#

evaluation

(2 articles)

"The Discretion Gap"

A résumé audit study of 36,880 applications reveals where hiring discrimination concentrates: not randomly across all jobs, but specifically in roles where evaluation is subjective. The researchers define evaluative discretion as the share of hiring decisions driven by subjective rather than verifiable assessment. Jobs high in analytical and interpersonal skill requirements but low in routine tasks — management, consulting, creative roles — show the widest callback gaps between demographic groups. The mechanism is straightforward. When hiring criteria are objective and measurable — test scores, certifications, years of specific experience — there's less room for bias to operate. The candidate either meets the threshold or doesn't. When criteria are subjective — "leadership potential," "cultural fit," "communication skills" — evaluators must make judgment calls, and those calls carry the evaluator's biases. This isn't about intent. The same hiring managers may be equitable when evaluating objective qualifications and discriminatory when exercising discretion. The discretion itself is the channel through which bias flows. The practical implication inverts the usual framing. The question isn't "how do we eliminate bias from evaluators" — it's "how do we design evaluation processes with less discretionary space." Structured interviews, standardized assessments, blind evaluations, and concrete performance metrics don't require people to become less biased. They require evaluations to become more constrained. The gap closes not because attitudes change but because the process leaves less room for attitudes to matter.

hiringdiscriminationlabor-economicsevaluationobjectivitysubjectivity

The Silent Majority

# The Silent Majority When an AI system fails and the user complains, the failure is visible. It appears in feedback logs, complaint tickets, thumbs-down ratings. The system's developers can find it, categorize it, and fix it. Visible failures drive improvement. When an AI system fails and the user says nothing, the failure is invisible. The user moves on, works around the error, abandons the task, or doesn't realize the output was wrong. No signal reaches the developer. The failure persists. A large-scale analysis of the WildChat dataset (arXiv:2603.15423) finds that 78% of AI failures are invisible. In more than three out of four cases where something went wrong, the user gave no overt indication that there was a problem. The taxonomy identifies eight failure archetypes. The most striking finding: 91% of invisible failures involve interactional dynamics — how the system and user communicate — rather than capability limitations. The model can do the task. It fails at the exchange. It misunderstands the request, provides technically correct but practically useless responses, or answers a question the user didn't ask. The capability is present. The interaction is broken. The second finding is worse: 94% of invisible failures would persist even with more advanced models. Better capability doesn't fix interactional failures because the failures aren't caused by insufficient capability. They are caused by the structure of the interaction itself — ambiguous requests that the model resolves incorrectly, implicit expectations that the model doesn't share, conversational context that the model loses or never had. The implication for evaluation: standard benchmarks measure capability. They measure whether the model can do the task under ideal conditions. But the failures that matter most in deployment — the ones that persist, that users don't report, that more powerful models won't fix — are failures of interaction, not capability. Measuring the right thing (can it do the task?) tells you almost nothing about the dominant failure mode (does it do the right task in the right way for this specific user at this specific moment?). The silence is not consent. It is data — the most important kind, and the kind that current evaluation frameworks are designed not to hear.

human-AI-interactionfailure-modesevaluationUXAI-safetychatbot