Disagreement is Data: Why AI Needs More Arguments, Not Fewer

A moderation queue looks simple until two reasonable reviewers disagree.

One reviewer sees a political comment as ordinary partisan sarcasm. Another sees the same sentence as offensive. A third is unsure, which is not the same as being confused. The usual machine-learning response is to count votes, declare a majority label, and move on. Very efficient. Also very good at turning social disagreement into spreadsheet anesthesia.

A 2026 paper, Learning Who Disagrees: Demographic Importance Weighting for Modeling Annotator Distributions with DiADEM, pushes against that habit.¹ Its central claim is not merely that annotators disagree. Everyone who has managed labeling work already knows this, usually after paying for the same item to be labeled five times. The sharper claim is that some disagreement is structured enough to be modeled, and that a model can learn not only what label is likely, but which kinds of annotators are likely to disagree about it.

That distinction matters. Most AI evaluation pipelines still behave as if disagreement is a quality-control problem: remove bad labels, resolve conflicts, find the gold answer. DiADEM treats disagreement as part of the target. It asks a less comfortable question: when people disagree about safety or offensiveness, what information should the model preserve rather than flatten?

The answer is not “let demographics decide truth.” That would be both technically lazy and socially radioactive. The answer is more precise: when the task is subjective, demographic and contextual information may help predict the distribution of human judgments. The model should learn how much those signals matter, where they matter, and how much uncertainty remains after they are used.

That is the useful managerial lesson. The paper is not just another benchmark win. It is a design argument for AI systems that handle subjective judgment.

The old pipeline quietly deletes the thing managers need to know

Traditional supervised learning likes a single label because a single label makes training simple. If five people review an item and three choose “safe” while two choose “unsafe,” the majority label becomes “safe.” The two dissenting judgments disappear from the training target.

This is tolerable when the task is objective and disagreement mostly comes from error. If a labeler misreads a digit or misses a named entity, aggregation can reduce noise. But content safety, political offensiveness, customer complaints, policy interpretation, and human-risk evaluation are not always like that. In those domains, disagreement may reveal ambiguity, group-specific perception, local norms, or simply an unresolved policy edge case.

The practical mistake is treating all disagreement as if it has the same origin. Some disagreement is noise. Some is ambiguity. Some is a minority perspective that the organization may legally, ethically, or commercially need to see. Majority vote cannot tell the difference. It only tells the organization which view had more seats in the annotation sample.

DiADEM enters at exactly this point. It does not ask the model to find the “real” label behind human disagreement. It asks the model to learn a distribution of responses conditioned on both the item and the annotator.

That sounds like a technical shift. It is also an organizational shift. The output is no longer just “this item is unsafe.” It becomes closer to: “this item is likely to be judged unsafe by these annotator profiles, safe by those profiles, and ambiguous overall.” For governance, that is a very different object.

DiADEM models the annotator, the item, and the collision between them

The paper builds DiADEM around a simple but powerful change: replace anonymous annotator IDs with demographic-aware annotator representations.

A conventional multi-annotator model can learn that annotator 127 is stricter than annotator 381. Useful, but brittle. If a new annotator appears, the model has no history for that person. It can only fall back toward the average. That is precisely where subjective AI systems become dull: they become “reasonable” by becoming majoritarian.

DiADEM instead encodes annotators through demographic attributes such as age, race, gender, education, locale, and political affiliation, depending on the dataset. Each demographic feature is projected into a representation, and the model learns an importance weight for each demographic axis. Those weights are normalized and trained end to end.

The mechanism matters because it creates two layers of learning:

The “fusion” part is easy to underestimate. DiADEM does not merely append demographic fields to text and hope the network behaves wisely. It combines item and annotator representations through both concatenation and Hadamard, or element-wise, interactions. Concatenation gives the model a broad joint representation. Hadamard interactions help capture more granular compatibility between item dimensions and annotator dimensions.

In plain business language: the system is not only learning that some reviewers are stricter. It is learning that some reviewers may be stricter on some kinds of content. That is the difference between crude segmentation and useful perspective modeling.

The disagreement loss is the paper’s most business-relevant trick

The paper’s most important design choice may be its item-level disagreement loss. DiADEM is not trained only to get individual labels right. It is also trained to match the observed variance among annotators.

That is a subtle but large change. A model can be fairly accurate while still misrepresenting disagreement. Suppose most annotators choose “No,” but a meaningful minority choose “Yes.” A majority-seeking model can score well by predicting “No” with high confidence. It looks competent. It has also erased the very cases a safety or policy team may want to inspect.

DiADEM’s additional loss penalizes mismatches between actual and predicted annotator variance. If humans strongly disagree on an item, the model is pushed to predict disagreement. If humans largely agree, the model should not invent controversy just to look sophisticated. We have enough systems doing that already.

This turns disagreement from an unfortunate residue into a training target. That is the paper’s architectural thesis.

The model has three output paths: aggregate item-level label distributions, per-annotator predictions, and annotator-level behavior distributions. This multi-objective design is important because subjective annotation is not a single problem. A deployment team may want the likely aggregate view, the likely response of a specific reviewer profile, and a signal of whether the case is likely to be contested.

In a compliance workflow, those are different decisions. Aggregate distribution helps prioritize. Per-annotator prediction helps audit perspective coverage. Disagreement prediction helps decide whether to escalate. One model output is a label. These outputs are a governance interface.

The strongest evidence appears when the model meets unseen annotators

The paper evaluates DiADEM on two datasets: DICES, a conversational AI safety dataset with demographic metadata, and VOICED, a political-offense dataset where annotators judge offensiveness and political affiliation is part of the perspective signal.

The authors use two split types. This matters more than it may first appear.

On the annotator-level DICES split, DiADEM is the strongest model on the standard classification metrics reported by the paper. It reaches 0.7337 accuracy, 0.4116 macro-F1, 0.6980 weighted-F1, 0.2450 Cohen’s kappa, and 0.2643 MCC. More importantly for the paper’s thesis, it also performs well on distributional and perspectivist metrics: JSD 0.0446, ER 0.4911, and ECE 0.0391.

Those numbers are not just leaderboard confetti. The annotator-level split is the cold-start setting: the model must predict responses from annotators it did not see during training. Anonymous ID embeddings are naturally weak here because there is no useful ID history for a new person. DiADEM’s demographic-aware encoding gives the model a way to transfer what it learned from similar profiles, while still conditioning on the item.

The same pattern appears on VOICED. In the annotator-level split, DiADEM reports 0.7646 accuracy, 0.5587 macro-F1, 0.7025 weighted-F1, 0.2616 MCC, and the best soft/perspectivist scores among the compared systems, including JSD 0.0923 and ECE 0.0129. The paper’s point is not that these absolute scores are magical. They are not. The point is that the architecture improves where the problem is hardest: predicting perspective-conditioned judgments for people the model has not already memorized.

That is also where the business case lives. A company rarely has rich historical labels for every future reviewer, customer segment, jurisdiction, or stakeholder group. A cold-start mechanism is not a benchmark detail. It is the difference between a system that can generalize to new human groups and one that quietly defaults to the majority it already knows.

Item-level results show useful gains, but not a clean sweep

The item-level split tells a slightly different story, and it is worth preserving the messiness.

On DICES item-level evaluation, DiADEM leads on accuracy, weighted-F1, MCC, JSD, MD, and ER. But LeWiDi slightly leads on macro-F1, Cohen’s kappa, and ECE. That is not a fatal weakness. It is a reminder that no single metric captures perspectivist modeling. Macro-F1 is sensitive to class balance, ECE measures calibration, and JSD/ER focus on distributional alignment.

The business reading should be: DiADEM is strongest on the metrics most aligned with its design goal, but deployment teams should not worship one metric because the paper said “disagreement.” Conveniently, the paper itself warns against reading conclusions from one score in isolation.

On VOICED item-level evaluation, the case is cleaner. DiADEM reaches 0.8000 accuracy and leads on the main soft and perspectivist metrics, including JSD 0.0250 and ECE 0.0144. DisCo remains competitive on some hard-label measures, but DiADEM’s distribution-sensitive gains better match the task the paper cares about: modeling uncertainty and disagreement rather than merely predicting the most common label.

This is an important distinction for procurement teams. A model that is “accurate” on a flattened label may still be poor at identifying controversial cases. If the product requirement is standard classification, a simpler model may be enough. If the requirement is perspective-aware triage, soft-label fidelity and disagreement calibration become first-class metrics.

Low divergence can be fake comfort

One of the paper’s most useful warnings comes from a failure mode in the DICES annotator-level results. The LeWiDi-style DisCo variant achieves an artificially low JSD because it collapses to predicting only one class, “No,” with zero recall on the other classes. The paper flags this as degenerate collapse, not genuine distributional alignment.

This is exactly the kind of metric trap that can make an AI dashboard look healthier than the underlying system. A model can appear distributionally close under one measure while failing to represent minority labels. In subjective domains, the harm is not just lower performance. The harm is false reassurance.

The operational lesson is straightforward: disagreement-aware systems need metric bundles. A useful evaluation should combine hard-label performance, soft distribution alignment, calibration, minority-class behavior, and direct variance or entropy tracking. If a model has excellent divergence but ignores minority responses, it is not perspectivist. It is just confidently boring.

The LLM role-play baseline fails for an instructive reason

The paper compares DiADEM against several LLM baselines, including prompted role-play setups where the model is given demographic information and asked to answer as a human annotator. The prompts even ask the LLM to reason step by step before producing a label.

This is a useful test because it targets a very current temptation: why build a specialized disagreement model if a large model can simply “pretend” to be different people?

The results are not flattering to that temptation. Across the reported DICES and VOICED annotator-level tables, the LLM baselines trail DiADEM on the key standard, soft, and perspectivist metrics. They can generate plausible-sounding demographic reasoning, but plausible role-play is not the same as recovering the empirical structure of human disagreement.

This should not be overread. The paper does not prove that LLMs can never help with annotation, evaluation, or synthetic perspective generation. It does show that prompting an LLM with demographic descriptions is not a substitute for training against observed annotator distributions. Persona simulation is cheap. Empirical grounding is not. Annoying, but reality often invoices separately.

For businesses, the implication is not “avoid LLM judges.” It is: do not treat LLM-as-a-judge outputs as equivalent to human-disagreement modeling unless you have tested that equivalence against human distributional data. Especially in safety, policy, and offensive-content settings, the difference is not academic. It changes escalation, audit, and risk visibility.

The learned weights are diagnostic, not destiny

DiADEM’s learned demographic weights are one of the paper’s more interpretable features. They show which demographic axes the model found useful for predicting disagreement in each dataset and split.

Across the reported weights, race is the highest-weighted demographic in DICES for both annotator-level and item-level splits, and also highest in VOICED’s annotator-level split. In VOICED’s item-level split, education receives the highest weight, with race, gender, age, and political affiliation following. Age is often close to the top, especially in VOICED annotator-level results.

That pattern is interesting, but it should be read carefully. The weights are model-learned importance signals for these datasets. They are not universal social laws. They do not prove that race or age “causes” disagreement. They indicate that, given the available metadata and task setup, those features helped the model predict annotator responses.

This difference matters. A careless organization could look at such weights and start building crude demographic rules. That would be the fastest route from “perspectivist AI” to “segmentation with legal problems.” The better use is diagnostic: identify where disagreement appears structured, check whether the annotation sample is balanced, and decide whether additional review, richer metadata, or policy clarification is needed.

The paper’s appendix reinforces this discipline. It explicitly notes demographic sparsity, split- and metric-dependence, limited baseline coverage, and the fact that demographics are incomplete contextual proxies. That limitation is not a decorative paragraph placed at the end to appease reviewers. It affects how the method should be used.

The appendix evidence supports robustness, not a second thesis

The appendix does several useful things. It provides confusion matrices, per-class F1/support plots, disagreement calibration plots, variance and entropy correlation results, and qualitative examples where similar annotator profiles still receive different predicted labels.

The variance-correlation results are especially informative. On DICES by-annotator evaluation, DiADEM reports strong alignment between predicted and empirical disagreement: variance Pearson 0.7544 and Spearman 0.7406; entropy Pearson 0.7292 and Spearman 0.7176. The DisCo baseline is near zero on the same DICES disagreement-correlation measures, while LeWiDi is competitive but lower than DiADEM.

VOICED is more modest. DiADEM’s by-annotator variance correlations are around 0.1878 Pearson and 0.1928 Spearman, still above the listed baselines but far from the DICES strength. This is where the paper becomes more useful precisely because it is not perfectly neat. Disagreement is partly predictable, but predictability depends on dataset, task, labels, sample structure, and available metadata.

The qualitative examples also have a limited purpose. They show that DiADEM does not simply collapse everyone with similar demographics into the same prediction. That is helpful as a sanity check. It is not statistical proof of fairness, causal reasoning, or subgroup reliability. In a serious deployment review, those examples would support model inspection, not replace subgroup audits.

What this changes for business AI systems

The most immediate use case is not a flashy product. It is better diagnosis.

Organizations already use human labels for moderation, safety evaluation, customer sentiment, compliance triage, insurance review, fraud investigation, HR classification, and policy interpretation. Many of those systems still train against flattened labels and then wonder why deployment creates stakeholder complaints. The answer is sometimes boring: the model learned the majority label because that is what the training pipeline told it to learn.

A DiADEM-like approach suggests a different workflow.

First, preserve annotator-level labels instead of destroying them during preprocessing. Second, collect perspective metadata only where lawful, ethical, consented, and genuinely relevant. Third, train models to predict both label distributions and disagreement. Fourth, use disagreement predictions as workflow signals: escalate contested cases, sample them for policy review, or route them to more diverse human panels. Fifth, evaluate models with metric bundles that include calibration and distributional fidelity, not just accuracy.

The return on investment is not simply “higher model accuracy.” That is the easy sales slide, and therefore suspicious. The better ROI is reduced blind spots:

Cognaptus’ inference from the paper is therefore practical but bounded: disagreement-aware modeling can improve visibility in subjective decision systems. It can help organizations know when their AI is confident because humans agree, and when it is confident because the pipeline erased disagreement before training began.

That is a very different form of assurance.

The boundary: do not turn perspective modeling into demographic determinism

The paper’s business relevance depends on respecting its limits.

DiADEM is tested on two subjective NLP datasets: DICES and VOICED. Both are relevant to safety and offensiveness, but they do not cover every business domain, language, culture, modality, or organizational policy environment. The learned demographic weights are dataset-specific. They may shift with sampling, label design, region, task framing, or annotation instructions.

There is also a governance issue. Annotator metadata is sensitive. Collecting it requires consent, minimization, secure handling, and a clear purpose. A company should not collect demographic fields just because a model architecture can use them. “The neural net wanted it” is not a privacy policy.

Finally, demographic metadata is incomplete. People do not judge content only through age, race, gender, education, locale, or political affiliation. They bring personal experience, professional background, language competence, platform norms, moral commitments, fatigue, and occasionally the simple desire to finish the labeling batch before lunch. Some of that can be modeled. Some of it should not be reduced to columns.

The right conclusion is not that demographics explain disagreement. The right conclusion is that disagreement can contain structured signal, and demographic-aware modeling is one imperfect tool for recovering part of that signal.

More arguments, better systems

The industry often talks about AI alignment as if the goal is to discover one clean human preference and encode it. That framing is convenient. It is also frequently false.

In many real tasks, humans do not converge neatly. They disagree because the item is ambiguous, because the policy is underspecified, because social experience changes interpretation, or because the world has not issued a final answer for the model to imitate. DiADEM is useful because it does not rush to hide that mess. It gives the model a way to learn from it.

The paper’s deeper message is managerial: do not ask AI systems to resolve disagreement before the organization understands it. A system that predicts only the majority label may be efficient, but it is not necessarily safer, fairer, or more useful. Sometimes the most valuable output is not the answer. It is the shape of the argument.

And in a world increasingly governed by automated judgments, knowing where reasonable people disagree is not a defect in the data.

It is the data.

Notes

Cognaptus: Automate the Present, Incubate the Future.