AI has become a partner in how people learn about, do, and engage with science, and the partnership takes three forms: a scientist works with a co-scientist whose output must be checked; a member of the public looks something up to decide whether a diet works or whether to fit solar panels; and a student takes up an inquiry with AI in a science class. Across all three, one thing decides whether the partnership helps or harms: whether the human evaluates what the AI returns or takes it on trust. I argue that this evaluation -- epistemic vigilance calibrated to how far a fallible source can be trusted -- is, given adequate prior knowledge, the binding constraint on productive augmentation. You can hand the AI a great deal precisely because you stay vigilant; vigilance makes generative partnership safe, so it licenses augmentation rather than restricting it. Vigilance is already invoked in science education but under-specified for the AI case; I specify its components, the mechanism tying it to learning, and a way to measure it without soliciting the evaluation it is meant to detect. What is distinctive is that the machine's fluent, confident prose reads as trustworthy whether or not it is, so its surface works against the human evaluating it. The argument bears hardest on education: the integrated conceptual knowledge instruction aims to foster forms only under deep processing, and vigilance sets how deeply a claim is processed, so it is the precondition for learning with AI. The design factors the field reports matter through whether they engage the learner's evaluation; none works around it. Untested is vigilance as a measured disposition, above all where the AI is confidently wrong. Because it is unevenly distributed, integrating AI uniformly is likely to widen the gap between better- and less-prepared students. I close on how it might be built by fading support as the learner takes over.

翻译：人工智能已成为人们学习、实践与参与科学的伙伴，这种伙伴关系表现为三种形式：科学家与需核验其输出结果的协同科学家合作；公众查阅信息以决定某种饮食是否有效或是否安装太阳能板；学生在科学课上借助AI开展探究。在这三种形式中，决定伙伴关系有益还是有害的关键因素，在于人类是否评估AI返回的内容还是盲目信任。笔者认为，这种评估——即根据不可靠来源的可信度校准的认知警惕性——在具备充分先验知识的前提下，是制约高效增强的核心约束条件。之所以能赋予AI大量自主权，恰恰是因为保持警惕；警惕性使生成式伙伴关系变得安全，因此它允许增强而非限制增强。警惕性在科学教育中虽已被提及，但在AI情境下仍定义不足；本文明确了其构成要素、将其与学习关联的机制，以及在不直接征求评估结果的前提下测量警惕性的方法。其独特之处在于，机器流畅而自信的文本无论真实与否都显得可信，这种表面特征会阻碍人类对其进行评估。这一论点对教育影响尤为显著：教学目标所指向的整合性概念知识只有在深度加工中才能形成，而警惕性决定了信息被加工的深度，因此它是与AI共同学习的先决条件。该领域报道的设计因素能否发挥作用，取决于是否激发学习者的评估行为；没有任何设计能绕过这一机制。目前尚未验证的是将警惕性作为可测量特质（尤其是在AI自信地犯错时）的效果。由于警惕性分布不均，统一整合AI教育很可能扩大学生之间的能力差距。最后，笔者探讨了如何通过逐步减少支持、让学习者自主掌控，来培养这一能力。