What journalists need to know about Beyond Performance

Why this research matters

Headlines often report that AI “improved” test scores, homework completion, or workplace output. This research supplies a method for testing whether that kind of improvement actually means the person got better at anything, or whether the gain disappears the moment the AI is removed.

What is new

A six-element declaration researchers must make before evaluating a claim (what functioning, what population, what scope, what contrast, what time window, what would count as failure), and a decision table that returns succeed, unproven, or fail for specific claim types — assisted performance, human learning, meaningful control, governed extension, or broad social progress.

What would be inaccurate to say

• That this paper ran the cited experiments. It did not — Bastani et al. and Strömberg et al. are independent, previously published studies the method is applied to as worked examples.
• That the method has been independently validated as reliable across evaluators or domains. It has not.
• That the paper concludes AI assistance is generally harmful to learning. It concludes the opposite is also unwarranted: a safeguarded design changed the outcome in the cited studies.
• That the method produces a single numerical score. It returns a disposition (succeed / unproven / fail) per declared claim, not a composite index.

Researcher contact

Micheal Charles Preble · Independent Researcher · micheal@perfinitive.com · ORCID

No quotations are pre-approved for press use on this page. Contact directly for comment or an approved quote.