AI and collective intelligence: help people think together

Start with the human process

A community has collected hundreds of suggestions about a public space. Some repeat each other. Some use different languages. Some contain a concern that appears only once but matters greatly to a small group. AI could help organize the material so participants can see themes and unanswered questions.

The first design decision should be what the community needs to understand—not which model is most impressive. Define the question, the people affected, the decision authority, and the evidence that will be used. Then ask where a tool could reduce a specific burden without removing meaningful human judgment.

Useful roles are bounded and inspectable

Translation drafts can help participants cross language barriers, provided a competent person can review important passages. A summary can help readers navigate a long record if it links to the original contributions. A model can suggest counterarguments, but those suggestions need to be checked against evidence.

A useful interface could show a theme beside the statements that support it, the statements that challenge it, and the editor’s changes. Participants should be able to say “this does not represent my view” and have a visible correction route. Confidence in the tool should come from that process, not from its fluent tone.

Keep the original material accessible to authorized reviewers. A summary should be a doorway into the conversation, not the only surviving version of it.

The risks are part of the architecture

NIST’s AI Risk Management Framework addresses risks to individuals, organizations, and society and encourages trustworthiness throughout AI design, use, and evaluation. It provides a useful starting point for assessing a proposed workflow. NIST · AI Risk Management Framework

For the community example, test whether the model omits less common views, misrepresents uncertain statements, or invents support for a recommendation. A summary that sounds balanced may still distort the record. Also examine who can access the underlying submissions and whether participants understood how their words would be processed.

Avoid entering private material into a service without an appropriate basis and clear expectations. More automated analysis is not always worth the privacy cost.

A pilot that can teach you something

Select a small set of contributions that reviewers can inspect in full. Ask two people to summarize them independently. Generate a machine-assisted draft. Compare coverage, attribution, omissions, and time spent correcting errors. Include at least one reviewer familiar with concerns that may be easy to miss.

Define failure conditions in advance. For example, if the tool repeatedly drops objections from a language group, do not use it to represent that group’s views. If reviewers cannot trace a claim back to a source, the summary is not ready for a decision record.

Publish an understandable account of what the tool did and what people checked. Do not present a small pilot as proof that a method is safe at every scale.

Keep legitimacy with people

A model can estimate, summarize, and recommend. It cannot by itself establish what people owe each other or whose interests should prevail. A technical output does not remove the need for consent, accountable authority, and the ability to contest a decision.

The positive possibility is that AI helps more people enter the conversation and makes complex material easier to inspect. The failure mode is a new hierarchy hidden behind an interface: the model speaks, the operator interprets, and everyone else is expected to comply.

Choose tools that leave people more capable after using them. The desired outcome is a community that understands its choices better, preserves its own memory, and can change the process when it fails.