Documentation / QueryTek Review glossary / AI Output Review
AI output review is structured human evaluation of model-generated content before it is relied on or published. It exists because generated text is fluent regardless of whether it is correct, so the usual signals readers use to judge reliability are absent.
The reason this needs its own discipline is that AI output fails differently from human output. Human error usually correlates with visible signs of difficulty — hedged phrasing, incomplete sections, awkward construction. Generated text is uniformly confident, so fluency carries no information about accuracy.
That changes what a reviewer looks for. Four failure modes dominate.
Fabrication. Plausible, specific, and false — invented citations, non-existent provisions, confident figures. This is the one fluency most effectively conceals.
Omission. What the output leaves out, which is harder to see than what it gets wrong. A summary that reads completely while dropping a material exception is more dangerous than one that is obviously partial.
Unsupported inference. Conclusions drawn beyond what the source material supports, presented in the same register as the supported ones.
Tone and policy fit. Output that is accurate but wrong for the audience, the register, or the organization's obligations.
Because the failure modes differ, the rubric differs. A rubric built for human drafting weights grammar and structure, which generated text rarely fails. A rubric for AI output weights verifiability and completeness — dimensions that require checking claims against sources rather than reading for quality.
Sampling is the practical question. Reviewing every output is often infeasible at generation volumes, so programs review a defined sample and escalate on findings. That is defensible when the sampling basis is stated and risk-weighted; it is not when "we review the output" means someone skims a fraction of it.
Organizations adopting generative tools inherit responsibility for what those tools produce. A regulator or customer harmed by a fabricated statement is not interested in which system generated it.
Structured review is also what makes an oversight claim demonstrable. "A human reviewed it" is unverifiable; a rubric, a sampling basis, and recorded findings are evidence.
QueryTek Review applies structured human evaluation to AI-generated material using rubrics weighted for verifiability and completeness, and records outcomes as review evidence. Sampling rates and rubric contents are set per engagement.
