If AI representatives exchange views faster than people can follow, how can participants inspect and contest a human-facing summary before a decision deadline?

Contribution to the handbook

Priority · Assessed 2026-10-30

The question identifies a consequential design and timing tradeoff: participants may need to inspect and influence a summary before it informs a decision, even when exchanges are too fast to follow directly. The proposed fast-representative scenario is not shown to work in practice, and a deliberative summary is not itself authorization for a decision.

Does this challenge the goal?

In-scope gap

The question asks how to make a possible AI-mediated process inspectable and contestable within a deadline. It does not contest the exploration’s goal or desired outcome.

Whether such a process can provide meaningful review remains unresolved; the definition does not prescribe a review mechanism.

Making and interpreting laws →

Authority, accountability and resilience →

Tests, stop conditions and unresolved work →

Identify what exchange record, summary provenance, reasons, and unresolved disagreements participants would need to inspect to check whether the human-facing summary reflects the exchange.

What we can say so far

A starting point would be an inspectable record of the AI representatives’ contributions, linked to the claims in the human-facing summary, so participants can compare the summary with what was said. Inspection should include the reasons and evidence used, the summary’s inputs and revisions, and views or disagreements it omits or describes as resolved. These are inferred requirements, not a documented procedure for fast AI-to-AI exchange. The supplied sources support examining evidence and reasoning, and report one study in which revisions used participants’ initial opinions, an initial statement and their critiques. They do not specify a complete exchange record or provenance mechanism that would establish summary fidelity.

What remains open

• No supplied procedure specifies how to capture the complete exchange and link each summary claim or omission to its source contributions.
• The evidence does not specify what provenance must be retained about summarization or how unresolved disagreements and their reasons are reliably identified.
• The evidence does not establish that access to a record makes an opaque system or its reasoning meaningfully inspectable.

Specify who may correct or challenge the summary, how a challenge can affect it or the pending decision, and what minimum opportunity should be protected when time is limited, including when a decision should be deferred.

What we can say so far

Participants in the reported study could submit written critiques, which the AI used to revise a group statement. Separately, Crawford and Schultz propose that people targeted by an automated system should be able to examine its evidence and reasoning and appeal to an impartial decision-maker. These provide starting points, but do not settle the proposed challenge procedure: the supplied material does not specify who must have correction rights, how an unresolved challenge affects a pending decision, a protected minimum opportunity under deadline pressure, or when the decision must be deferred.

What remains open

• Which participants or affected people are entitled to correct or challenge the summary.
• Whether and how a challenge must change the summary, be recorded as unresolved, or affect the pending decision.
• The minimum meaningful opportunity for review and challenge when time is limited.
• The circumstances and authority for deferring a decision, including how to handle an urgent deadline.

Explain how inspection, responses to challenges, and any final authorization could fit before a decision deadline without treating a nominal chance to object as meaningful review.

What we can say so far

The supplied evidence suggests part of a sequence: give participants a human-facing statement to inspect, collect their critiques, and use those critiques alongside their original views to revise the statement. In one study, a deliberation round on one question took about 15 minutes; this does not establish that meaningful review can always be completed in that time. [s-habermas-machine-406eb21715f7a4b3] A meaningful deadline would also have to leave enough time and access for participants to understand and challenge the account, not merely provide a formal opportunity to object. The supplied material does not specify how to allocate that time, ensure challenges receive responses, handle unresolved objections, or obtain final authorization before the deadline. Nor does it establish what should happen if adequate review cannot be completed in time.

What remains open

• No supplied procedure allocates the available time among inspection, challenges, responses, revision, and final authorization.
• No supplied mechanism ensures that challenges receive a response or records how unresolved or conflicting objections are treated before the deadline.
• The evidence does not specify who authorizes a final decision, how that authorization occurs in time, or what happens if meaningful review cannot be completed.
• No supplied evidence establishes an adequate review window or demonstrates effectiveness of such a process for fast AI-representative exchanges.

Compare a proposed fast-exchange process with a relevant slower or human-paced alternative on decision time, affected people’s understanding, and their ability to change the summary or outcome; distinguish proposed gains from tested effects.

What we can say so far

A proposed fast process has AI representatives exchange views and give people a summary; a slower alternative includes people directly in the discussion. The proposed process could reduce exchange time, but the sources do not compare the two on total decision time, participants’ understanding, or their influence over the summary or outcome. Direct discussion may retain nonverbal cues and relationship-building opportunities, but can also take too long for a deadline. A related, but different, study of AI-assisted group statements reported statement production in seconds versus several minutes for human mediators; this is not a comparison of complete decisions or agent-to-agent exchanges. Its approximately 15-minute rounds let participants critique an initial statement, and participants preferred revised statements incorporating critiques. Outside raters judged AI statements clearer and more informative than human-written ones, but that is not a test of affected participants’ understanding. These findings do not establish comparative influence over a decision outcome. Authors’ broader claims about efficiency, scale and inclusion remain proposed gains, not established effects.

What remains open

• No comparative measurement of total decision time for fast representative exchange versus human-paced deliberation.
• No comparative measurement of affected people’s understanding or their ability to change the summary or decision outcome.
• No evidence establishing that the study’s statement-revision findings transfer to fast exchanges between personal AI representatives.