When you score an AI separately for each type of user, does it reveal winners and losers that one average hides?

This explores whether scoring a system separately for each kind of user, instead of averaging everyone into one number, exposes trade-offs like helping one group while failing another, which a single aggregate would bury. The corpus backs the premise but has no study that directly shows those trade-offs with numbers. What it has is strong circumstantial evidence, a structural argument, and some warnings about the cost.

The most direct support is the argument behind MatrAIx. It says outcome-only benchmarks abstract away how diverse users phrase requests and judge results, so it keeps user variation separate from fixed scoring, using a population of 8.3 billion persona records (Can simulated users reveal what offline benchmarks miss?). A neighboring note shows how badly a single number can mislead. Judge pass rates rose from 23.1 to 80.0 percent while task-facing defect detection stayed exactly where it was (Can a higher evaluation score hide poor task performance?). That case is about optimizing against the metric, not about user variation. Still, the shape is the same: the score climbs while the thing you care about doesn't. The reward-hacking note names the common cause as optimization against signals that only partly represent the real task (Does reward hacking always stem from the same failure?). An average over all users is one of those partial signals.

The corpus also gives a structural reason to expect the trade-offs to be real. PReF treats each person's preferences as a set of coefficients over a shared pool of base reward functions, and ten well-chosen questions are enough to pin those coefficients down (Can user preferences be learned from just ten questions?). My reading, not the note's claim, is that a single aggregate reward is then just one particular blend of those coefficients. A model tuned to it serves the average blend, and users whose coefficients sit elsewhere pay for that. The trade-off comes with the setup, and only per-user scoring makes it visible.

Aggregates aren't useless, though. Chatbot Arena's 240K+ crowdsourced votes agree with expert raters because the questions are diverse and discriminating (Can crowdsourced votes reliably rank language models?). That makes the average good at saying which model is better overall, but it can't say better for whom. Per-user tracks answer a different question rather than replacing that one. One caution comes from the A/B-testing work: behavior-based personas predict which way a test goes 75 to 90 percent of the time, but are least trustworthy when the true effect is near zero (Can behavior-based personas predict A/B test outcomes?). A flat average is exactly where you can't tell whether nothing is happening or opposing effects are cancelling out. That is where simulated per-user tracks are the least reliable.

The last caveat is that splitting scores by user doesn't make the hard evaluation problems go away. The trajectory-level evaluation note argues that comparability, reproducibility, and the link from evidence to judgment relocate into a higher-dimensional space instead of disappearing, and that shared protocols matter more than the format (Do interactive evaluations actually solve the benchmark comparison problem?). Reusing one persona population across surveys, chatbots, and other interfaces is a practical step toward repeatable per-user comparison (Can one persona population evaluate different application types?). So the answer is probably yes in principle, and the corpus explains why. Whether the hidden trade-offs are meaningful in practice is still unshown here, because nothing in it runs the same system through per-user and aggregate scoring side by side.