Does the U-shaped distribution of raters compound the negativity bias from public posting?
This explores whether two separate distortions in online ratings — the bimodal 'only the delighted and the furious bother to rate' selection effect, and the tendency of people to skew negative when posting publicly — stack on top of each other to push ratings down further than either would alone.
This reads the question as bundling two distinct mechanisms: a selection effect (who shows up to rate at all, the U-shape) and a behavioral effect (how those people shift their scores once an audience is watching). The corpus doesn't have a single note on the U-shaped distribution itself, but it has sharp material on the second mechanism and on how rating distortions compound — enough to answer the spirit of the question, which is really 'do these biases reinforce each other?'
The public-posting negativity effect is well-documented here. Why do online reviewers publish negative ratings despite positive experiences? shows people *lower* their ratings in public even after a genuinely positive experience — not because they felt worse, but because negative reviewers read as more intelligent, so there's a self-presentational payoff to going negative. Crucially, private raters show no such shift. That last detail matters for your question: the negativity is an audience effect layered on top of whatever opinion the rater actually held. So if the pool of raters is already skewed toward the extremes (the U-shape), the public-posting effect doesn't just add a fixed amount of negativity — it preferentially recruits and rewards the negative tail.
The compounding piece is where the corpus gets interesting. Do online ratings actually reflect independent customer opinions? decomposes ratings into baseline quality, social-dynamics influence, and noise, and finds prior ratings measurably drag subsequent ones — a feedback loop where today's distorted average becomes tomorrow's anchor. So even a small public-posting nudge isn't a one-time tax; it propagates forward. Notably, the same work finds that *high opinion variance eventually dampens the distortion* — which is the U-shape cutting the other way. A genuinely polarized (high-variance) rater pool can actually break the social-dynamics spiral rather than feed it. So the two effects don't cleanly compound; they can partially cancel.
Underneath all of this is the question of whether raters are even measuring the same thing. Do all annotation responses measure the same underlying thing? separates genuine preferences from 'non-attitudes' and constructed-on-the-spot preferences. A U-shaped distribution may partly be an artifact of who has a genuine attitude strong enough to act on — the indifferent middle simply has no signal to report. And Do different recommender types shape opinion convergence differently? adds that different products attract different audience segments with different prior expectations, meaning the shape of the rater pool isn't fixed — it's manufactured by how people arrive at the product in the first place.
Worth knowing the mirror image too: the same negativity that humans amplify in public, language models systematically *suppress*. Why do LLMs generate polite reviews even when users hated products? and Can user history override an LLM's politeness bias in reviews? show RLHF-trained models default to polite, inflated reviews and need fine-tuning plus real user history to produce an honestly negative one. So as AI-generated reviews enter these pools, they may push the distribution back toward the bland middle — a counter-pressure to both the U-shape and the public-negativity effect, and a sign that the rating ecosystem's biases are about to get a new, opposite-signed input.
Sources 6 notes
Posters systematically reduce their ratings in public when exposed to negative reviews, even with positive personal experience—because negative reviewers appear more intelligent. Private raters show no such shift, revealing a self-presentational mechanism tied to multiple-audience communication.
Moe and Trusov decomposed ratings into baseline quality, social-dynamics influence, and error, finding that prior ratings meaningfully affect subsequent ones. These effects have both immediate sales impact and long-term compounding effects through future ratings, though high opinion variance can eventually dampen the distortion.
Behavioral science reveals that annotations contain genuine preferences, non-attitudes, and constructed preferences—distinguishable by consistency across measurement conditions. Treating them uniformly contaminates reward model training and downstream alignment.
Research shows that frequently-bought-together and co-viewed recommendation networks produce different opinion convergence patterns. The mechanism: each recommender type attracts different audience segments with different prior expectations, shaping both who sees products together and how they rate them.
Off-the-shelf LLMs generate inappropriately positive reviews due to alignment-training politeness bias. Combining user review history, rating signals as satisfaction indicators, and supervised fine-tuning successfully redirects the model to generate negative reviews when warranted.
Review-LLM defeats the politeness bias inherent in RLHF-trained models by aggregating user behavior sequences (prior reviews, item ratings) in the prompt and fine-tuning on these contextualized examples. This dual intervention—personalized context plus explicit satisfaction signals—allows the model to generate authentically negative reviews matching user dissatisfaction.