Demonstrable Understanding Under AI

Definition

Demonstrable understanding under AI is the requirement that a person explain, defend, adapt, and evaluate an artifact when fluent written production no longer reliably shows who understands or originated it.

Current Synthesis

The episode applies this problem to mathematics. A deep, correct paper once supplied strong evidence that its writer understood the relevant ideas; generative systems weaken that inference by enabling people to submit sophisticated work they may not be able to explain. Assessment therefore shifts toward oral examination, discussion, lecture, live defense, and questions that expose reasoning, while preserving written work as evidence rather than treating it as self-authenticating proof of mastery.

This is not an anti-AI rule. AI-assisted output can be useful and correct, but institutions still need to distinguish artifact quality, personal understanding, contribution, and responsibility. Those are related judgments, not one score.

Key Claims

• Polished or technically correct output is no longer sufficient evidence of personal mastery.
• Oral defense and live explanation can test whether a person can give reasons, answer challenges, and transfer an idea to a new case.
• Artifact quality, authorship, originality, understanding, and responsibility should be evaluated separately.
• Hiring and graduate education need transparent norms for acceptable AI assistance and evidence of contribution.
• Human understanding remains necessary for judging importance, usefulness, and explanatory value even when formal verification is available.

Evidence

• Broken output-understanding inference - What it means when AI helps solve complex math problems says generative AI lets a person produce work that experts judge correct without demonstrating deep understanding.
• Assessment response - What it means when AI helps solve complex math problems predicts more weight on oral examinations, discussions, lectures, explanation, and defense.
• Institutional stakes - What it means when AI helps solve complex math problems connects the distinction to future mathematics hiring and graduate training.
• Mathematical boundary - What it means when AI helps solve complex math problems distinguishes machine verification from human interpretation and integration.

Counterevidence & Qualifications

Oral performance can be affected by language, disability, anxiety, status, and evaluator bias, and it does not by itself prove originality or exclude concealed assistance. Written work remains valuable for careful reasoning, revision, and durable communication. A credible assessment system should combine declared tool use, process evidence, artifact review, live explanation, and task-appropriate accommodations rather than replacing one single signal with another.

What Changed

• Created the concept to separate output quality from personal understanding, authorship, and responsibility.
• Added oral defense as one useful but incomplete evidence channel.

Related Concepts

• AI University Assessment Reform - broader institutional redesign of assessment under routine AI use.
• Knowledge Externalization Tradeoff / 知识外化权衡 - distinguishes access to external output from internal mastery.
• AI-Generated Proof Governance - applies the distinction to mathematical acceptance, credit, and integration.
• Human Judgment Under AI - retains evaluation and responsibility around machine assistance.
• Research Taste - judgment needed to identify valuable questions and results.
• Examination Systems - wider tradeoff among depth, fairness, standardization, and scale.