A project meeting ends with broad support for launching a pilot. Six people discuss benefits, schedules and staffing. One participant raises a quieter objection: the pilot’s error cannot be reversed after data has been shared. The AI-generated summary records the enthusiasm, the deadline and the action owners. The objection is gone.
This is an illustrative scenario, not a report of a particular meeting. It identifies a failure that matters beyond note-taking. A summary does not merely shorten a conversation; it decides which parts of the conversation become the working record. If a model preserves repeated agreement but drops a single material objection, the group may later remember consensus that never existed.
Research on opinion and long-document summarisation suggests that this is a testable system behaviour, not just a vague fear. The practical response is to evaluate AI meeting tools with a dissent test: deliberately place one consequential, well-supported objection in a controlled transcript and check whether the workflow preserves it accurately.

Compression is an editorial act
A transcript can contain thousands of words; a useful meeting note may contain a few hundred. The model must therefore select. Repeated topics provide many overlapping cues. A lone objection has less textual mass, even when it has more decision value. “Most frequently mentioned” and “most consequential if ignored” are different ranking functions.
This does not mean every model literally counts votes. It is a causal hypothesis about the task: when a prompt requests a short, coherent summary, the model has incentives to favour central, repeated and easily integrated material. A minority position may be compressed away because it appears once, disrupts the dominant narrative or cannot be resolved within the requested length.
Evidence from deliberation research makes the risk concrete. A 2025 preprint, Can AI Truly Represent Your Voice in Deliberations?, assembled human-grounded opinion and judgement data and evaluated 18 language models. The authors report persistent under-representation of minority positions as well as sensitivity to input order. The work concerns large-scale deliberation rather than ordinary staff meetings, so its results should not be transferred mechanically. The shared mechanism is selection under compression.
A peer-reviewed EACL 2024 study of diversity in argument summarisation similarly found that general-purpose language models and specialised systems struggle to represent arguments held by few people. And a NAACL 2025 study of long-form summarisation found a positional pattern: information in the middle of inputs was treated less faithfully than information near the beginning or end. Frequency and position can therefore interact. One objection, spoken once in the middle, is an especially demanding case.
Fluency hides the omission
An invented statement can be noticed because it conflicts with the transcript. An omission leaves no awkward sentence to challenge. The result may be grammatical, concise and factually compatible with everything it includes. Its failure lies in what the reader no longer knows to look for.
This is why ordinary quality questions—“Is the summary clear?” or “Does it capture the main points?”—are insufficient. The “main points” are exactly what the summariser has selected. A better review asks whether the note preserves every contribution that could change the decision, its conditions or its reversibility.
Meeting-summary research also shows that omissions are not easy to repair automatically. A COLING 2025 study created a human-annotated dataset of errors in generated meeting summaries and tested a two-stage process that first identifies mistakes and then refines the summary. The approach improved summaries, but partial omissions remained harder to identify than several other error types. Asking a model to review its own polished output is useful; it is not proof that missing material has been recovered.
Separate the summary from the dissent register
The simplest workflow change is structural. Keep the concise summary, but generate a separate dissent register from the transcript before finalising it. Each entry should contain:
- the decision, assumption or action being challenged;
- the objection in neutral language;
- the reason, evidence or condition attached to it;
- whether the group resolved, deferred or did not address it;
- a link or timestamp back to the source passage.
The register should not infer motives or manufacture balance. It records a position because of its relevance to the decision, not because every meeting needs two opposing camps. A second pass can then compare the draft summary with the register and ask: which material objections are absent, distorted or presented as resolved when the transcript leaves them open?
This complements the lesson in Why Example Order Can Change an AI Answer. Input order is not harmless plumbing. If a workflow’s output changes when the same objection moves from the opening to the middle, the meeting record depends on position rather than solely on substance.
Run a controlled dissent test
A team can test its summarisation workflow without using confidential meeting data. Write a synthetic transcript containing a clear majority view and one material objection. The objection should affect a real decision variable—for example, a safety constraint, an irreversible disclosure, an unbudgeted dependency or a condition for rollback. Keep the facts fictional and avoid personal data.
Create three versions with identical content but place the objection near the beginning, middle and end. In a fourth, paraphrase it in a different style. Generate summaries several times with the normal prompt and settings, then repeat with the two-pass summary-plus-register workflow.
Score four things:
- Coverage: is the objection present?
- Faithfulness: are its reason and scope preserved without exaggeration?
- Status: does the note correctly state whether it was resolved?
- Position stability: does moving the passage change those results?
This is a proposed experiment, not one Alkemata has run. The useful output is not a single “accuracy” score but a small failure matrix: omitted, weakened, misattributed, falsely resolved or preserved. Repeat the test when the model, prompt, transcript pipeline or summary length changes.
Dissent is not a veto
The strongest objection to this method is that recording every disagreement can turn a meeting note into a catalogue of resistance. A trivial reservation may receive the same visual weight as broad evidence, and participants could learn that objecting guarantees prominence.
The answer is materiality, not silence. A dissent belongs in the register when accepting it would change the decision, required evidence, risk owner, rollback plan or condition for proceeding. Recording it does not grant it equal evidential weight or a veto. The note can state the group’s response and final decision. What it should not do is erase the existence of the decision-relevant challenge.
A generated summary becomes organisational memory surprisingly quickly: it feeds task lists, later briefings and future AI searches. Before granting it that role, a team can make one small demand. Move the consequential objection, rerun the system and see whether the record still remembers what the group had reason to worry about.
What concrete example would confirm or challenge this article?
Choose a reaction, write a comment, or ask your AI assistant to send a proposal for review. Comments are reviewed before publication.