1. 定义：当读者阅读时，“认知负荷”究竟是什么意思#
不是把你压垮的负荷——而是把你建起来的负荷。
教育心理学里有一个区分，很少进入大众关于阅读的讨论，但下面的论证必须依赖它。在技术意义上，John Sweller 在认知负荷理论里确立的“认知负荷”，并不是一个单一的东西。1 当读者遇到困难文本时，至少有两种认知劳动在同时运行，而它们与“阅读最终产出什么”之间的关系完全不同。
第一种是噪音。它是读者花在与糟糕句子对抗上的努力：一句话写得别扭、指代含糊，导致你反复读一段；作者没有筛选与组织，让你不得不在工作记忆里同时抱着太多分心的想法。Sweller 把这称为 extraneous load（外在负荷/无关负荷）——不是材料本身要求你做的工作，而是材料的呈现方式强加给你的工作；它会消耗工作记忆，却不建起任何耐用的东西。1 教学设计者几十年来一直在研究怎么降低它。他们是对的。
第二种，是当材料本身很复杂时，真正“理解”所需要的工作——你把前提放在工作记忆里，同时追踪“第二章前半段的一个判断如何限制了此刻这段话的结论”；你注意到作者在证据尚未齐备时就先抛出了结论；你感受到两条尚未解决的论证之间的张力。Sweller 的框架把这称为：对内在负荷的 germane processing（建构性加工/促成性加工）——把认知资源投入到长期记忆中的图式（schema）建构上；理解不是“被递送过来”的，而是“被建起来”的。1
这两者不是一回事。我们关心的是第二种。不是对困难的偏见，也不是把“吃苦”浪漫化成目的本身——而是一个精确的区分：某一种认知工作是否在完成某一种“形成”，以及当工具把它移走时会发生什么。
Maryanne Wolf 早已指出，深度阅读不是被动接收。2 她说它像一场“精心编排的马戏表演”——左右半球、多个脑叶、整个大脑的多层结构同时参与：建构图像、整合推断、权衡批判性的替代解释。工作本身，就是那个“东西”。
2. 读者脑中握着的模型（以及被告知应该握着的模型）#
主流共识有研究。主流共识有机制解释。直到最近，主流共识几乎没有遇到过强有力的反论证。
大多数审慎的知识工作者对 AI 与阅读的看法，并不天真。大体上，它有上一节引用的教育心理学文献作支撑。区分“有益/无益认知工作”的同一套框架，也在三十多年的教学研究里反复记录：降低外在负荷会提升学习。当学生被糟糕的组织、丑陋的排版、无关的枝节搞得迷糊时，把这些东西清理掉，会把工作记忆腾出来给理解用。从这个角度看，清理文本、压缩文本的 AI 工具，做的正是好的教学设计会做的事：减少噪音，让信号变得可达。
这种看法背后还有一项规模可观、同行评审的随机对照试验（RCT）。2025 年，Kestin 等人在哈佛发表了一项随机对照试验（N=194），比较 AI 辅导与传统课堂主动学习在物理教学上的效果。3 AI 辅导带来的增益超过两倍——后测分数中位数 4.5 对 3.5，p<10⁻⁸——而且用时更少：49 分钟对 60 分钟。这是一个大效应、严谨设计、并且发表在 Nature 体系期刊的结果。了解这项研究的读者，有充分理由认为：这套共识不是“有头衔的人随便说说”。
这套共识逻辑自洽，有证据支持。而在另一个认知领域里，已经有一件“共识恰好反过来”的事情被记录了下来。
3. 空间认知文献早就研究过的案例#
当一个人通过开车在城市里穿行来学会导航时，大脑会通过一张特定的网络来编码环境：海马旁回皮层（parahippocampal cortex）与扣带后区/脾周皮层（retrosplenial cortex）形成的是“路线层面”的表征——地标与顺序线索，让你能够沿着同一条路再走一遍。4 但如果你是通过研究一张地图来学习这座城市，另一套网络会更显著地被调动起来——当学习者建立一种 allocentric（以环境为中心的）“概览式（survey-level）表征”时，下额回（inferior frontal gyrus）的激活更强。4 2012 年 Zhang、Copara 与 Ekstrom 的 fMRI 研究把这件事说得非常明确：路线学习与地图学习形成的“基于神经的表征”，依赖的是“部分可分离的脑系统”。4
这不是两条不同路线通往同一个目的地；这是两种不同的认知成就。路线学习让你能够“再次走回同一条路”。地图式学习让你能够在熟悉空间里规划新的路径——因为“概览式表征”让你能计划那些你从未走过的路线。
依赖一种模式、放弃另一种模式，其行为后果是可测量的。Dahmani 与 Bohbot 在 2020 年的研究，考察了人们习惯性使用 GPS 与空间记忆之间的纵向关系。5 在他们的横断样本（N=50）中，GPS 依赖越重，就越少依赖以海马为基础的空间记忆策略，越多依赖刺激—反应式的导航——也就是“跟着走”，而不是“建地图”。纵向部分样本很小（N=13），这一点必须说清——但他们发现：“自第一次测试以来 GPS 使用更频繁的人，其认知制图能力下降得更快”。5 导航工具不只是替你换了一条路；它还与它所替代的能力的“渐进性萎缩”相关。
N=13 并不是一个大的纵向样本。选择效应也无法完全排除——空间制图能力本就在下降的人，可能更倾向使用 GPS，而不是 GPS 导致下降。设计中的被试内成分部分缓解了这个问题，但不能彻底消除它。这个发现是方向性的，而且与 Zhang 等人“网络分离”的机制图景一致——但它还不是一个已经坐实的因果主张。
如果同样的区分也成立于论证理解呢？——“为一个论证建图”与“沿着别人已经建好的路线导航”，会不会也是两件不同的事？
4. 深度阅读究竟对阅读大脑做了什么#
大脑并不是“天生就会读”的。这不是比喻；这是 Stanislas Dehaene 的“神经元再利用（neuronal recycling）”假说的锚点发现。视觉词形区（visual word form area，位于枕颞皮层）是从腹侧视觉通路“挪用”出来的——在文盲个体中，这套网络主要用于面孔与物体识别。6 阅读习得通过长时间练习，把这一区域招募来完成字母与词的识别，而且这个过程具有可观测的竞争性：一项 2018 年的纵向 fMRI 研究发现，“阅读习得并没有替换掉那些最初反应，但阻断了它们的发展”。6 阅读大脑更像一次“翻修工程”，而不是从零开始的建造。
阅读回路是通过练习获得的——这意味着它可能发展得并不充分，也可能在练习中断时退化。底层基底不是固定不变的。建起来的东西，可以被维护，也可以被放任变弱。
Wolf 对这个回路在“正常运转”时到底做什么的描述，则补上了功能维度。她写道，深度阅读是一场“精心编排的马戏表演”——左右半球、多个脑叶、整个大脑的多层结构同时参与；这个过程把“背景知识与共情”整合起来，也把“推断与批判性分析”整合起来。2 对文本的不同可能解释会“来回移动”；读者把备选解释悬置在张力之中、权衡它们，并构造出一个属于读者自己的回应——它建立在读者长期累积的模型之上。
这不是一个“被动接收系统”的描述；这是一个“主动建构过程”的描述，而它确实要付出相当高的认知成本。
回路是后天习得的；它是主动建构的；而且它会因为不用而退化——Wolf 自己就经历过这一点，结尾我们会回到它。这些都把一个问题推到了读者现成的工作模型面前：如果这个回路是通过使用而建起来的，那么阅读里究竟是哪一部分在“做建造”？
5. 反转：真正做功的是哪一种负荷#
当你在工作记忆里抱着一条长论证的前提，同时追踪“第三章的一个主张如何依赖第一章的一个主张”，与此同时第七章引入的反驳还悬在那里等待评估——这种认知劳动，并不是阅读的低效；它就是“形成”本身。
在 Sweller 2019 年对认知负荷理论的再表述里，germane processing 并不是叠加在内在负荷之上的另一种“额外负荷”。1 它是把工作记忆资源“转向”对内在负荷的加工——而这种加工，具体来说，就是在长期记忆中建构耐用图式（schema）。追踪一条长论证、抱住它的依赖关系、感受逻辑在哪里扣不住——这就是对“高元素交互（high-element-interactivity）”内在负荷的建构性加工。这里每一个词都有分量：高元素交互，意味着大量相互依赖的元素必须被同时托住——不仅“难”，而且“结构上复杂”，这些元素在整篇文本跨度里彼此交织；内在，意味着复杂度属于材料本身，而不是糟糕呈现方式带来的；建构性加工，意味着工作记忆被用来做图式建构：把一个耐用的心智模型“建在长期记忆里”。
这才是值得被捍卫的那种特定认知工作。不是一般意义上的“困难”。也不是外在负荷——糟糕的组织、分心的排版、没必要的术语。这里并不是在为“写得烂的书”辩护。这里要捍卫的是：读者在几百页跨度里“随着论证展开而建一张心智地图”的劳动——所有依赖、张力、反向压力都保持原样。
AI 总结并不是在减少外在负荷。那样做当然很好，也几乎不会引发争议。AI 总结往往是在把建构性加工整段拿掉：它把已经完成的图式——那张“地图”——交付给你，却省略了“制图”的工作。读者拿到了建造过程的产出，却没有运行建造过程本身。用 Sweller 的术语说：图式是在外部被生产出来并递送给学习者的，绕开了本应在内部把它建起来的那份工作记忆投入。1
空间认知文献里的“制图 vs 导航”类比，让这个区分变得直观——但请注意，这里是类比，不是直接的实验结论。神经科学证据在空间领域；而即便在空间领域，纵向行为数据样本也很小（前面已经承认）。把它应用到论证理解，是一种建立在“可被记录的平行关系”之上的推断，而不是由它直接证明。空间领域的案例，帮我们把阅读回路文献（Wolf、Dehaene）更含糊描述的东西讲清：建构一个空间的表征，与沿着既有路径走一遍，是两件不同的事；差异会体现在“之后大脑还能做什么”上。
认知卸载（cognitive offloading）文献在“记忆层面”补充的一点是：被卸载掉的东西，就不会被内部编码。7 Risko 与 Gilbert 的综述记录了：当人们把信息交给外部工具存放时，就更不可能投入内部资源去保留它。但这个发现谈的是“记忆”，不是“能力形成”。空间认知案例更能承载“形成”这一层的重量——它讨论的不是“记住了什么”，而是当你选择不同参与模式时，某种能力被建起来或没被建起来。
读者沿着 AI 构造出来的论证摘要“走一遍”，并没有为这个论证建起一张认知地图；读者只是穿过了一条别人已经画好的路线。而那个真正建过地图的人——扛住论证的负荷、追踪依赖、感受逻辑紧绷处——能够在没有预制路线时也去导航新的论证。这不是同一种认知成就换了不同效率的包装；这是两种完全不同的认知成就。
AI 拿走的那种负荷，恰恰是做功的负荷。
如果最强的反证只是“缺乏研究”，这篇文章会是另一种样子。但事实不是。
6. 第一条 steelman：AI 导师在“交付”上胜过人类教学#
Kestin 2025 年的 RCT 值得被认真对待，而不是只被当成一个“顺手一提”。
这项研究很严谨：N=194，哈佛本科生，随机分组，领域是物理。AI 辅导组的后测中位数是 4.5，而课堂主动学习组是 3.5；差异达到 p<10⁻⁸。3 学习发生在 49 分钟而不是 60 分钟。按教育研究的任何标准来看，这个效应量都很大。这不是一项可以被轻描淡写、被最小化、被挥手略过的研究。
而它也确实没有得到这样的对待。Kestin 测量的是一种非常具体的东西，而这种具体性很重要：结果评估落在 Bloom 分类学的第 2 到第 4 层——理解、应用、分析——材料是第一次接触的；场景是一节 60 分钟的单次课程；领域是结构化的（入门物理）。3 这测的是教学的“传递/交付功能（transfer function）”：学生是否学会了这节课递送的内容？这是一个重要问题，而 AI 辅导给出了令人印象深刻的答案。
但“形成/塑造功能（formation function）”是另一个问题：这种参与是否会建起一种能力——在不熟悉的领域里评估新论证的能力？在没有结构化课堂、没有 AI 引导、当读者独自面对一篇没有老师、没有标准答案的文本时，这种能力是否会被建起来？Kestin 的研究没有测量这一点，作者也直接承认了这一空白。论文自己的说法是：这些收益“可能无法泛化到多个概念的复杂综合与更高阶的批判性思维”。3 作者暗示的那条缝——“传递功能（学会这次内容）”与“形成功能（建起一种结构性的评估能力，使你能够普遍地判断论证质量）”——正是值得被继续展开的区分。
这不是挪球门。本文的核心主张从来不是“AI 不擅长交付内容”。Kestin 反而表明：在受控条件下，AI 非常擅长交付内容，甚至比人类教学更强。但深度阅读复杂论证文本所独特建起的，并不是“内容交付”。
Kestin 是更“容易”的那条反证。还有更难的一条。
7. 更难的 steelman：苏格拉底式 AI 在专家脚手架内发展批判性思维#
Kao、Grant 与 Woltering 的 2025 年预印本，描述了一项三组随机对照试验（N=90，高中十年级科学课）：对照组、没有 AI 的 Argument-Driven Inquiry（ADI）框架组、以及引入 AI 的 ADI 组（使用配置为苏格拉底式对话的 ChatGPT Study Mode）。8 AI-苏格拉底组的学生，在科学论证与批判性思维上的提升显著高于另外两组。研究自己对其意义的表述是：“首次给出实验性证据，表明苏格拉底式 AI 导师能够在真实课堂中增强青少年的推理与参与度。”8
这不是“内容交付”的发现；这是“论证与批判性思维”的发现。AI 在这里并不是在一个结构化课堂里给本科生讲物理；它在促进探究、提出问题、把假设端上台面、要求学生评估证据并考虑反驳。结果指标测量的是技能——论证、批判性思维——它们更接近本文要讨论的那种“制图能力”。了解这项研究的读者会追问：如果苏格拉底式 AI 能在课堂里发展批判性思维，那为什么同样的参与不能发展出“独立评估论证”的能力？
这个问题值得被回答，而不是被驳回。答案在结构上。
Kao 的研究并没有测试“当读者独自面对文本时会发生什么”；它测试的是：当读者身处课堂，AI 由已经知道这份认知工作形状的人来设计时，会发生什么。关键结构特征在于：AI-苏格拉底条件并不是“学生随便用 ChatGPT”。ChatGPT 的 Study Mode 被学习科学家、教学法专家与教育者在 ADI 框架内进行了配置——这是一个结构化的教学系统，有明确目标、定义清晰的能力项、以及专家设计的进阶路径。8 苏格拉底式对话被脚手架化、结构化，并被框架内嵌的专业知识引导。学生是在这种专业知识的“在场”中，发展出论证与批判性思维技能的。
“制图”的区分（前面已承认这只是类比，但它很精确）指出了这里没有被测试的东西：在没有专家脚手架的情况下，面对一个陌生、非结构化的论证，读者能否建起这份论证的认知地图——读者必须自己发现结构、识别依赖关系、自己决定评估准则，因为没有人替你设计路线。按这个类比，脚手架化的苏格拉底式对话，更像是在专家设计的地图上进行结构化的路线导航：旅程被策展过，学生沿着设计者认定“有效”的路径前进。本文要捍卫的能力——为一段陌生论证自己建图——并不是 Kao 的研究设计所测试的对象。
这是一个论证，而不是一条被报道的实验结论；区分这一点很重要。本文要捍卫的能力，并没有像 Kestin 与 Kao 所用的量表那样、一个已经 RCT 验证过的测量工具。底层基底是明确的——Wolf、Dehaene、Sweller 以及阅读回路文献已经把地基铺好。可直接的行为测试——重度 AI 摘要使用者在面对新材料时，是否表现出更弱的“论证建图能力”，相对于匹配的深度阅读者——尚未被做过。这类测试需要一个能测“制图能力”的工具，而这种工具目前还不存在为一个被验证的量表。
Kao 测试的是：在专家设计的探究路径里，脚手架化的苏格拉底式 AI 能为学生做什么。而我们讨论的能力，是当“没有人替你提问”时读者能建起什么——结构要自己发现，依赖要自己映射，评估准则要在没有预设框架的情况下自己确立。这是两种不同的认知成就。区分是真实的，而目前还没有研究同时测试了两者。
8. 真正诚实的承认，以及在承认之后仍然成立的东西#
目前没有研究直接测试：成人读者习惯性使用 AI 摘要，是否会削弱他们为陌生、非结构化论证建认知地图的能力。本文的主张，是一个“汇聚式推断”（convergent inference）：四条证据线，各自承载一部分，靠汇聚去承载任何单条证据线都承载不起的重量。
Sweller 命名了这份工作：对高元素交互内在负荷的建构性加工，会在长期记忆中建构图式——这一机制在三十年的教学设计研究中被反复建立。1 Wolf 与 Dehaene 命名了基底：一种后天获得的神经重布线，它在建构中是主动的、跨区域的，并且既会因使用而发展，也会因不用而退化。26 机制需要作用对象；阅读回路文献表明：这个基底存在，而且它不是固定不变的。空间认知案例给出了一个已被记录的平行关系：在另一个领域、另一种工具下，习惯性 GPS 使用与可测量的认知制图能力下降相关；当这种模式出现在类比两端时，就更难把它当成巧合。45 Risko 与 Gilbert 提供了更一般的框架——被卸载的认知不会被内部编码。7 但他们的发现谈的是“记忆”，不是“形成”，本文不会把重量压在它一条上。
这些碎片是拼得起来的。推断是多来源、跨领域、汇聚式的：AI 总结移除了那份在长期记忆中建构论证图式的建构性加工；而读者的深度阅读回路正是靠做这份工作建起来的；空间认知案例显示，在一个平行领域里，类似的“工具替代”与能力衰退相关。没有任何单一研究能把这一点钉死；在论证理解这一领域里，直接测试尚未发生。
而这恰恰是智识严肃性有时所要求的：在足够多的领域里指认已知，使得就我们手里的证据而言，这种推断是最合理的解读。
9. 为什么这件事现在尤其重要：一个早已被削弱的基底#
当下有三种条件同时出现，而在以往任何一次媒介转型里都没有同时出现过。
此前的转变——印刷、广播、电视、互联网——改变的是文本如何流通、哪些文本能抵达读者。它们各自以不同方式侵蚀注意力。但没有一种技术宣称要替代“阅读这件事本身”。AI 是这一谱系里第一种公开说：我可以提取信息、压缩论证、把结论交付给你。阅读的“传递/交付功能”（也就是大多数成年人绝大多数时候阅读复杂文本的理由）如今第一次在规模上变得可被替代。当“传递”被外部处理时，“形成/塑造”就成为更具相对价值的那种认知工作——因为“形成”正是 AI 无法替读者完成的那一部分。
与此同时，迎接这场替代的那群人，并不是以满血状态抵达的。
NAEP 2023 年长期趋势评估发现：13 岁孩子的阅读成绩自 2020 年以来下降了 4 分，过去十年累计下降 7 分，大致回到了 1971 年的基线——抹去了半个世纪的表面进步。9 美国时间使用调查（American Time Use Survey）覆盖 2003 到 2023 年的 236,270 名受访者，显示：在任意一天里“为娱乐而阅读”的美国成年人比例从 28% 降到 16%——二十年里相对下降 43%。10
在青少年群体中，Jean Twenge 对 Monitoring the Future 调查数据的分析显示了一条持续塌陷的轨迹：在 1970 年代末，大约 7% 的 12 年级学生“为了娱乐一本书都不读”；到 2021–22 年，这一比例上升到 33%，甚至在计划读研究生的学生中也是如此。11 这里的数字来自 Twenge 对调查的二次解读，而不是直接的原始表格统计；但它与周边指标指向同一方向。
在挪威，对征兵数据做的家庭内分析发现：IQ 分数在 1975 出生队列达到峰值，随后在 1991 出生队列之前一路下降——其原因来自环境，而不是遗传。12 更广义的“一般认知能力”在一个研究充分的国家人口中，已经在一个如今迈入中年的出生队列上呈现下行。IQ 不是阅读能力本身，这个差异很重要——这里提供它，是作为背景，不是要把它当作同一主张。
这些都不是 AI 造成的。ATUS 的窗口（2003–2023）基本处在 AI 普及之前。挪威 IQ 的下降发生在 1975–1991 出生队列上。青少年阅读的塌陷从 1980 年代就开始了。AI 没有造成这些。AI 是落在一个“防御已经被削弱的人群”身上的加速器——它不是始作俑者，却是在最糟糕时刻到来的新压力测试。
NAEP 回到了 1971。ATUS 在二十年里几乎腰斩。计划读研的那群人里，“一本书都不读”的比例自 1976 年以来增长了三倍。
关于“下滑”的证据并不是本文的中心主张，但它解释了为什么中心主张会在此刻变得紧迫。
10. 练习上的含义：一个开放问题#
Wolf 自己经历过退化。
在多年高强度的数字化阅读之后，她说自己失去了与复杂、持续性散文相处所需的“认知耐心（cognitive patience）”。她描述自己在“表面、而且非常快”地读；“事实上，我读得太快，以至于无法理解更深的层次”。2 她发现自己无法忍受去重读的那本书，是黑塞的 Magister Ludi。Wolf 很清楚发生了什么，因为她一生都在研究阅读大脑。她也知道该做什么：两周的刻意、专注的深度阅读练习，把这种能力恢复了回来。2
这里的主张不是“不可逆的认知损伤”。Wolf 描述的退化——以及她实现的恢复——指向的是一种更精确、也更可行动的理解：深度阅读是一种练习，它会建起并维护一种能力；当练习中断时，能力会变弱。至少对那些已经建起阅读回路的成年人而言，这种中断是可恢复的。两周的刻意投入，就把 Wolf 在多年数字阅读中被侵蚀掉的东西带了回来。
关键区分在于：能力 vs 练习。Wolf 的快速恢复暗示：退化的是练习习惯，而不是底层神经基底。回路还在；练习断了。对已经建起阅读回路的成年人而言，这是个好消息。重度 AI 摘要使用造成的练习缺口，不一定是结构性的神经损失；它更像是“练习的损失”。而练习的损失，是可恢复的。
对于那些回路尚未建成的孩子，情况不同——那将属于另一套论证：关于儿童期的阅读教学与形成性发展。对成人读者而言，这里的信息不是末日，而是诊断。
从这一点并不能直接推出处方。我们没有一个被验证的协议、没有循证的“剂量”、没有同行评审的“训练方案”，来维持通过深度阅读形成的论证评估能力。我们拥有的，是 Wolf 关于退化与恢复的第一人称记录，是 Wolf 与 Dehaene 建立起来的神经科学基底，以及 Sweller 命名“哪一种认知工作是有生产力的”的框架。从这些材料里，我们能看到一个练习上的含义，尽管它还无法被精确规定：能力靠使用来维持；使用就是做那份认知工作；而当你导航的是摘要、而不是亲自建构论证时，那份认知工作就没有发生。
不是对屏幕的道德化批判。不是对纸质书的怀旧。也不是把处方伪装成证据。把这些剥离之后，剩下的是一个值得一直握着的问题。
AI 在信息交付上非常强。这句让步是真诚的，而且之所以反复强调，是因为它确实需要被承认。问题不是“要不要用 AI”。问题是：哪些认知工作必须由读者自己完成——读者选择承担哪一种负荷——而这种选择会在时间尺度上建起什么、或者建不起什么。
在这个框架里，AI 拿走的负荷不是一个该被消灭的麻烦；它是一种练习——每一次读者面对复杂论证时，是做还是不做，全看你是否把那份工作交给了会主动替你做的工具。
那么，什么练习值得我们去做？
References#
Sweller, J.(2010)."Element interactivity and intrinsic, extraneous, and germane cognitive load." Educational Psychology Review 22(2): 123–138.https://link.springer.com/article/10.1007/s10648-010-9128-5; Sweller, J., van Merriënboer, J. J. G.,&Paas, F.(2019)."Cognitive Architecture and Instructional Design: 20 Years Later." Educational Psychology Review 31: 261–292.https://link.springer.com/article/10.1007/s10648-019-09465-5↩↩↩↩↩↩
Wolf, M.(2018).Reader, Come Home: The Reading Brain in a Digital World. Harper Collins. — verified via Slate review and WBUR interview.↩↩↩↩↩
Kestin, G.et al.(2025)."AI tutoring outperforms in-class active learning: an RCT introducing a novel research-based design in an authentic educational setting." Scientific Reports. PMC12179260.https://pmc.ncbi.nlm.nih.gov/articles/PMC12179260/↩↩↩↩
Zhang, H., Copara, M.,&Ekstrom, A. D.(2012)."Differential Recruitment of Brain Networks following Route and Cartographic Map Learning of Spatial Environments." PLOS ONE. PMC3445610.https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0044886↩↩↩
Dahmani, L.,&Bohbot, V. D.(2020)."Habitual use of GPS negatively impacts spatial memory during self-guided navigation." Scientific Reports 10: 6310.https://www.nature.com/articles/s41598-020-62877-0↩↩↩
Dehaene, S.,&Dehaene-Lambertz, G.(2018)."From the Processing of Digits to the Processing of Words: The Neural Recycling Hypothesis." Trends in Cognitive Sciences.https://doi.org/10.1016/j.tics.2018.04.005 and associated VWFA longitudinal evidence cited in-text.↩↩↩
Risko, E. F.,&Gilbert, S. J.(2016)."Cognitive offloading." Trends in Cognitive Sciences 20(9): 676–688.https://doi.org/10.1016/j.tics.2016.07.002↩↩
Kao, C., Grant, L.,&Woltering, S.(2025)."Socratic AI tutors enhance adolescents'reasoning and engagement in real classrooms." Research Square preprint. DOI: 10.21203/rs.3.rs-8118546/v1.https://www.researchsquare.com/article/rs-8118546/v1↩↩↩
National Center for Education Statistics.(2023).NAEP Long-Term Trend Assessment Results: Reading and Mathematics (Age 13).https://www.nationsreportcard.gov/highlights/ltt/2023/↩
Bone, J. K.et al.(2025)."The decline in reading for pleasure over 20 years of the American Time Use Survey." iScience(Cell Press). PMC12496190.https://pmc.ncbi.nlm.nih.gov/articles/PMC12496190/↩
Twenge, J. M.(2023)."Are books dead? Why Gen Z doesn't read." Generation Tech(Substack), citing Monitoring the Future Survey data (University of Michigan/NIDA).https://www.generationtechblog.com/p/are-books-dead-why-gen-z-doesnt-read — As Twenge's secondary analysis of MtF data; MtF primary data not directly fetched.↩
Bratsberg, B.&Rogeberg, O.(2018)."Flynn effect and its reversal are both environmentally caused." PNAS 115(26): 6674–6678. PMC6042097.https://pmc.ncbi.nlm.nih.gov/articles/PMC6042097/↩
Further Reading#
Paas, F.&van Merriënboer, J. J. G.(2020)."Cognitive-Load Theory: Methods to Manage Working Memory Load in the Learning of Complex Tasks." Current Directions in Psychological Science 29(4): 394–398.https://journals.sagepub.com/doi/full/10.1177/0963721420922183 — 认知负荷理论的方法论侧；为“有益 vs 无益负荷”的区分提供底座，并澄清为什么本文捍卫的是“内在负荷的建构性加工”，而不是一般意义上的困难。
Wolf, M.(2007).Proust and the Squid: The Story and Science of the Reading Brain. Harper. — Wolf 更早、更技术性的著作，解释阅读大脑如何在习得中被建起来；为本文关于“回路可塑性”的主张提供发育学基底。
Carr, N.(2010).The Shallows: What the Internet Is Doing to Our Brains. W. W. Norton. — 对互联网时代注意力碎片化的经典论述；为第 9 节关于“AI 不同于以往媒介转变”的论证提供前 AI 背景。
Shumailov, I.et al.(2024)."AI models collapse when trained on recursively generated data." Nature 631(8022): 755–759.https://pubmed.ncbi.nlm.nih.gov/39048682/ — 当 AI 摘要在规模上替代人类“推理密度高”的文本生产时，可能引发供给侧的二阶后果：递归训练导致的模型崩塌。
Kosmyna, N.et al.(2025)."Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task." arXiv preprint 2506.08872.https://arxiv.org/abs/2506.08872 — 在写作任务中使用 LLM 辅助时神经参与度下降的 EEG 证据（预印本、小样本，且已有方法学批评）；方向上与本文机制一致，但不足以单独承重。
简述（In brief）#
认知负荷并不是单一概念。Sweller 的框架区分了外在负荷——糟糕文字、含混指代、分心版式带来的噪音——与对内在负荷的建构性加工：工作记忆的投入把耐用图式建进长期记忆。教学设计几十年来一直在降低前者；而后者是另一种动物：它不是“读难论证的低效”，它是读者评估论证能力被建起来的机制。
关于 AI 与阅读的主流立场，确实依托一个真实发现。Kestin 2025 年的哈佛 RCT 显示：AI 辅导在更短时间内让学习增益超过课堂主动学习两倍，p 低于 10⁻⁸。结果大、严谨，无法被挥手略过。但它测量的是结构化领域里 Bloom 2–4 层级的内容交付——教学的传递功能。Kao 的三组 RCT 把问题推得更硬：在学习科学家搭建的脚手架内，苏格拉底式 AI 能提升十年级课堂中的批判性思维。两条发现都是真的。但两者都没有测试：当没有人替你设计路线时，读者独自面对一段长、陌生、非结构化论证，究竟能建起什么。
空间认知文献早已研究过一个平行案例。Zhang 与 Ekstrom 的 fMRI 研究记录了两套可分离网络：路线学习编码顺序线索；地图学习建构 allocentric 的概览式表征。Dahmani 与 Bohbot 的纵向数据样本不大、方向性强于定论，但它把更重 GPS 使用与认知制图能力更陡的下降联系起来。Wolf 把深度阅读称为“精心编排的马戏表演”；Dehaene 显示阅读回路是被“再利用”的神经地产，靠使用建成，也会因不用而退化。Risko 与 Gilbert 提供卸载框架：被委派出去的东西不会被内部编码。
没有任何单一研究能确立本文的中心主张。这里的论证是一个汇聚式推断：Sweller 命名机制，Wolf 与 Dehaene 命名基底，空间认知案例命名已被记录的平行模式。AI 总结并不是在减少外在负荷（那当然欢迎）；它是在把建构性加工整段拿掉：把完成的图式交付给你，却省略制图工作。读者拿到地图，却没有亲手画过地图。
之所以必须在此刻讨论，是因为 AI 并不是落在“中性地面”上。NAEP 的 13 岁阅读成绩回到 1971 基线；ATUS 显示成人“为娱乐而读”在 2003–2023 之间几乎腰斩；Twenge 的分析显示 12 年级学生中“零本娱乐阅读”的比例从约 7% 上升到约 33%；挪威征兵数据则显示 IQ 在 1975 出生队列见顶后下行。这些都不是 AI 造成的。AI 是落在一个早已被削弱的基底上的加速器。
Wolf 自己失去了阅读黑塞所需的认知耐心；两周刻意练习把能力带了回来。能力靠使用建起、靠使用维持、也靠使用恢复——至少对已经建好回路的成年人如此。没有处方可给。最后剩下的是一个问题：当工具主动提出“我替你扛”，每一次你选择让它扛走哪一种负荷时，你究竟在时间里建起了什么，或者让什么没能建起来？
The Load That Does the Work
1. Definition: what cognitive load actually means when a reader reads#
Not the load that overwhelms — the load that builds.
There is a distinction in educational psychology that rarely makes it into popular conversations about reading, but what follows depends on it. Cognitive load, in the technical senseJohn Sweller established in cognitive load theory, is not a single thing.1 When a reader encounters difficult text, at least two kinds of cognitive labor are running simultaneously, and they have very different relationships to what that reading produces.
The first kind is noise. It is the effort spent fighting a badly constructed sentence, re-reading a paragraph whose referents are ambiguous, holding too many distracting ideas in mind because the writer did not curate them. Sweller calls this extraneous load — work imposed not by the material itself but by the presentation of the material, work that consumes working memory without building anything durable.1 Instructional designers have spent decades finding ways to reduce it. They are right to.
The second kind is the work of comprehension itself when the material is complex — holding premises in working memory while tracking how a claim two chapters back conditions a claim being made right now, noticing when a conclusion is stated before all its evidence has been assembled, feeling the tension between two arguments that have not yet resolved. Sweller's framework names this germane processing of intrinsic load: the cognitive investment that constructs schema in long-term memory, the work through which understanding is not received but built.1
These are not the same thing. The second kind is what concerns us here. Not a prejudice against difficulty, not a romanticization of effort for its own sake — a precise distinction. The question is whether a specific kind of cognitive work is doing a specific kind of formation, and what happens when a tool removes it.
Deep reading,Maryanne Wolf has established, is not passive reception.2 It is an"elaborately choreographed circus performance" — both hemispheres, multiple lobes, all layers of the brain simultaneously engaged, constructing images and integrating inference and weighing critical alternatives. The work itself is the thing.
2. The model the reader holds (and has been told to hold)#
The consensus has a study. The consensus has a mechanism. The consensus has, until recently, had nothing seriously arguing against it.
The position most thoughtful knowledge workers hold about AI and reading is not naive. It is, in broad outline, well-supported by the educational psychology literature the previous section invoked. The same framework that distinguishes productive from unproductive cognitive work has also documented, over three decades of instructional research, that reducing extraneous load improves learning. When a student is confused by poor organization, ugly typography, or irrelevant tangents, clearing that away frees working memory for comprehension. The AI tools that clean up and compress a text are, by this account, doing exactly what good instructional design does. They reduce noise. They make the signal accessible.
This position has a large, peer-reviewed RCT behind it. In 2025, Kestin and colleagues published results from a randomized controlled trial at Harvard (N=194) comparing AI tutoring against traditional in-class active learning for physics instruction.3 AI tutoring produced more than double the gains — median post-test score 4.5 versus 3.5, p<10 to the−8 — in less time: 49 minutes versus 60. This is a large effect, a rigorous design, and a nature-portfolio journal publication. The reader who knows this study has good reasons to think the consensus is not merely credentialed opinion.
The consensus is coherent. It has support. And there is a documented case, in a different cognitive domain, where the consensus turns out to be exactly the wrong way around.
3. The case the spatial-cognition literature has already studied#
When someone learns to navigate a city by driving its streets, the brain encodes the environment through a particular network: parahippocampal cortex and retrosplenial cortex register route-level representations, the landmarks and sequential cues that allow a traveler to follow the same path again.4 When someone learns the same city by studying a cartographic map, a different network comes online — inferior frontal gyrus shows greater activation, as the learner builds an allocentric, survey-level representation of the space.4 The 2012 Zhang, Copara, and Ekstrom fMRI study documented this precisely: "neural-based representations formed following route and map learning rely on partially dissociable brain systems."4
These are not two routes to the same destination. They are two different cognitive achievements. Route learning builds the ability to re-navigate the same path. Cartographic learning builds the ability to navigate new paths through known space — because the survey representation allows the learner to plan routes that were never traveled.
The behavioral consequence of relying on one mode while foregoing the other is measurable. Dahmani and Bohbot's 2020 study examined habitual GPS use and its longitudinal relationship to spatial memory.5 In their cross-sectional sample (N=50), heavier GPS reliance was associated with reduced dependence on hippocampus-based spatial memory strategies and greater reliance on stimulus-response navigation — which is route-following without map-building. The longitudinal arm — small (N=13), a limitation worth naming — found that"those who used GPS more extensively since the initial visit exhibited a steeper decline in their cognitive mapping abilities."5 The navigation tool did not merely route users differently; it was associated with the progressive atrophy of the capacity it substituted.
N=13 is not a large longitudinal sample. Selection effects cannot be fully ruled out — people whose spatial mapping is already declining may gravitate toward more GPS use, rather than GPS use causing the decline. The within-subject component of the design partially addresses this, but it does not eliminate it. The finding is directional and corroborated by the mechanistic picture from the Zhang et al. network-dissociation study — not a settled causal claim.
What if the same distinction holds in argument comprehension — building the map of an argument versus navigating one that has been built?
4. What deep reading actually does to the reading brain#
The brain was not born to read. This is not a metaphor; it is the finding that anchorsStanislas Dehaene's neuronal recycling hypothesis. The visual word form area — located in the occipitotemporal cortex — is repurposed from the ventral visual pathway, a network that in illiterate individuals is devoted to face and object recognition.6 Reading acquisition recruits this region for letter and word identification through extended practice, and the process is demonstrably competitive: a 2018 longitudinal fMRI study found that"reading acquisition did not displace those initial responses but blocked their development."6 The reading brain is a renovation project, not a construction from scratch.
The reading circuit is acquired through practice — which means it can fail to develop fully, or deteriorate when practice lapses. The substrate is not fixed. What is built can be maintained or allowed to weaken.
Wolf's account of what this circuit does when it is working adds the functional dimension. Deep reading is, she writes, an"elaborately choreographed circus performance" — both hemispheres, multiple lobes, all the brain's layers simultaneously engaged in a process that integrates"background knowledge with empathy and inference with critical analysis."2 Different possible interpretations of the text"move back and forth";the reader holds alternatives in tension, weighs them, and constructs a response that is the reader's own, built from the reader's own accumulated models. This is not a description of a passive reception system. It is a description of an active construction process running at considerable cognitive cost.
The circuit is acquired, it is active, and it deteriorates from disuse — Wolf herself experienced this, and the closing will return to it. All of which raises a question the reader's working model has not yet had to answer: if the circuit is built by use, what specifically about reading is doing the building?
5. The inversion: the load that does the work#
The cognitive work of holding a long argument's premises in working memory while tracking how a claim in chapter three depends on a claim from chapter one, while a counter-argument introduced in chapter seven is waiting to be assessed — this is not the inefficiency of reading. It is the formation.
In Sweller's 2019 reformulation of cognitive load theory, germane processing is not a separate additive load stacked on top of intrinsic load.1 It is a redirection of working memory resources toward the processing of intrinsic load — the kind of processing, specifically, that constructs durable schema in long-term memory. The cognitive work of tracking a long argument, holding its dependencies, feeling where the logic does not close — this is germane processing of high-element-interactivity intrinsic load. Each of those words carries weight.High-element-interactivity means many interdependent elements held simultaneously — not just difficult, but structurally complex, elements woven together across the full span of the text.Intrinsic means the complexity belongs to the material itself, not to bad presentation.Germane processing means working memory redirected toward schema construction: the building of a durable mental model in long-term memory.
This is the specific kind of cognitive work worth defending. Not difficulty in general. Not extraneous load — badly organized prose, distracting layout, needless jargon. This is not a defense of poorly written books. It is a defense of this: the cognitive labor of building a mental map of an argument as it unfolds across hundreds of pages, with all the dependencies and tensions and counter-pressures intact.
AI summarization does not reduce extraneous load. That would be good and largely uncontroversial. AI summarization eliminates the germane processing entirely. It delivers the completed schema — the map — without the cartographic work. The reader receives the output of the construction process without running the construction process. In Sweller's terms, the schema is produced externally and delivered to the learner, bypassing the working-memory investment that would have built it internally.1
The cartography-vs-navigation analogy from the spatial-cognition literature makes this distinction vivid — but this is an analogy, not a finding. The neuroscientific evidence is in the spatial domain — and even there the behavioral longitudinal leg is a small sample, conceded as such earlier — and the application to argument comprehension is an inferential move grounded in the documented parallel, not established by it. The spatial case makes legible what the reading-circuit literature (Wolf, Dehaene) describes more vaguely: there is a difference between building a representation of a space and following a path through one, and the difference shows up in what the brain can do afterward.
What the cognitive offloading literature adds — at the memory level — is that what is offloaded is not internally encoded.7 Risko and Gilbert's review documented that when people store information through external tools, they are less likely to devote internal resources to retaining it. But this finding is about memory specifically, not about skill formation. The spatial cognition case carries more of that weight — it documents not just what is remembered but what capacity is built or not built depending on which mode of engagement was practiced.
The reader who navigates an AI-constructed summary of an argument has not built a cognitive map of that argument; the reader has traversed a path someone else mapped. The reader who built the map — who held the argument under its own load, tracked the dependencies, felt where the logic strained — can navigate new arguments without a pre-built path. These are not the same cognitive achievement, dressed up in different efficiencies. They are different cognitive achievements entirely.
The load that AI removes is the load that does the work.
All of which would be one kind of essay if the strongest available counter-evidence were silence. It is not.
Kestin's 2025 RCT deserves more than a preview.
The study is rigorous: N=194 Harvard undergraduates, randomized, with physics as the domain. AI tutoring produced a median post-test score of 4.5 versus 3.5 for in-class active learning, a difference that reached p<10 to the−8.3 The learning happened in 49 minutes rather than 60. The effect size is large by any standard in educational research. This is not a study to dismiss, minimize, or wave past.
It gets none of that. What Kestin measured was specific, and the specificity matters: outcomes were assessed at Bloom's Taxonomy levels 2 through 4 — understanding, applying, analyzing — for material encountered for the first time, in a single 60-minute session, in a structured domain (introductory physics).3 These are the transfer-function outcomes of instruction: did students learn the content delivered in this session? That is an important question and AI tutoring answered it impressively.
The formation function is a different question: does this engagement build the capacity to evaluate novel arguments in unfamiliar domains, without a structured session, without an AI guide, when the reader is alone with a text that has no teacher and no answer key? Kestin's study did not measure this, and the authors acknowledged the gap directly. The paper's own language: the benefits"may not generalize to complex synthesis of multiple concepts and higher-order critical thinking."3 What the authors gestured at — between the transfer function (learning this content) and the formation function (building the structural evaluation capacity that enables assessment of argument quality in general) — is the distinction worth developing.
This distinction is not a goalpost move. The central claim was never that AI is bad at content delivery. Kestin shows AI is excellent at content delivery, better than human instruction under controlled conditions. Content delivery is not what deep reading of complex argumentative text uniquely builds.
Kestin is the easier version of the counter-evidence. There is a harder one.
7. The harder steelman: Socratic AI develops critical thinking — in expert-designed scaffolding#
Kao, Grant, and Woltering's 2025 preprint describes a three-arm randomized controlled trial (N=90,10th-grade science) comparing a control condition, an Argument-Driven Inquiry framework without AI, and an AI-powered ADI condition using ChatGPT's Study Mode configured for Socratic dialogue.8 Students in the AI-Socratic condition showed significantly greater gains in scientific argumentation and critical thinking than either of the other two groups. The study's own description of its significance: "the first experimental evidence that Socratic AI tutors can enhance adolescents'reasoning and engagement in real classrooms."8
This is not a content-delivery finding. This is an argumentation and critical thinking finding. The AI was not teaching physics to undergraduates in a structured session; it was facilitating inquiry, posing questions, surfacing assumptions, asking students to evaluate evidence and consider counter-arguments. The outcome measures assessed skills — argumentation, critical thinking — that land much closer to the cartographic capacity at the center of this argument. The reader who knows this study will press: if Socratic AI can develop critical thinking in a classroom, why wouldn't the same engagement develop the capacity for independent argument evaluation?
The question deserves an answer, not a dismissal. The answer is structural.
Kao's study did not test what happens when the reader is alone with the text; it tested what happens when the reader is in a classroom with an AI designed by people who already know the shape of the cognitive work. The key structural feature: the AI-Socratic condition was not ChatGPT used autonomously. ChatGPT's Study Mode was configured by learning scientists, pedagogical experts, and educators within the Argument-Driven Inquiry framework — a structured pedagogical system with specific objectives, defined competencies, and expert-designed progression.8 The Socratic dialogue was scaffolded, structured, and guided by the expertise embedded in the framework. Students developed argumentation and critical thinking skills in the presence of that expertise.
The cartography distinction — an analogy, conceded as such earlier, but a precise one — specifies what was not tested: the capacity to build a cognitive map of a novel, unstructured argument encountered without expert scaffolding, in which the reader must discover the structure, identify the dependencies, and determine the evaluative criteria themselves, because no one has designed the route. Scaffolded Socratic dialogue is, in the analogy's terms, structured route navigation through an expert-designed map. The journey has been curated; the student navigates paths already identified by the designers as productive. The capacity worth defending — building one's own map of a strange argument — is not what the Kao study's design tested.
This is an argument, not a reported finding, and the distinction matters. The capacity being defended has no RCT-validated measurement instrument equivalent to the ones Kestin and Kao used. The substrate is established — Wolf, Dehaene, Sweller, the reading-circuit literature lay out the ground. The direct behavioral test — do heavy AI-summary users show weaker argument-mapping on novel material than matched deep readers? — has not been conducted. That test would need an instrument measuring cartographic capacity, and that instrument does not yet exist as a validated tool.
Kao tested what scaffolded Socratic AI can do for students guided through expert-designed inquiry. The capacity at issue is what a reader builds when no one is asking the questions — when the structure must be discovered, the dependencies mapped, the evaluative criteria determined without a framework provided in advance. These are different cognitive achievements. The distinction is real, and no study yet has tested both.
8. The honest concession, and what survives it#
No study has directly tested whether habitual AI-summary use degrades the capacity to build a cognitive map of a novel, unstructured argument in adult readers. The claim is a convergent inference — four sources, each bearing a piece, the convergence carrying what no single source can carry alone.
Sweller names the work: germane processing of high-element-interactivity intrinsic load constructs schema in long-term memory, a mechanism established across thirty years of instructional design research.1 Wolf and Dehaene name the substrate: an acquired neural rewiring, active in construction, multi-region, subject to both development through use and deterioration through disuse.26 The mechanism needs something to act on; the reading-circuit literature establishes that the substrate is there — and that it is not fixed. The spatial-cognition case names the documented parallel: when habitual GPS use correlates with measurable decline in cognitive mapping ability, in a different domain with a different tool, the pattern across the analogy becomes harder to dismiss as coincidence.45 Risko and Gilbert add the most general frame — offloaded cognition is not internally encoded7 — though their finding is about memory, not formation specifically, and the argument does not rest weight on it.
The pieces fit. The inference is multi-source, multi-domain, convergent. AI summarization removes the germane processing that constructs argument schema in long-term memory, in a reader whose deep reading circuit was built by doing that work, in a pattern the spatial-cognition case shows playing out in a parallel domain. No single study establishes this. The direct test in the argument-comprehension domain has not been done.
That is what intellectual seriousness sometimes requires: name what is known across enough domains that the inference is the most plausible reading of what we have.
9. Why this matters now: the pre-weakened substrate#
Three conditions co-occur now that have not co-occurred in any prior media transition.
Prior shifts — print, radio, television, the web — changed how texts circulated and which ones reached readers. Each eroded attention in its own way. None proposed to substitute for the act of reading itself. AI is the first technology in the lineage that says: I can extract the information, compress the argument, and deliver the conclusion. The transfer function of reading — the reason most adults engage with complex texts most of the time — is now substitutable at scale. When transfer is handled externally, formation becomes the differentially more valuable cognitive work, because formation is the specific thing AI cannot perform on the reader's behalf.
Meanwhile, the population arriving at this substitution is not arriving at full strength.
NAEP's 2023 long-term trend assessment found that 13-year-old reading scores had declined 4 points since 2020 and 7 points over the preceding decade, returning to approximately the 1971 baseline — erasing a half-century of apparent progress.9 The American Time Use Survey, across 236,270 respondents from 2003 to 2023, found that the proportion of US adults reading for pleasure on any given day fell from 28 percent to 16 percent: a 43 percent relative decline over twenty years.10
Among adolescents, Jean Twenge's analysis of Monitoring the Future Survey data shows a trajectory of sustained collapse: the proportion of 12th graders reading zero books for pleasure rose from approximately 7 percent in the late 1970s to 33 percent by 2021-22, even among students planning graduate school.11 The figure is Twenge's secondary reading of the survey rather than a direct tabulation, and it points the same direction as the measures that bracket it.
In Norway, within-family analysis of military conscript data found that IQ scores peaked with the 1975 birth cohort and declined through the 1991 cohort — environmental in origin, not genetic.12 General cognitive capacity, broadly measured, has been declining in a well-studied national population since a birth cohort now in middle age. IQ is not reading specifically, and that difference matters — the pattern is offered as context, not as the same claim.
None of this was caused by AI. The ATUS window (2003-2023) is largely pre-AI. The Norwegian IQ decline ran 1975-1991. The adolescent reading collapse began in the 1980s. AI did not cause this. AI is the accelerant arriving on a population whose defenses are already weakened — not the originating force, but the new stress test arriving at the worst possible moment.
NAEP returned to 1971. ATUS halved in twenty years. The graduate-school-track proportion reading zero books tripled since 1976.
The body of evidence about decline is not the central claim here, but it is the reason the central claim arrives when it does.
10. The practice implication: an open question#
Wolf herself experienced the deterioration.
After years of intensive digital reading, she reports having lost the"cognitive patience"to engage with complex sustained prose. She describes reading"on the surface and very quickly; in fact, I read too fast to comprehend deeper levels."2 The text she found herself unable to bear rereading was Hesse's Magister Ludi. Wolf knew exactly what was happening because she had spent her career studying the reading brain. She also knew what to do: deliberate, concentrated reading practice, sustained over two weeks, restored the capacity.2
The claim here is not irreversible cognitive damage. The deterioration Wolf describes — and the recovery she achieved — points toward something more precise and, for this reader, more actionable: deep reading is a practice that builds and maintains a capacity, and the capacity weakens when the practice lapses. The lapse, at least for adults who have already built the circuit, is recoverable. Two weeks of intentional engagement returned to Wolf what years of digital reading had eroded.
The distinction that matters is between capacity and practice. Wolf's rapid recovery suggests what degraded was the practice routine, not the underlying neural substrate. The circuit was intact; the practice had lapsed. For adults who have already built the reading circuit, this is good news. The practice deficit that heavy AI-summary use creates is not necessarily a structural neural loss. It is a practice loss. And practice losses are recoverable.
For children whose circuits are not yet built, the situation is different — that concern belongs to a different argument, about reading instruction and formation development in childhood. For adult readers, the message is not doom but diagnosis.
No prescription follows from this. There is no validated protocol, no evidence-based dose, no peer-reviewed regimen for maintaining argument-evaluation capacity through deep reading. What exists is Wolf's first-person account of deterioration and recovery, the neuroscientific substrate Wolf and Dehaene established, and the Sweller framework that names what the productive cognitive work is. From these, a practice implication is visible without being fully specified: the capacity is maintained by use, the use is the doing of the cognitive work, and the cognitive work is not done when a summary is navigated instead of an argument constructed.
Not the moralizing about screens. Not the nostalgia for the physical book. Not the prescription dressed up as evidence. What remains, stripped of those, is a question worth holding.
AI is excellent at information delivery. That concession is genuine and has been repeated because it needs to be. The question is not whether AI should be used. The question is which cognitive work the reader does themselves — which load the reader chooses to carry — and what the answer builds or fails to build over time.
The load that AI removes is, in this framing, not a nuisance to be eliminated. It is a practice — done or not done, every time the reader engages a complex argument with a tool that offers to do the work instead.
What practice, then, is worth doing?
References#
Sweller, J.(2010)."Element interactivity and intrinsic, extraneous, and germane cognitive load." Educational Psychology Review 22(2): 123–138.https://link.springer.com/article/10.1007/s10648-010-9128-5; Sweller, J., van Merriënboer, J. J. G.,&Paas, F.(2019)."Cognitive Architecture and Instructional Design: 20 Years Later." Educational Psychology Review 31: 261–292.https://link.springer.com/article/10.1007/s10648-019-09465-5↩↩↩↩↩↩
Wolf, M.(2018).Reader, Come Home: The Reading Brain in a Digital World. Harper Collins. — verified via Slate review and WBUR interview.↩↩↩↩↩
Kestin, G.et al.(2025)."AI tutoring outperforms in-class active learning: an RCT introducing a novel research-based design in an authentic educational setting." Scientific Reports. PMC12179260.https://pmc.ncbi.nlm.nih.gov/articles/PMC12179260/↩↩↩↩
Zhang, H., Copara, M.,&Ekstrom, A. D.(2012)."Differential Recruitment of Brain Networks following Route and Cartographic Map Learning of Spatial Environments." PLOS ONE. PMC3445610.https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0044886↩↩↩↩
Dahmani, L.&Bohbot, V. D.(2020)."Habitual use of GPS negatively impacts spatial memory during self-guided navigation." Scientific Reports. PMC7156656.https://pmc.ncbi.nlm.nih.gov/articles/PMC7156656/↩↩↩
Dehaene, S.(2009).Reading in the Brain. Penguin Viking. — primary claims verified via PLOS Biology 2018 longitudinal fMRI study (https://journals.plos.org/plosbiology/article?id=10.1371/journal.pbio.2004103) and MIT News coverage of related research.↩↩↩
Risko, E. F.&Gilbert, S. J.(2016)."Cognitive Offloading." Trends in Cognitive Sciences 20(9): 676–688.https://doi.org/10.1016/j.tics.2016.07.002↩↩
Kao, C., Grant, L.,&Woltering, S.(2025)."Socratic AI tutors enhance adolescents'reasoning and engagement in real classrooms." Research Square preprint. DOI: 10.21203/rs.3.rs-8118546/v1.https://www.researchsquare.com/article/rs-8118546/v1↩↩↩
National Center for Education Statistics.(2023).NAEP Long-Term Trend Assessment Results: Reading and Mathematics (Age 13).https://www.nationsreportcard.gov/highlights/ltt/2023/↩
Bone, J. K.et al.(2025)."The decline in reading for pleasure over 20 years of the American Time Use Survey." iScience(Cell Press). PMC12496190.https://pmc.ncbi.nlm.nih.gov/articles/PMC12496190/↩
Twenge, J. M.(2023)."Are books dead? Why Gen Z doesn't read." Generation Tech(Substack), citing Monitoring the Future Survey data (University of Michigan/NIDA).https://www.generationtechblog.com/p/are-books-dead-why-gen-z-doesnt-read — As Twenge's secondary analysis of MtF data; MtF primary data not directly fetched.↩
Bratsberg, B.&Rogeberg, O.(2018)."Flynn effect and its reversal are both environmentally caused." PNAS 115(26): 6674–6678. PMC6042097.https://pmc.ncbi.nlm.nih.gov/articles/PMC6042097/↩
Further Reading#
Paas, F.&van Merriënboer, J. J. G.(2020)."Cognitive-Load Theory: Methods to Manage Working Memory Load in the Learning of Complex Tasks." Current Directions in Psychological Science 29(4): 394–398.https://journals.sagepub.com/doi/full/10.1177/0963721420922183 — The methods arm of cognitive load theory; grounds the productive-vs-unproductive load distinction and clarifies why the article defends germane intrinsic load rather than difficulty in general. Informs the E14 deployment throughout.
Wolf, M.(2007).Proust and the Squid: The Story and Science of the Reading Brain. Harper. — Wolf's earlier, more technical account of how the reading brain is built through acquisition; provides the developmental substrate for the circuit-plasticity claims made via E02. Background reading that shaped the article's framing of formation as an acquired, not innate, capacity.
Carr, N.(2010).The Shallows: What the Internet Is Doing to Our Brains. W. W. Norton. — Named the fragmentation-of-attention concern in relation to the web; provides pre-AI context for the trajectory argument in§9. The article's three-condition unprecedented framing depends on distinguishing AI's functional substitution from the web's attentional fragmentation that Carr documented.
Shumailov, I.et al.(2024)."AI models collapse when trained on recursively generated data." Nature 631(8022): 755–759.https://pubmed.ncbi.nlm.nih.gov/39048682/ — A second-order consequence of population-scale AI-summary use substituting for human-generated reasoning-dense text: the recursive training collapse it produces is the supply-side mirror of the demand-side argument this piece makes.
Kosmyna, N.et al.(2025)."Your Brain on ChatGPT:Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task." arXiv preprint 2506.08872.https://arxiv.org/abs/2506.08872 — EEG evidence of reduced neural engagement under LLM assistance for writing tasks. Preprint, small N, and a published methodological critique (Stankovic et al., arXiv: 2601.00856) — directionally consistent with this piece's mechanism claim, but not yet load-bearing on its own.
In brief#
Cognitive load is not one thing. Sweller's framework distinguishes extraneous load — the noise of bad prose, ambiguous referents, distracting layout — from germane processing of intrinsic load, the working-memory investment that constructs durable schema in long-term memory. Instructional design has spent decades reducing the first. The second is a different animal: it is not the inefficiency of reading a difficult argument, it is the mechanism by which a reader's capacity to evaluate arguments gets built.
The consensus position on AI and reading rests on a real finding. Kestin's 2025 Harvard RCT showed AI tutoring more than doubling learning gains over in-class active learning in less time, with p below 10⁻⁸. The result is large, rigorous, and impossible to wave past. What it measured, however, was content delivery at Bloom levels 2 through 4 in a structured domain — the transfer function of instruction. Kao's three-arm RCT pushes harder, showing that Socratic AI scaffolded by learning scientists can raise critical thinking in tenth-grade classrooms. Both findings are real. Neither tested what a reader builds alone with a long, unstructured argument when no one has designed the route.
The spatial-cognition literature has already studied the parallel. Zhang and Ekstrom's fMRI work documents two dissociable networks: route-learning encodes sequential cues, cartographic learning builds an allocentric survey of the space. Dahmani and Bohbot's longitudinal data — modest in sample, directional rather than settled — links heavier GPS use to steeper decline in cognitive mapping. Wolf calls deep reading an elaborately choreographed circus performance; Dehaene shows the reading circuit is recycled neural real estate, built by use, subject to atrophy from disuse. Risko and Gilbert add the offloading frame: what is delegated is not internally encoded.
No single study establishes the central claim. The argument is a convergent inference: Sweller names the mechanism, Wolf and Dehaene name the substrate, the spatial-cognition case names the documented parallel. AI summarization does not reduce extraneous load — that would be welcome. It eliminates the germane processing entirely, delivering the completed schema without the cartographic work. The reader receives the map without making it.
The trajectory matters because AI did not arrive on neutral ground. NAEP's age-13 reading scores have returned to the 1971 baseline. The American Time Use Survey shows adult pleasure-reading nearly halved between 2003 and 2023. Among twelfth graders, by Twenge's analysis of the same survey, the proportion reading zero books for pleasure rose from roughly seven percent to thirty-three. Norwegian within-family conscript data finds IQ peaking with the 1975 cohort and declining since. None of this was caused by AI. AI is the accelerant arriving on a pre-weakened substrate.
Wolf herself lost the cognitive patience to read Hesse; two weeks of deliberate practice restored it. The capacity is built by use, maintained by use, recoverable by use — for adults whose circuits are already in place. No protocol follows. What remains is a question: which load is worth carrying, every time a tool offers to carry it instead?