A Human Capability Test for Learning Outcomes in the AI Era

Synthesis: This paper argues that generative AI has broken the traditional link between submitted work and student capability, so universities can no longer assume that the quality of submitted work reliably demonstrates what a student can do. The author proposes a human capability test built around three questions: what must a student demonstrate independently, what can be strengthened through AI augmentation, and what must the student verify, defend, and take responsibility for. An engineering material-balance problem illustrates how the test operates at the level of an individual learning outcome. The framework shifts the Assessment question from whether AI was used to what capability remains the learner's own.

The Problem

Permission policies for GenAI increasingly recognize its implications for university learning, but they do not provide evidence of what students have actually learned. This breaks the traditional assumption that submitted work quality demonstrates student capability.

The Human Capability Test

The proposed test reframes Assessment around three questions: what must a student demonstrate independently; what can be strengthened through AI augmentation; and what must the student verify, defend, and take responsibility for. It is intentionally non-prescriptive and adaptable across disciplines. An engineering material-balance problem illustrates how the test operates at the level of an individual learning outcome.

What this means for practice

• Instructors. Answer the test's three questions for every learning outcome before writing the assignment: what must students demonstrate independently, what may AI augment, and what must they verify, defend, and take responsibility for.
• Instructors. Require multiple forms of evidence rather than one AI-assisted submission — an oral defense, supervised problem formulation, laboratory decision, or transfer problem each contributes, and no single format is the test.
• Assessment designers. Distribute the balance across the program: more independent demonstration in foundational courses, deliberate AI-supported synthesis and transfer in intermediate courses, and defended consequential judgments in advanced courses.
• Assessment designers. Use the test at program level for assurance of learning, mapping each outcome's independent, augmented, and accountable conditions against accreditation requirements such as ABET's expectation that students solve complex engineering problems.

Limitations

• This is a conceptual proposal, not an empirical study: no students, courses, or assessment data are reported, so the test's reliability and validity are untested.
• The framework is illustrated with a single example, an engineering material-balance problem, and the author states it is intentionally non-prescriptive, leaving no scoring rules or validated instrument for others to reuse.
• The argument rests on institutional policy reports from MIT, Brown, and the University of Toronto plus the ABET criteria, not on evidence about whether the proposed evidence formats actually distinguish independent from AI-augmented capability.

Citation

Saleh, N. B. (2026). A Human Capability Test for Learning Outcomes in the AI Era. EdArXiv preprint.