生成式人工智慧往往会创造出一种能力的幻觉，从而降低认知参与度。超过二十年的研究显示，科技使用者可分为自动化者、验证者和 cyborgs 等类别。对使用人工智慧的学生进行脑电图（EEG）监测发现，对大多数人而言，高频伽马（gamma）脑波在数分钟内就会崩塌，将认知状态转向被动消费。然而，在人工智慧被设定为以问题和情境而非直接答案来提示使用者的实验中，主动且高伽马参与的学生比例增加了一倍以上，显示出摩擦对于激发学习是必要的。

这种学习效率的差异在2019年的一项 Harvard 研究中得到了凸显，在该研究中，与主动解决问题奋斗的学生学到了更多，但由于缺乏认知流畅度，他们回报的学习成就感较低。同样地，由 Norma Ming 于2012年至2016年间进行的一项教育研究追踪了约 60,000 名大学生和 MBA 学生，结果显示，获得最高分的学生是在讨论论坛中最常出错的人。相反地，90%的学生（十分之九，nine out of ten）选择了安全、防御性的参与，以确保及格而未参与更深层的学习。

人工智慧基准测试衡量单一模型的能力，无法反映人机协作的动态。欧盟《人工智慧法案》第14条等政策框架错误地依赖人类监督来防止错误，这往往会导致自动化偏见。为了解决这个问题，像是欧盟的人工智慧办公室和美国的 CAISI 等测试机构，应该使用混合智慧指数来评估人机协同效应。评估这些互动至关重要，因为能够主动挑战流畅机器的人与被动盲目批准输出的人之间的经济差距将随著时间推移而加剧。

Generative artificial intelligence often creates an illusion of competence that can reduce cognitive engagement. Over two decades of research shows that technology users diverge into categories such as automators, validators, and cyborgs. Electroencephalogram (EEG) monitoring of students using AI reveals that for most, high-frequency gamma brain waves collapse within minutes, shifting cognitive state to passive consumption. However, in experiments where the AI was programmed to prompt users with questions and context rather than direct answers, the ratio of active, high-gamma engaged students more than doubled, showing that friction is necessary to stimulate learning.

This discrepancy in learning efficiency is highlighted by a 2019 Harvard study where students who struggled with active problem-solving learned more but reported feeling less successful due to the lack of cognitive fluency. Similarly, an education research study conducted by Norma Ming between 2012 and 2016 tracking approximately 60,000 undergraduate and MBA students revealed that top marks were achieved by those who were most frequently wrong in discussion forums. Conversely, 90% of students (nine out of ten) chose safe, defensive participation to secure passing grades without engaging in deeper learning.

AI benchmarks measuring isolated model capabilities fail to reflect human-AI collaboration dynamics. Policy frameworks like Article 14 of the EU’s AI Act mistakenly rely on human oversight to prevent errors, which frequently results in automation bias. To address this, testing bodies like the EU’s AI Office and America’s CAISI should evaluate human-machine synergy using a Hybrid Intelligence Index. Evaluating these interactions is vital, as the economic divide between individuals capable of actively challenging fluent machines and those who passively rubber-stamp outputs will compound over time.