Can discourse communities collectively detect disruptions individual readers miss?
This explores whether the collective — many readers, averaged annotations, a community's shared interpretive posture — catches breaks in meaning, false claims, or norm violations that any single reader would slide past, and what the corpus says about where that collective advantage comes from and where it breaks down.
This reads the question as: is detection a group capacity rather than an individual one — can a discourse community register disruptions (incoherence, false presuppositions, off-key AI text) that lone readers miss? The corpus leans toward yes, but for a reason more interesting than 'crowds are smarter.' The collective isn't just more eyes; it's a different kind of signal. When researchers average human judgments of where one event ends and another begins, that consensus boundary is sharper and more reliable than any individual annotator's — so much so that GPT-3's segmentation tracks the *averaged* humans better than real humans track each other llms-segment-narrative-events-closer-to-human-consensus-than-individual-han. The community computes something no member holds alone.
But the same corpus warns against treating consensus as the whole story. Disagreement among readers is often not noise to be averaged away — it's information. Interpretations of socially loaded sentences are irreducibly multiple because readers occupy different social positions, and the *spread* of those readings carries meaning a single 'correct' answer would erase Why do readers interpret the same sentence so differently?. So a discourse community detects disruptions in two opposite ways at once: by converging (consensus surfaces the event break no one reader pins down) and by diverging (the pattern of who-reads-it-how flags the socially charged spot). A disruption can be precisely the place where the community's readings scatter.
The sharper finding is about *what only the collective can do.* An AI can predict social appropriateness better than any individual human, yet it structurally cannot enter the community process that creates and validates norms — it pattern-matches the output without participating in the making Can AI predict social norms better than humans?. Detecting a norm *disruption* is an act of that same participatory process. This is why calibration is collective, not individual: models flag irony as a pattern but wildly overestimate how often it actually occurs, because they never absorbed the community's lived sense of how rare it is in ordinary use Do language models overestimate how often irony appears?. Frequency-in-use is knowledge that lives in the community, not the text.
Where this becomes urgent is AI-generated discourse itself. Every established source — advertising, journalism, a politician's speech — carries an interpretive posture the public learned to apply: a 'discount' that filters how the words land. AI text arrived too fast and shifts too quickly for any such collective posture to form, so it circulates without the protective skepticism we automatically bring to interested speech How do we learn to read AI-generated text critically?. That's the clearest case for the question's premise: an individual reader, lacking the communal discount, can't reliably detect the disruption of machine-authored persuasion — and right now the *community* can't either, because it hasn't yet built the shared stance that detection depends on.
There's a final twist the corpus offers: comprehension itself runs on layers a lone reader holds in tension — tracking linguistic segments, the speaker's purposes, and shifting salience all at once, where a failure in any one disrupts the whole How do readers track segments, purposes, and salience together?. Some disruptions register only at the semantic level — contradiction, broken coreference, creeping irrelevance — and can be caught by formal representations no casual reader applies What semantic failures break dialogue coherence most realistically?. The takeaway you didn't know you wanted: collective detection works not because the group is wiser, but because meaning was always a social settlement — and a community can sense when that settlement has been quietly broken, even when no single member can say where.
Sources 7 notes
Interpretation Modeling research shows that disagreement on socially embedded sentences reflects valid differences in reader perspective, not annotation failure. Structured human disagreement in NLI benchmarks confirms that interpretation distributions carry meaningful information.
GPT-4.5 outperforms all individual humans at predicting social appropriateness, yet structurally cannot enter the community processes that establish and validate norms. This reveals a critical gap between pattern-matching and authentic participation in knowledge-making.
GPT-4o assigns significantly higher irony scores than humans (p < .001), revealing that LLMs detect irony as a pattern but miscalibrate its prevalence because ironic examples are more salient in training data than in actual use.
Every established discourse source carries an interpretive posture that filters how publics receive it. AI-generated text arrived too recently and shifts too quickly to anchor such a posture, allowing it to spread without the protective skepticism we automatically apply to interested speech.
Discourse processing demands parallel recognition of linguistic segments, intentional structure, and attentional salience—not sequential processing. These three layers constrain each other during comprehension, and failures in any single layer disrupt overall understanding.
Research using Abstract Meaning Representation identified four distinct incoherence types: contradiction, coreference inconsistency, irrelevancy, and decreased engagement. AMR-trained classifiers detect these semantic failures while text-level manipulations alone cannot.