Why do online ratings fail to represent independent individual preferences?
This explores why the star ratings you see online don't add up to a clean signal of what each person independently thinks — and the corpus has a lot to say, because the distortions come from several directions at once.
This explores why online ratings fail to represent independent individual preferences — and the short version from the corpus is that a rating is never really independent: it's shaped by who came before it, who chose to show up, and how moody the rater was that day.
Start with social contagion. When researchers pulled ratings apart into baseline quality, social influence, and noise, they found that prior ratings meaningfully push later ones — the score you give is partly an echo of the scores already on the page, and that echo compounds as it feeds future raters Do online ratings actually reflect independent customer opinions?. Layered on top is a selection problem: only people who already expected to like a product buy it and bother to review, so the pool of raters is filtered twice before anyone clicks a star. What looks like 'product quality' is really the satisfaction of self-selected buyers, and summary statistics can even slow down honest quality discovery Do online reviews actually measure product quality or just buyer preferences?.
Then there's the instability of the individual rater themselves. The same person, rating the same item in different sessions, can shift by multiple stars — driven by mood, anchoring, and personal rating style rather than any change in preference Why do the same users rate items differently each time?. So even a single rating mixes 'what I think of this' with 'how I happen to rate things' and 'what I just saw rated before me.' A deeper cut from the alignment world makes this concrete: responses don't all measure the same thing. Some are genuine preferences, some are non-attitudes (no real opinion, answered anyway), and some are preferences constructed on the spot by the question itself — and you can only tell them apart by checking consistency across conditions Do all annotation responses measure the same underlying thing?.
Here's the part you might not have come looking for: the platform isn't a neutral mirror, it's an active participant. Different recommender types route different audiences to the same product, which changes whether its ratings converge or diverge — the system decides who rates what Do different recommender types shape opinion convergence differently?. Ranking systems must explicitly model this selection bias or they collapse into feedback loops that amplify their own past choices Why do ranking systems need to model selection bias explicitly?, and undersized embeddings quietly tilt everything toward already-popular items, starving niche tastes of the exposure they'd need to register Does embedding dimensionality secretly drive popularity bias in recommenders?. So 'independent individual preference' is being bent by the recommender before the rating is even entered.
The payoff insight: this isn't just a measurement nuisance, it's a statistical fact about the data. Preference data is not independent-and-identically-distributed across raters — each person carries their own distribution, so getting a real signal depends on the number of distinct raters, not just the volume of ratings Does preference data need more raters than examples?. And the fix can backfire: when you personalize hard enough to escape the noisy average, you lose the averaging that kept sycophancy and echo chambers in check Does personalizing reward models amplify user echo chambers?. Independence, in other words, is something these systems systematically dissolve — by design.
Sources 9 notes
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.
Only consumers expecting satisfaction purchase and review, creating two selection filters. Research shows early reviewers shape later perceptions, altruism affects learnability, and summary statistics can actually slow quality discovery. Observed ratings misrepresent the satisfaction distribution of all potential buyers.
Amatriain et al. found that the same user gives substantially different ratings to the same item across sessions, shifting by multiple stars. This noise stems from temporal inconsistency, rater-specific biases, and anchoring effects—making ratings reflect both preference and rating-behavior rather than stable preference alone.
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.
YouTube's multi-objective ranker uses MMoE for conflicting objectives and a shallow position tower to remove selection bias from training data. Without both mechanisms, models converge on degenerate equilibria that amplify their own past decisions.
Research shows that when user/item embedding dimensions are too small, recommender systems overfit toward popular items to maximize ranking quality. This compounds over time as niche items receive insufficient exposure, and cannot be fixed post-hoc without treating dimensionality as a fairness hyperparameter.
Preference data is not i.i.d. across raters with different preferences. PAC bounds for personalized reward models decompose into terms depending on both examples per rater and number of raters, showing rater diversity matters as much as data volume.
Specializing reward models per user removes the averaging effect of aggregate models, allowing systems to learn sycophancy and reinforce polarization at scale, mirroring recommender-system failures.