User Feedback on Generative AI: Three Measured Flaws

A multi-phase study filed on 2 October 2026

On 2 October 2026, at 00:39 UTC, four researchers filed on arXiv a piece of work on the feedback users send to generative AI systems. The text carries the names of Alisa Frik, Julia Bernd, Amitis Karami and Mohammad Tahaei, and comes out of a collaboration between academic research and eBay (arXiv:2610.02631^[1]).

It is an empirical study of AI systems that starts from an interface detail and arrives at the quality of measurement. The core of the work is a benchmark of current industry approaches: the authors compared the real interfaces through which a person reports a model error. Public rules call for user involvement after release, while practical guidance on how to design those mechanisms remains scarce.

The outcome is a trio of recurring flaws. On the basis of those findings the group wrote best-practice recommendations, then built and tested a prototype collection tool. The same version of the paper remains readable on alphaXiv^[2].

• Hard-to-discover feedback mechanisms
• Obscure terminology in the labels offered to the user
• Little attention to the value returned to whoever writes the report

The channel few users find

Hard discovery of the mechanism weighs before anything else: a channel few people see collects few voices.

The flaw counts as a sampling problem. Whoever reaches the hidden button is a self-selected minority, driven by strong irritation or technical curiosity. The team thus receives the tail of the distribution and reads that tail as the voice of the public.

The consequence touches the budget. A group measuring perceived quality through an invisible channel pays for staff, dashboards and meetings in exchange for a signal that describes a handful of extreme users.

The paper places this flaw among those common to the products examined. The exact frequency in the sample stays outside the public abstract. A more insistent prompt raises the volume of responses and leaves the question of their representativeness open.

Words the user reads another way

The second flaw concerns the vocabulary of the labels. «Helpful», «wrong», «inappropriate», «off-topic»: each item carries a different meaning in the head of the person clicking and in the table of the person analyzing.

An ambiguous label ruins the signal upstream of the pipeline. Two users press the same button for two distinct reasons, and the count merges the two cases into a single metric. That number rises and falls for reasons the analyst struggles to reconstruct.

Here terminology is a measurement choice, before it is an interface choice.

The authors classify the problem among those common to the products observed. Their alternative taxonomy proposal lives in the full text, outside the public summary.

A clear label costs one glossary sentence and returns categories comparable over time. The value sits in the stability of the data, more than in the richness of the menu.

The value for whoever writes the report

The third flaw is the subtlest: the mechanisms examined pay little attention to what the user gets back in exchange for their time.

Writing a useful comment takes attention, memory of the context and a few minutes. The return, in the products observed, stays opaque: the report enters a silence. A user learns fast and stops.

The authors tie effective feedback to two promises made to the person: a quick and flexible gesture, and an experience that leaves a positive feeling. The prototype under test builds on this axis.

On the opposite side sits the product team's requirement, which calls for rich data in a usable format. The tension between the two requirements is the real object of the design.

The paper describes the collaboration with eBay as the frame of the test, on a consumer product with real traffic. The scale of that test stays outside the abstract.

The prototype, and what the abstract leaves unsaid

The study arrives at an artifact: a collection tool designed, built and put to the test. The abstract states the tool's objectives and keeps the outcome figures out.

This reticence deserves a note on method. The 2 October 2026 filing weighs 7,655 KB and contains the full text. The public summary carries zero samples, zero intervals, zero measured deltas.

Four authors, one industry partner and a tested prototype point to an applied study. The weight of the evidence lives inside the intermediate phases, so serious reading goes through the PDF.

How far the result extends beyond the eBay case remains an open question, and the authors leave it open. Carrying their conclusion into a different domain requires another study.

Caution counts double here: a qualitative benchmark of products ranks the flaws, instead of quantifying their impact.

Feedback as a measurement signal

The subject touches the measurement of models, beyond the graphics of a button. Human judgment is the signal that has trained and corrected generative systems for years.

The work on InstructGPT, published in 2022, documented a 1.3-billion-parameter model preferred by human raters over a 175-billion one (doi:10.52202/068431-2011^[3]). The human signal weighed more than the scale of the parameters. A signal collected from a handful of extreme users, with ambiguous labels, enters the model with the same weight.

This desk has spent months measuring the distance between evaluation and production. AgentGym2 scores 46.15 on real tasks against idealized scores, and FinProBench brings the comparison onto 1,723 professional deliverables.

The behavior under examination makes the picture worse. Broken promises rise from 9% to 87% when the model's belief about what the judge rewards changes, and Science measured sycophancy 49% higher across eleven models. A badly designed feedback channel adds noise to an already fragile measurement chain.

Where it touches budget allocation

For an investment committee the useful reading concerns the spending line on measurement, more than the line on the model.

The evaluation market races toward 9.57 billion and sells automation as the arbiter. JuryProbe shows that a panel of model judges shares its blind spots, so it works as an instrument with systematic bias. Feedback from real users remains the complementary source, and this paper measures how fragile it is in its current design.

For a head of data the question turns infrastructural: where the reports end up, under which schema, with which link to the conversation episode that generated them. Feedback stripped of technical context counts as a survey.

For a board the thesis under examination is the one about continuous improvement driven by users. The 2 October 2026 work describes it as a practice to redesign, before describing it as a practice acquired.