§1The question before the search box在搜索框之前,机会已经被筛选过一次
“MIT has free courses. If you wanted to learn, why didn't you look?” The accusation feels reasonable only after one important fact has been silently supplied: the listener knows MIT publishes courses, knows where they are, and has enough context to decide that one of them might help. The difference between a resource existing and a learner being able to use it begins before anyone types into a search box.
“MIT 都把课程免费公开了,你想学为什么不自己去看?”这句话听起来有道理,却悄悄加入了一个前提:被质问的人早就知道 MIT 公开了课程,知道去哪里找,也有能力判断其中哪门课可能适合自己。资源客观存在,与一个人实际能够使用它,并不是同一件事。差别甚至在打开搜索框之前就出现了。
Consider two students who both want to learn calculus. One has a teacher who recommends an introductory OpenCourseWare class, explains what a prerequisite is, and points to problem sets. The other hears only that they should study harder, searches for isolated explanations, and cannot distinguish a complete course from a few attractive videos. These are illustrative students, not measured populations. Their contrast shows why an instruction to “take initiative” cannot be a complete account of opportunity.
设想两个都想学微积分的学生。甲有老师介绍 MIT OpenCourseWare,告诉他从哪门课开始、需要补哪些先修知识、如何配合习题。乙从未接触过这些信息,只被告知“你应该努力学习”,于是反复搜索零碎的讲解视频,却无法判断哪个系列完整、哪个难度合适。这是用于阐明机制的假设情境,不是关于两类学生群体的统计结论。它说明:一句“你要主动”,不足以解释学习机会为什么不同。
This does not eliminate personal responsibility. Someone who knows of an appropriate, accessible resource still has to choose whether and how to study it. A learner might even discover the same material through a broad search without knowing MIT beforehand. But responsibility for choosing among known options should not be confused with responsibility for having been exposed to the options in the first place.
这并不意味着个人毫无责任。一个人已经了解某项适合自己的资源、具有实际使用条件之后,仍然需要自己选择、投入时间并完成练习。他也可能通过更一般的关键词搜索,偶然发现从未听过的公开课程。但“在已知选项中作出选择的责任”,不能被偷换成“理应事先知道所有选项的责任”。
§2A free course can still be hard to reach免费开放,不等于真正抵达
MIT OpenCourseWare makes materials from more than 2,500 MIT courses available free of charge. Its own guidance says no account or registration is required; it also says OCW does not award credit, provide instructors, or offer certificates1. This is a genuine public good, not an example of a hidden paywall. Yet the catalog can still leave a self-directed learner with questions: Which course is introductory? Which lectures are accompanied by assignments? What mathematics should come first? Is a different teacher's course actually a better starting point?
MIT OpenCourseWare 免费提供超过 2,500 门 MIT 课程的教学材料。MIT 官方说明无需注册或付费,同时明确它不提供学分、教师互动和结业证书1。这是一项真正有价值的开放教育实践,而不是另一道收费墙。但对一个独立学习者来说,目录本身仍然留下不少问题:应该先学哪门?有没有完整作业?函数和代数需要掌握到什么程度?更适合我的课程,会不会其实来自另一所大学?
It helps to separate three conditions. A resource is available when someone has published it under usable terms. It is discoverable when an appropriate learner can encounter it without already knowing the institution, title, or acronym. It is usable when its language, sequence, prerequisites, format, and exercises allow that learner to make progress. A fourth condition—whether it improves independent learning—requires evidence beyond a visit or a download.
我们至少应区分三个条件:开放意味着资源已经以明确方式提供;可发现意味着不知道提供者名称的人,也有合理机会找到它;可利用意味着语言、难度、先修知识、结构与练习能够支持具体的人真正开始学习。至于有没有提高独立学习能力,那又是需要另外验证的第四个问题。一次访问、一次下载,均不能直接等同于学习成效。
- 01Available开放Published under usable terms资源已经以明确方式提供
- 02Discoverable可发现Found without already knowing the institution, title or acronym不知道提供者名称的人,也有合理机会找到它
- 03Usable可利用Its language, sequence, prerequisites, format and exercises let this learner make progress语言、难度、先修知识、结构与练习,能支持具体的人开始学习
- 04Improves independent learning提高独立学习能力Needs evidence beyond a visit or a download需要另外验证:一次访问、一次下载不等于学习成效
Public promotion, then, need not mean advertising an elite brand. A good library orientation, school resource guide, or search system might start with “I need to understand functions before calculus,” not “I want MIT.” It should also be free to say that a locally produced text or a less prestigious institution offers the better lesson for this person. Making a resource easy to find is part of making an opportunity real.
因此,推广优秀教育资源不等于给名校打广告。图书馆的新生导览、学校的资源指南、公共学习检索系统,可以从“我还没理解函数,想开始学微积分”出发,而不是要求用户先知道“MIT”三个字母。它还应该允许自己得出一个不讨喜的结论:对眼前这个学生来说,一套普通高校教师制作的讲义也许比名校公开课更好。让人找得到合适资源,本身就是把教育机会落到实处。
§3Finding an item is not the same as obtaining it找到了一份资料,仍然可能拿不到
Now imagine that the learner discovers exactly what they need: a past mathematics examination, a journal article cited in a textbook, or a dataset required to reproduce a result. The link exists, but it may lead to an unofficial scan, a purchase screen, or an institutional login. These obstacles often get grouped together as “closed knowledge.” Their causes differ, however, and they should not be assigned a single easy remedy.
接下来设想:学生已经准确找到了自己需要的东西——一份历年数学试卷、一篇教材引用的外文论文,或者复现某项研究所需的数据。链接确实存在,点击后却可能出现非官方扫描件、单篇收费页面或学校身份认证。它们常被笼统地描述为“知识被封锁”,但背后的原因并不一样,解决办法也不应只有一句“全部免费”。
Consider examination papers. A trustworthy archive would identify the organizing body, exam year, original questions, corrections, and the status of any answer key. Reconstructed papers can be useful, but a polished PDF circulating on exam night is not proof that the organizer leaked its source file. At the same time, the availability of commercial reconstructions does not answer whether an authoritative version can be obtained at reasonable cost.
先看考试真题。可靠的公共档案应当标明考试组织方、年份、原始试题、勘误记录,以及所谓“标准答案”究竟是否由官方发布。民间还原卷有使用价值,但考试当晚流通的精美 PDF 并不自动证明存在内部泄题:考场记忆重建、照片转录与命题机构原始文件是不同的证据。与此同时,市场上有收费整理版,也无法替代公众对权威试题来源的合理需求。
After an exam, what may be published depends on its rules, rights, security period, and whether questions are reused. Some organizers already publish materials; others may have reasons to delay or limit release. The policy proposal is therefore precise rather than absolute: establish a public release schedule for eligible papers, explain exceptions, publish corrections, preserve the original files in searchable accessible formats, and distinguish official answers from third-party solutions. Payment can reward teachers for original analysis without requiring a monopoly on access to the underlying eligible paper.
考试结束后能否公开,还取决于具体规则、版权、保密期限以及试题是否计划重复使用。有的考试已经公布部分材料,有的可能需要延迟。因此更准确的公共政策目标是:为符合条件的试卷制定明确的发布时点,说明例外,公开原件与勘误,提供可检索的长期档案,并区分官方答案和第三方解析。教师可以因原创讲解获得报酬,但这不意味着符合公开条件的原题只能经由商业资料获得。
§4Why the university card opens a different door为什么一张校园卡能打开另一道门
A research paper is another case. A university library often buys a license allowing a defined population to use a publisher's database, rather than ownership of its contents. Peking University Library's remote-access notice identifies affiliated cardholders who may use its licensed electronic resources from off campus, and excludes users without the required status2. Tsinghua University Library's policy prohibits passing acquired documents or access credentials to people outside the university3. Such restrictions are real; suggesting credential sharing would not fix the licensing problem.
学术论文面对的是另一套机制。高校图书馆通常购买的是限定合法用户范围的数据库使用许可,而不是买下全部论文版权。北京大学图书馆的校外访问须知列出了可以远程访问电子资源的校内身份,并明确其他用户不具有同等远程权限2;清华大学图书馆则规定,不得将取得的文献或个人账号提供给校外人员使用3。这是真实存在的授权边界,也说明借用名校账号不是合规解决方案。
The moral question is not whether these librarians are keeping knowledge from outsiders for pleasure. They are generally obligated to follow contracts that fund campus access. The structural question is why two people pursuing the same legitimate scholarly inquiry can have such different routes to full text simply because one holds a particular institutional identity. Admission and employment are imperfect proxies for intellectual need.
真正需要讨论的,并不是指责图书馆员故意排斥校外求知者。他们往往必须遵守学校与出版商之间的合同。结构性问题在于:两个对同一篇论文有正当学术需求的人,可能仅仅因为是否属于某所高校,就要经过完全不同的获取路径。录取身份和任职单位,并不能完整衡量一个人的研究需求与理解能力。
Research data require even more care. A published article, a de-identified reusable dataset, patient-level records, restricted interviews, and a confidential industry dataset are not interchangeable. Privacy, consent, research ethics, security and intellectual-property duties can make some restrictions justified. More equitable access means predictable, lawful processes matched to the material—not indiscriminate release.
研究数据还需要更细致地对待。已经发表的论文、去标识化且允许复用的数据集、患者原始记录、受伦理审批约束的访谈资料以及企业保密数据,不能混为一谈。隐私、同意范围、科研伦理、安全与知识产权,都可能构成合理的访问限制。更公平的目标是为不同材料建立透明、合法、可申请的获取机制,而不是不加区分地要求全部公开。
§5The missing measure in academic procurement采购的成果,应该是获得了多少真正需要的知识
Collective purchasing and document delivery are not imaginary solutions still waiting to be invented. CALIS's e得 service already integrates discovery, document delivery, interlibrary borrowing and related fulfillment services4. China's National Science and Technology Library reports some foreign journal resources negotiated under a national-license model for eligible non-profit academic organizations, subject to conditions5. These arrangements matter because a serious reform must begin with what is already working.
联合采购和文献传递并不是等待发明的新概念。CALIS 的 e得已经整合检索、馆际互借、文献传递等获取方式4;国家科技图书文献中心 NSTL 也说明,其部分外文期刊资源采用“国家许可、全国开通”的模式,面向符合条件的非营利学术机构提供服务5。这些机制值得认真研究,因为改革应当从已有成果出发,而不是假装此前什么都没有做过。
Nor does national purchasing automatically make every subscription cheaper. Publishers may reprice according to the newly covered audience; usage varies enormously between disciplines; a contract covering every institution may still omit the article a researcher needs. Buying more database packages is an input metric, not proof of better scholarly access. The question is what proportion of legitimate requests ultimately end in access to the requested full text.
同样不能预先断言“国家统一买一定便宜”。出版商可能按照更大的用户覆盖面重新定价;不同学科使用需求差异很大;即使覆盖所有高校,一个新购买的资源包也可能仍然没有研究者想要的那篇文章。采购了多少数据库,是投入指标,不等于文献需求得到了满足。我们真正应该关心的,是提出了正当全文需求的人,最后有多少成功获得了全文。
Imagine a privacy-preserving log of requests for particular DOIs or chapters. For each request, record only what is needed: whether an authorized open copy exists; whether an institution has the relevant license; whether lawful document delivery was tried; the time, user-facing cost, and outcome. Define a fulfillment rate as valid requests legally fulfilled within a specified period divided by all valid requests. Compare it across types of institutions, fields, and unaffiliated users without publishing personal reading histories. Such evidence could help test a layered arrangement: broadly needed national licenses, subject consortia, local supplements, open repositories, and on-demand delivery. A claim that this saves money should be made only after contracts and actual costs have been compared.
可以设想一套保护个人阅读隐私的需求记录:某篇 DOI 论文或章节是否有合法开放版本?机构是否已有授权?是否尝试合规文献传递?等待了多久、支付多少、最终是否取得?把“在约定时限内合法满足的有效请求数/全部有效请求数”定义为文献需求满足率,再比较不同学校、学科和独立研究者的差异。这样才有可能检验分层方案:普遍需求考虑国家许可,专门领域通过学科联盟采购,学校补充自身特色资源,其他缺口借助开放仓储和按需传递。至于能节省多少钱,必须先拿到合同、价格和真实需求数据,不能拍脑袋承诺。
Fulfilment rate
文献需求满足率
valid requests legally fulfilled within a specified period在约定时限内合法满足的有效请求数all valid requests全部有效请求数
Recorded for each request, and nothing more
每条需求只记录
- Is there an authorised open copy?是否有合法开放版本?
- Does the institution hold the licence?机构是否已有授权?
- Was lawful document delivery tried?是否尝试合规文献传递?
- Time, cost to the reader, outcome等待多久、支付多少、最终是否取得
§6A path through knowledge, not a ranking of universities最好的课程,不应该是名校排行榜
Suppose discovery and licensed access work perfectly. We would still have to answer a pedagogical question: can the learner use what has been found? Course prestige does not settle this. A proof-oriented calculus class might be excellent for one student and counterproductive for another who is missing basic algebra. A course designed for exams might serve a time-constrained candidate better than one designed for mathematical exploration; the reverse may also be true.
再向前想一步:即使找得到资料,也能免费合法取得,就必然学得会吗?并不。对已经掌握基础的人,一门强调严格证明的微积分课程可能极其优秀;对连函数图像都不熟悉的学生,它却可能成为新的挫败来源。同样,备考者可能需要概念辨析、题型变化和限时练习,探索数学理论的人则可能更需要证明、反例与开放性问题。没有脱离具体学习者的“绝对最优课程”。
A useful course catalog would describe prerequisites, instructional goals, evidence of quality, language, accessibility, exercise coverage, solutions, and opportunities for feedback. It would then map those features to an individual learner's current understanding rather than simply listing universities by reputation. The result might be one complete course, or a coherent path through several. Either choice requires guarding against fragmented recommendations that never build a cumulative curriculum.
真正有用的课程目录,应当描述先修要求、教学目标、质量证据、语言和无障碍支持、练习覆盖、答案与反馈机制,再把这些特征与个人当前的理解程度联系起来,而不是简单按大学声望排序。匹配结果可能是一门完整课程,也可能是几位教师的课程共同组成的连续路径。但把名校视频碎片随意拼接,同样可能破坏知识的内在顺序。
This is the point where a public knowledge infrastructure becomes a learning infrastructure. A library search answers “Where is the document?” A pedagogical system must also answer “Should I study it now, what do I need first, and what will show that I understood?” Those are not the same design problem, but they meet in the experience of one person trying to learn.
这时,公共知识基础设施才开始延伸为学习基础设施。图书馆检索首先回答“文献在哪里”;教学系统还需要回答“我现在适不适合学它、应该先补什么、怎样才能证明自己学会了”。它们是两个不同的技术与制度问题,却汇集到同一个学习者身上。
§7Where AI helps—and where it can make things worseAI 应该降低学习阻力,而不是替学生完成学习
An AI system could help a learner describe a goal in ordinary language, discover lawful high-quality material, explain why it was selected, and propose a prerequisite-aware sequence. During a lesson it could answer questions about a specific paragraph or diagram, distinguish a computational slip from a conceptual misunderstanding, suggest a small next step, and return the learner to the primary source. The course remains a coherent source of teaching; the AI supports the learner's encounter with it.
AI 可以使前面的环节连起来。学习者只需用自然语言描述目标,系统就帮助发现合法可用的高质量课程,说明推荐理由和先修知识,并形成可调整的学习路径。学习过程中,AI 能针对当前章节或图示回答问题,区分计算失误和概念误解,给出足以推动思考的下一步提示,再把学生带回原始教材与练习。课程维持严谨、连贯的知识主线,AI 处理个人遇到的障碍。
But immediate assistance can create an illusion of mastery. In a published 2025 PNAS field experiment involving nearly 1,000 high-school mathematics students in Turkey, an unrestricted GPT-based assistant improved assisted-practice performance yet was associated with worse unassisted test performance than a no-AI condition. A more deliberately designed tutoring interface with teacher-authored safeguards largely mitigated that negative effect in the study6. This does not establish that all AI tutoring harms or benefits learning. It demonstrates why getting today's answer correct is an inadequate success measure.
但“随时得到答案”有可能制造掌握知识的假象。2025 年发表于《美国国家科学院院刊》(PNAS)的一项现场实验,涉及土耳其一所高中近千名数学学习者:无充分教学保护的 GPT 辅助提高了有 AI 时的练习表现,却使移除 AI 后的独立测验表现低于无 AI 对照组;采用教师设计提示等约束的辅导版本,在该研究中很大程度上缓解了负面效应6。这不是对所有 AI 教育产品的普遍判决,却提醒我们:今天答对,不等于明天能够独立完成。
A better loop would diagnose a starting point, choose material, let the learner attempt it, provide the least help that moves understanding forward, gradually fade support, and test a fresh problem without AI. A delayed check could reveal whether the idea lasts. The AI also needs a reliable way to say it does not know, to cite its source, and to avoid inventing a confident explanation. Personalization is meaningful only when its effect on unaided learning can be tested.
因此,学习过程更应该是:诊断起点,匹配课程,让学生尝试,提供必要而不过度的帮助,逐步撤去提示,再用新题独立验证,并在更久之后检验是否遗忘。AI 还应能够承认不确定、引用资料来源,避免编造看似专业的解释。个性化不是把题目换成学生喜欢的背景,而是用证据判断何时需要帮助、什么帮助有效,以及帮助能否最终退出。
§8Turn the argument into tests that can fail让这些主张接受可能失败的检验
The preceding claims contain different kinds of uncertainty. We have documentary evidence that OCW exists and that campus licenses restrict access. We do not have evidence here that every disadvantaged learner is unaware of OCW, that national procurement will cut costs, or that one AI learning design will work for every subject. Treating those propositions as established facts would repeat the very knowledge-quality problem this essay criticizes.
前面的观点包含不同层次的确定性。我们有官方资料证明 OCW 公开存在,也有高校规则证明授权范围确实受限;但本文没有证据证明所有弱势学生都不知道 MIT 课程,没有证据证明全国采购一定省钱,也没有证据证明某一种 AI 学习系统适合所有学科。若把后面这些设想直接写成既成事实,就会重复我们批评的知识质量问题。
Three complementary studies could move the discussion forward. First, a discoverability task: recruit learners with varied institutional backgrounds, give them realistic goals without naming any provider, and record whether they identify a suitable free course, how long it takes, and what prerequisites they miss. Second, a document-request audit: sample valid scholarly requests, measure lawful fulfillment rate, delay, and cost, while separately identifying publication licensing and privacy restrictions. Third, an instructional comparison: randomize comparable learners between a conventional course directory and a recommendation-plus-tutoring design; assess not just engagement but later, unaided transfer on unfamiliar problems.
至少可以设计三个可检验的研究。第一,资源发现实验:招募具有不同教育背景的学习者,只给出真实学习目标而不提供机构名称,观察他们能否找到合适的免费课程、花费多少时间、是否遗漏重要先修要求。第二,学术文献需求审计:抽取有效文献请求,衡量合法全文获取率、等待时间和成本,并把版权限制与涉及隐私的数据访问要求分开。第三,教学效果比较:让基础相近的学习者随机使用传统课程检索方式,或使用课程匹配加 AI 辅导;评价不仅包括体验和投入,还必须包括延迟的、离开 AI 后的独立迁移表现。
These tests should report failures as carefully as successes. If users find courses but cannot complete them, discovery was not enough. If collective procurement raises total costs without improving fulfillment, that policy version should be changed. If students produce better AI-assisted answers but weaker independent solutions, a tutoring design has not achieved the intended goal. Evidence is valuable precisely because it can overturn our attractive intuitions.
这些研究不仅应该报告成功,也必须报告失败。如果人们找到了课程,却无法完成学习,说明可发现性还不够;如果联合采购增加成本却没有提升全文获取率,方案就应调整;如果 AI 让练习答案更好看,却降低独立解题能力,那么这一版辅导设计没有实现目标。证据的价值,恰恰在于它能推翻一个原本令人心动的设想。
- 01Discoverability task资源发现实验Records记录Whether learners with varied backgrounds find a suitable free course, how long it takes, which prerequisites they miss不同背景的学习者能否找到合适的免费课程、花多少时间、遗漏了哪些先修要求It has failed if失败信号They find courses but cannot complete them找到了课程,却无法完成学习
- 02Document-request audit学术文献需求审计Records记录Lawful fulfilment rate, delay and cost for a sample of valid requests抽样有效请求的合法全文获取率、等待时间和成本It has failed if失败信号Collective purchasing raises costs without improving fulfilment联合采购增加了成本,却没有提升全文获取率
- 03Instructional comparison教学效果比较Records记录Later, unaided transfer on unfamiliar problems, not just engagement不只看体验和投入,还看延迟的、离开 AI 后的独立迁移表现It has failed if失败信号AI-assisted answers improve while independent solutions weakenAI 让练习答案更好看,却降低了独立解题能力
§9Who is responsible for the gap?究竟该由谁承担责任?
Individual initiative remains important: people should learn how to search, compare explanations, and sustain difficult practice. Institutions also face real limits: publishers have rights, examinations have security requirements, teachers have finite time, and public budgets have competing demands. None of those facts justifies treating every missed opportunity as an individual's personal defect.
个人的主动性当然重要:每个人仍需要学习检索、比较解释、坚持练习。制度也确实有边界:出版商有权利,考试涉及保密,教师的时间有限,公共经费还有其他用途。但这些约束都不能推出“没有接触到优质资源,就一定是个人不努力”。
Responsibilities should follow the points where intervention is possible. Resource creators can publish transparent metadata, prerequisites and accessible formats. Schools and libraries can actively introduce legitimate sources instead of waiting for students to know their names. Funders and consortia can investigate wider licensing and measure requests that went unfulfilled. Researchers and education-technology builders can test whether discovery and personalized support produce independent understanding. Learners still choose what to study; the surrounding system should give that choice real content.
责任应当与行动能力相匹配。内容生产者可以给出清楚的标题、元数据、先修要求与可访问格式;学校和图书馆可以主动介绍可靠来源,而不是等学生自己说出资源名称;资助机构与采购联盟可以尝试扩大授权,并公开尚未满足的有效需求;教育技术研究者可以检验资源发现和个性化辅导是否真正促进独立理解。学习者仍然拥有决定学什么的自主权,系统则应尽可能让这种选择有实质内容。
The underlying ambition is modest in wording and demanding in practice: a person's chance to encounter, obtain, understand, and independently use good knowledge should depend less on the school they happened to enter and more on their genuine learning needs. Making a course free is progress. Making it discoverable, lawfully obtainable, appropriately taught, and eventually unnecessary to lean on is a larger, measurable educational goal.
我们期待的目标,表述起来不复杂,实现起来却很艰难:一个人接触、获取、理解并独立使用优秀知识的机会,不应过度取决于他恰好进入哪所学校。让课程免费,是进步;让课程被需要的人发现、以合法方式取得、以合适的路径学懂,最终能够不依赖帮助地运用,才是更完整、也更值得测量的教育目标。
Sources and further reading资料来源与延伸阅读
- MIT Open Learning: MIT OpenCourseWare (free courses and terms of use)MIT Open Learning:MIT OpenCourseWare(免费课程与授权信息)
- Peking University Library: off-campus access notice (who may use licensed resources)北京大学图书馆:校外访问须知(合法用户范围)
- Tsinghua University Library: notice on protecting the intellectual property of electronic resources清华大学图书馆:关于保护电子资源知识产权的公告
- CALIS Smart Services: e得 document delivery and interlibrary sharingCALIS 智慧服务:e得文献与馆际共享
- National Science and Technology Library: foreign journals opened nationwide国家科技图书文献中心:全国开通的外文期刊资源
- Bastani et al. (2025), PNAS: Generative AI without guardrails can harm learningBastani et al. (2025), PNAS:Generative AI without guardrails can harm learning





