or calculations does not remove the responsibility of the people applying them.

The useful conclusion is not that one technique is universally better, but that each addresses a different class of task through a different mechanism. **Choose the method for the problem, evaluate its result against an alternative, and preserve the limits of what the trial can establish.**

Decision checklist

  1. 01Is the objective defined as a relevant, measurable outcome?
  2. 02Is the mechanism human collaboration, algorithmic search, content generation, or a combination?
  3. 03Do the data and constraints match the real use?
  4. 04Is there an appropriate comparison with usual practice or another alternative?
  5. 05Will quality, total costs, human review, and possible adverse effects be measured?
  6. 06Are oversight, error handling, and the authority to stop the process clear?
  7. 07Are planned conclusions limited to the conditions actually evaluated?
10

Keep exploring

Sources consulted

03

Corrections and transparency

If you spot incorrect or outdated information, send us a correction with the page and source we should review.

Submit a correction