Work

Goodhart's Law in the Classroom and the Algorithm: Why Static Benchmarks Fail Both Human Learners and Artificial Intelligence

Year: 2026
Type: preprint
Abstract: Standardised assessments and artificial intelligence benchmarks can produce high performance scores without necessarily demonstrating an equivalent ability to apply knowledge under changed conditions. This conceptual working paper examines that measurement problem through a comparison of secondary-school examinations and large language model mathematical reasoning benchmarks.Drawing on CBSE Circular Acad-30/2024, the GSM-Symbolic research presented at ICLR 2025, and scholarship on performance measurement, the paper considers how familiarity with predictable question structures may complicate the interpretation of assessment results. It distinguishes human learning from machine evaluation while identifying a shared concern: whether a measured score adequately captures performance when relevant assumptions or problem conditions change.The paper proposes three practical assessment modifications: controlled re-parameterisation, assumption reversal, and explanation-based error diagnosis. It also outlines a potential classroom evaluation framework addressing transfer, assessment validity, accessibility, and marking reliability.The proposed modifications are intended for further investigation rather than presented as experimentally validated improvements.Publication status: Independent, non-peer-reviewed conceptual working paper (November 2026). No original classroom experiments, participant studies, or model benchmark tests were conducted. AI-assisted drafting and editorial support are disclosed in the manuscript.
Source: Figshare
Language: English
Subfield: Education
Open Access status: green
OpenAlex ID: W7234094648
Cites:
Cited by:
Related to: 10