AI as a Partner in Learning about, Doing, and Engaging with Science: Vigilance as the Key to Productive Augmentation

Synthesis: Argues that epistemic vigilance — the human evaluation of AI output calibrated to how far a fallible source can be trusted — is the binding constraint on productive augmentation. AI's fluent, confident prose reads as trustworthy whether or not it is, making evaluation harder. Vigilance sets how deeply a claim is processed and is thus the precondition for learning with AI. Design factors (prompts, feedback, scaffolding) matter only through whether they engage the learner's evaluation. Because vigilance is unevenly distributed, uniform AI integration risks widening achievement gaps.

Key Findings

• The paper identifies epistemic vigilance — the human evaluation of AI output calibrated to how far a fallible source can be trusted — as, given adequate prior knowledge, the binding constraint on productive augmentation in learning with AI.
• The AI partnership takes three forms — the scientist working with a co-scientist, a member of the public checking a claim such as whether a diet works or whether to fit solar panels, and a student taking up an inquiry with AI inside a science class — and in all three the deciding factor is whether the human evaluates what the AI returns or takes it on trust.
• Vigilance is what licenses augmentation: because the human stays vigilant, generation, retrieval, and drafting can be delegated safely, so vigilance expands rather than restricts what can be handed to the AI.
• The AI case is distinctive because the machine's fluent, confident prose reads as trustworthy whether or not it is, so the default surface of the output works against the human doing the evaluating.
• Vigilance sets how deeply a claim is processed, making calibrated vigilance the precondition for productive learning with AI; design factors such as prompts, feedback, and scaffolding matter through whether they engage the learner's evaluation, and none works around it.
• Candidates that might seem to make vigilance dispensable — the learner's own content knowledge, a neighboring competence, or a more trustworthy AI — do not remove the need for it.
• Because the disposition to evaluate is unevenly distributed, integrating AI uniformly across a classroom is likely to widen achievement gaps.

The Argument

The paper specifies the components of vigilance, the mechanism that ties it to learning, and a way to measure it without soliciting the very evaluation it is meant to detect. It also distinguishes judging from producing: each capacity is built by exercising it, so what is handed over to the AI is never the exercise the lesson exists to provide — a learner who evaluates a derivation deeply is practicing judgment, not derivation. Existing evidence anchors the processing-depth half of the claim; what remains untested is vigilance as a measured disposition, above all in the regime where the AI is confidently wrong.

What this means for practice

• Instructors. Keep the evaluation with the learner. Generation, retrieval, and drafting can be handed over, but never the judgment the lesson exists to teach: a learner who evaluates a derivation is practicing judgment, not deriving.
• Make the check go wide, not just deep enough to follow the steps. Ask students to test the AI's answer against their prior knowledge and to explain why it holds, since that processing is where the learning happens.
• Treat Critical Thinking and Hallucination Risk awareness as central to instructional design rather than peripheral: prompts, feedback, and scaffolding work only insofar as they engage the learner's evaluation of the output.
• Aim Scaffolding and differentiated support at the disposition to evaluate rather than at tool access, because vigilance is unevenly distributed and uniform AI integration is likely to widen the gap between better- and less-prepared students.
• Fade the support as learners take the evaluation over, so that the disposition to evaluate is built rather than assumed.

Limitations

• This is a theoretical account, not a study: there is no sample, no intervention, and no data — the argument is developed from existing cognitive and science-education literature.
• The author states that what remains untested is vigilance as a measured disposition, above all in the regime where the AI is confidently wrong; the proposed way to measure vigilance without soliciting the very evaluation it is meant to detect is offered, not validated.
• Only one half of the claim is anchored in existing evidence — that vigilance governs how deeply a claim is processed — while the causal claim that vigilance is the binding constraint on productive augmentation rests on argument.
• The equity prediction that uniform AI integration widens achievement gaps follows from the argued uneven distribution of the disposition rather than from any measurement of it.

Citation

Marcus Kubsch (2026). AI as a Partner in Learning about, Doing, and Engaging with Science: Vigilance as the Key to Productive Augmentation. arXiv preprint (physics.ed-ph).