The notice came too late to read. The words were in another tongue. I stood there with the others all and what I said, it wrote me down.
The plural record survives the meeting. What kills it is usually the next step: the report, the digest, the intake queue, the summarized case file. Any process that turns many accounts into one output makes choices among them, and modern systems make those choices silently, at scale, with no record that a choice was made.
The mechanism is ordinary. A summarizer produces one output. A deduplication pass collapses near-matches. A consensus tool reports what "the" account says. Each operation resolves plurality by averaging or picking, and the losing account does not vanish from the world; it vanishes from the record that downstream processes read. Nobody chose. The losing account simply failed to survive the pipeline, and the pipeline produces the single account that the systems of the previous arcs all assumed they were being handed.
What the pipeline drops
The failure is specific and observable. Peters and Chin-Yee tested LLM summarization of scientific research and found a strong tendency to overgeneralize: summaries reported conclusions as firmer and broader than the sources stated them, and the tendency increased in newer models (Generalization bias in large language model summarization of scientific research, Royal Society Open Science, 2025). One finding in that paper deserves a slower reading than it usually gets: asking the models explicitly for faithful, accurate summaries made the overgeneralization worse, and the models overgeneralized more than human science communicators summarizing the same material. The instruction to be faithful did not protect the record. It polished the distortion.
The intake and digest pipelines in grievance practice inherit this. A model that compresses a consultation record resolves the disagreement inside it toward whichever account dominates the surface of the text. The minutes beat the lived account because the minutes arrive already formatted like an official record, and a compressor reads formatting as authority.
The same failure has a duplicate-detection cousin. Two accounts that agree in substance but differ in detail or register are candidates for "deduplication", and the pass that merges them chooses which details are noise. In consultation records the merged details are often the entire difference between consent and objection. Deduplication of testimony is an unrecorded adjudication wearing a compression's clothes.
Where the picking happens
A human editor who reconciles two accounts makes a decision that can be asked about. A pipeline that resolves them statistically makes a determination that nobody decided, in the exact sense The Duty Survives the Tool located the duty elsewhere: the authority was exercised, and no party held it.
The design response is mundane and strict. Pipelines that touch plural accounts must preserve attribution at the grain of the account: every reduction records what entered it, and any output that stands for several accounts names them and keeps them retrievable. Where a determination must be made, the rule that made it belongs in the record with the same explicitness the Account arc demanded of a threshold. The alternative is the grove filed by a clerk who kept only one testimony because it was the average of the other six.
The digest at the top of the file
Picture the ordinary version. A grievance team receives forty submissions after the March meeting and runs them through a summarizing tool so the monthly report can carry a paragraph instead of a stack. The paragraph is competent. It says community members raised concerns about the consultation process, including timing and language access, and that the company provided responses at the meeting. Every clause is supported by something in the stack. The paragraph has still decided the case, because "raised concerns" and "provided responses" place the two records in the relationship of question and answer, which is the relationship the minutes proposed and the community rejected.
Nobody reading the report sees the decision. The regional manager sees a paragraph. The board sees a line in a dashboard (we do love a dashboard) reporting consultation grievances as "addressed". The forty submissions remain on a server somewhere, findable in principle, and nobody downstream will open them, because the summary exists so that nobody has to.
Categorization does the same work with less prose. A grievance logged under "communication" has already been read as a misunderstanding, and one logged under "consent" has been read as a dispute about authority. The taxonomy is an adjudication drafted in advance, usually by whoever built the intake system, and it is applied at the moment of entry by whichever staff member or model fills the field. Trend reports then count the categories, and the count inherits every choice the field made.
Companions
- The address where the second account arrived: The Account Must Have an Address.
- The system that certified resemblance: The Container Is the Tell.
- Peters, U., & Chin-Yee, B., "Generalization bias in large language model summarization of scientific research", Royal Society Open Science 12(4):241776 (2025).
- The fiction: Akutagawa, "In a Grove", on Wikisource.
These notes come out of Sociable Systems, a practice that reads AI-shaped documents the way a hostile reviewer will, before a lender or a court finds the gap. The argument has an operational form: the Interim Protocol sets out four rules for AI use in environmental and social deliverables, covering disclosure at touch-point grain, evidence custody, the phrases no automated screening may settle, and a hostile read before anything ships. Free, and written to be cited or retired once institutional guidance arrives.