🤖 AI 教育市值 200 亿美元，但没人知道它到底有没有用

The $20 Billion AI Education Industry Has an Evidence Problem

"Strictly speaking, we do not know for sure if AIED 'works' or not."

「严格来说，我们并不确定 AI 教育到底有没有用。」

📐 The Elephant in the Classroom

Right now, there are more than 30 multi-million-dollar AI education companies operating worldwide. The market is projected to exceed US$ 20 billion within five years. Schools are buying AI tutoring systems, adaptive learning platforms, and automated essay graders. Governments are writing AI-in-education policies.

Here's what almost nobody mentions: the evidence that any of this works is, at best, mixed — and at worst, points in the wrong direction.

In 2022, Wayne Holmes (UCL) and Ilkka Tuomi published a comprehensive review of the entire AI-in-education field. They didn't just list the technologies. They asked the question the industry doesn't want asked:

If we've been building AI tutors since 1958 — why, sixty years later, can't we prove they work?

📏 教室里的那头大象

全球目前有超过30 家融资过千万美元的 AI 教育公司。市场预计五年内突破200 亿美元。学校在买 AI 辅导系统、自适应学习平台、自动作文评分器。各国政府在写 AI 教育政策。

几乎没人提的一件事：这一切有效的证据，充其量只能说「好坏参半」——说难听点，有些证据指向反方向。

2022 年，伦敦大学学院的 Wayne Holmes 和芬兰学者 Ilkka Tuomi 发表了一篇覆盖整个 AI 教育领域的综述。他们没有只罗列技术。他们问了一个行业不想被问到的问题：

如果从 1958 年起我们就在造 AI 老师——为什么六十年后，还是证明不了它们有效？

⏳ AI 教育不是新东西 / AI in Education Is Not New

Here's a timeline that might surprise you:

这是你可能不知道的时间线：

重点不是技术不新。重点是营销话术也不新。Skinner 1958 年说的「效果像私人家教」，跟今天每一家 AI 教育公司的官网首页，用的是同一套叙事。

📖 The Sacred Text: Bloom's 2-Sigma / 产业圣经：Bloom 的 2-sigma

Almost the entire AI-in-education industry is built on one finding from 1984.

Benjamin Bloom — the same Bloom who gave us Bloom's Taxonomy — compared three teaching formats:

This finding became the North Star of AIED: if one-on-one human tutoring is 2 sigma better, and AI can simulate one-on-one tutoring, then AI should be able to deliver those same gains — at scale. This is the logic behind virtually every Intelligent Tutoring System ever built.

The problem? As Holmes and Tuomi point out, the 2-sigma effect has become a "rarely contested implicit model" — an assumption so deeply embedded that few researchers question it. And when independent evaluators did question it, the results were not pretty.

几乎整个 AI 教育产业都建立在 1984 年的一个发现上。

Benjamin Bloom——对，就是提出 Bloom 分类学的那位——比较了三种教学形式：

这个发现成了整个 AI 教育领域的北极星：一对一真人辅导能提升 2 个标准差 → AI 能模拟一对一辅导 → AI 应该能达到同样效果，而且能规模化。这就是几乎每一套智能辅导系统背后的核心逻辑。

问题在哪？Holmes 和 Tuomi 指出，2-sigma 效应已经成了一个「几乎不受质疑的隐性假设」——深植于领域内部，几乎没人挑战。而当独立第三方真去挑战的时候，结果不太好看。

💥 The Evidence, Under the Microscope / 证据放大镜

Holmes and Tuomi catalogued the evidence landscape, and what they found is a pattern you'd recognize from supplement marketing: lots of small, positive studies funded by the people who built the product. Almost no large, independent ones.

Then they pointed to the gold standard: the What Works Clearinghouse, run by the U.S. Department of Education. The Clearinghouse only counts randomized controlled trials and high-quality quasi-experiments — the same standard you'd demand for a medical treatment.

Here's what it found for Cognitive Tutor, one of the most famous and widely deployed AI tutoring systems ever built:

Let that sink in. The poster child of AI tutoring — the system that launched a thousand startups — might actually make students worse at geometry.

Holmes and Tuomi's conclusion is devastatingly understated: meta-analyses show "modest learning gain impacts across many relatively small settings." And then the punchline: "Strictly speaking, we do not know for sure if AIED 'works' or not. More importantly, it is not entirely clear what such 'working' would comprise."

Holmes 和 Tuomi 梳理了整个证据版图，发现的模式你可能在保健品广告里见过：大量由产品开发者资助的小型正面研究，几乎没有大规模的独立验证。

然后他们指向了金标准：美国教育部运营的What Works Clearinghouse（什么有效信息中心）。这个机构只认随机对照试验和高质量准实验——跟药品审批用的是同一套标准。

它怎么评价 Cognitive Tutor——有史以来最有名、部署最广的 AI 辅导系统之一？

想一下：AI 教育产业的招牌产品——催生了无数创业公司的原型——可能让学生几何学得更差。

Holmes 和 Tuomi 的结论克制得近乎残酷：元分析显示「在较多小规模场景中有适度的学习收益」。然后是致命补刀：「严格来说，我们并不确定 AI 教育到底有没有用。更重要的是，我们甚至不确定'有用'到底是什么意思。」

🗂️ The Taxonomy: What's Actually Out There / 全景图：到底有哪些 AI 教育产品

Despite their skepticism, Holmes and Tuomi provide a clear map. They classify AIED into three categories — student, teacher, and institution — with each category containing multiple product types. Here's the landscape:

尽管持怀疑态度，Holmes 和 Tuomi 还是给出了一张清晰的全景图。他们把 AI 教育分成三类——面向学生、面向教师、面向机构——每类下面又细分出多种产品：

The paper also notes that many AI tools used in education weren't even designed for it — Google Docs, WeChat, YouTube, TikTok have all been "repurposed" by students and teachers. This is the invisible AIED.

论文还指出，很多在教育场景中使用的 AI 工具根本不是为教育设计的——Google Docs、微信、YouTube、TikTok 全部被师生「挪用」了。这才是隐形的 AI 教育。

💡 So What? / 所以呢？

Holmes and Tuomi aren't saying AI in education is hopeless. They're saying the conversation has been captured by a very narrow idea of what education is — one where the only thing that matters is whether a student can answer test questions correctly. This is what they call the "implicit model" problem: if your definition of success is test scores, and the AI marginally improves test scores in small studies, then the AI "works."

But education is more than test scores. It's about socialization, individuation, critical thinking, and becoming a person who can navigate a complex world. No AI tutoring system has ever been designed to measure — let alone improve — any of that.

The real risk isn't that AI education doesn't work. It's that it works for a definition of education so narrow that "working" becomes meaningless.

Holmes 和 Tuomi 没有说 AI 教育没希望。他们说的是：关于 AI 教育的讨论，被一种极其狭窄的教育观绑架了——在这种观念里，唯一重要的是学生能不能答对考试题。

但教育不只是分数。它关乎社会化、个体成长、批判性思维、成为一个能在复杂世界中游刃有余的人。没有任何 AI 辅导系统曾经被设计去测量——更别说提升——这些东西。

真正的风险不是 AI 教育没用。而是它「有用」的标准太窄，以至于「有用」这个词本身失去了意义。

📚 References / 参考文献

Holmes, W., & Tuomi, I. (2022). State of the art and practice in AI in education. European Journal of Education, 57(4), 542–570. https://doi.org/10.1111/ejed.12533

📖 延伸阅读：文中关键概念解释

🔤 Intelligent Tutoring System（ITS，智能辅导系统）

你可以把它理解成一个 AI 家教 App。它给学生推送知识点 → 出题 → 根据答案判断学生掌握程度 → 推送下一个知识点。整个过程由算法驱动，不需要真人老师介入。Cognitive Tutor 和国内的「作业帮」类产品都属于这个范畴。是 AI 教育中最主流、融资最多的产品类型。

📊 Bloom 的 2-sigma 效应

Benjamin Bloom 在 1984 年发现：接受一对一真人辅导的学生，平均成绩比大班教学的学生高出两个标准差（2 sigma）。一个标准差大概意味着从第 50 名跳到第 84 名。两个标准差意味着从第 50 名跳到第 98 名。这个数字太大了，以至于成了整个行业追逐的圣杯。问题在于：这个发现来自 1984 年的小规模实验，后续大规模独立验证从未复现过这个效果。

🏛️ What Works Clearinghouse（什么有效信息中心）

美国教育部下属的独立评估机构。它的工作很简单：搜集某个教育产品的所有高质量研究（只认随机对照试验和严格准实验），然后给出一个结论——有效、无效、还是证据不足。医学领域有 FDA 审药，教育领域有它审教学法。因为标准极其严格，很多被广泛使用的教育产品在它这里拿不到「有效」评级。

🧠 Mastery Learning（掌握学习）

一种教学策略：不允许学生带着漏洞往前走。每个学生必须在一个知识点上达到预设的掌握水平（比如正确率 80%），才能进入下一个知识点。Bloom 在 1968 年提出的理论，至今仍是大多数 ITS 的底层教学设计逻辑。听起来很合理，但问题在于——它把教育简化成了「知识点 → 测试 → 通过 → 下一个」的流水线。

👁️ 教育的社会化功能 vs 资格化功能

荷兰教育哲学家 Gert Biesta 提出，教育有三个功能：(1) 资格化——教学生知识和技能，让他们能做事；(2) 社会化——把学生培养成能融入社会和文化的人；(3) 主体化——帮助学生成为独立、有批判意识的个体。Holmes 和 Tuomi 的核心批评就是：AI 教育产业只关心资格化，完全忽视了后两者。

Reviewed with healthy skepticism, published with ❤️.

以健康的怀疑精神审校，以爱发布。

#AI教育 #教育科技 #EdTech #人工智能
#教育研究 #智能辅导系统 #批判性思维 #Bloom分类学
#教育公平 #科技伦理 #今天你学到了吗
#不要问AI能不能教而是问教育到底要教什么