Is a median sometimes better than a mean?

It can be more resistant to extreme numerical estimates, but the appropriate choice depends on the task and error distribution.

Can a crowd be made entirely of models?

Yes, as in an ensemble, although the independence and diversity of their errors remain important.

Does disagreement mean the process failed?

No. It can reveal uncertainty or conflicting assumptions that deserve investigation before a decision.

AI Glossary / AI and machine learning / Beginner

Wisdom of the crowd

Definition

Wisdom of the crowd is the idea that combining judgments from multiple contributors can produce a useful collective estimate. In machine learning, ensembles provide a related mechanism by combining the outputs of several models.

What it means

An individual estimate may be too high or too low. If different contributors make partly independent errors, aggregation can cancel some of those errors. The combined result may be more stable than a single contribution.

The idea is conditional, not a rule that the majority is always right. Shared misinformation, similar assumptions or pressure to conform can make errors move in the same direction.

For AI, the word crowd may describe several models, repeated runs or human reviewers. Those are different arrangements with different degrees of independence. Ten answers from the same model are not automatically ten independent sources of knowledge.

For business readers, the useful question is what diversity the contributors bring. Combining outputs is valuable when it improves evidence or error balance, not merely when it increases the number of agreeing voices.

How it works

The team defines the quantity or decision to estimate and selects contributors. Their information sources, methods and incentives affect whether their errors are likely to be shared.

An aggregation rule combines the results. Numerical estimates might be averaged or summarized by a median. Category decisions might use a vote. More elaborate methods may assign weights based on evaluated reliability.

The collective output is tested against an appropriate reference or later outcome. The team compares it with the strongest individual contributor and a simple baseline.

Disagreement is also informative. It can identify ambiguous cases or assumptions requiring investigation, but agreement should still be checked when all contributors rely on the same limited evidence.

Three practical examples

Forecasting: combine estimates from models that represent different patterns, then evaluate the aggregate on later periods.
Review workflows: collect independent judgments before discussion so early opinions do not shape every reviewer.
Classification: use an ensemble while examining whether the component models make complementary or shared errors.

Think of it this way

Think of measuring a table with several imperfect rulers. Small errors in different directions might partly cancel when measurements are combined.

The analogy has a limit. If every ruler has the same manufacturing defect, taking more measurements will preserve the same bias. Likewise, a crowd sharing one flawed source can become consistently wrong rather than collectively wise.

Worked example

When averaging helps and when it does not

Suppose three invented forecasts for demand are 90, 100 and 110 units, while the later observed demand is 100. Their average is 100, and the errors of minus ten and plus ten cancel.

Now suppose three other forecasts are 110, 115 and 120 because all use the same overstated growth assumption. Their average is 115, still 15 above the outcome.

The first example demonstrates possible error cancellation; the second demonstrates shared bias. Neither small set establishes what will happen across future forecasts.

To evaluate a crowd approach, the team needs many representative cases and a consistent aggregation rule. It should also investigate why contributors differ, since diversity created by random noise is not necessarily useful expertise.

Illustrative case study

A fictional editorial team asks three AI assistants to verify a claim. All repeat the same statement from a widely copied but incorrect summary, and the team initially treats agreement as confirmation.

The editor traces the evidence to its origin and introduces a process that asks reviewers to identify independent supporting records. The team records uncertainty when multiple answers depend on one underlying source.

The lesson is that agreement and independent corroboration are different. A crowd can amplify a shared error when its members draw from the same information pathway.

Common misconceptions and mistakes

More contributors always improve the answer. Additional biased or redundant contributions may add little value or worsen it.
Majority agreement establishes truth. It establishes agreement under the chosen group and process.
Repeated AI outputs are independent experts. They may share model parameters, prompts and source material, creating strongly correlated errors.

Business use

The idea can improve forecasting, review and model design when contributors offer complementary evidence or methods. It is also a useful warning against relying on one uncertain estimate.

Design for independent initial judgment where appropriate, document aggregation and evaluate actual outcomes. Keep the strongest individual baseline in the comparison. For factual research, trace source independence rather than count how many systems repeat a claim. Quality depends on the information structure as well as the number of contributors.

Check your understanding

Five models all use the same incorrect input figure and produce similar estimates. Does their agreement provide five independent confirmations?

Answer: No. The shared input creates a common error source. The figure needs verification regardless of the number of agreeing models.

Related terms

Ensemble
Voting Classifier
Voting Regressor
Aggregation
correlated errors
independent evidence

Learn next

Read ensemble and voting regressor. Compare contributors' errors across the same cases, and trace whether apparent agreement comes from independent evidence or a common source.

Sources

Google, Machine Learning Glossary, wisdom of the crowd: https://developers.google.com/machine-learning/glossary#wisdom_of_the_crowd. Ensemble analogy consulted. Forecast arithmetic and editorial case are original explanatory examples.

Previous: Winsorized samples Next: Wishart distribution