AI 学的是思维的产物,而非思维本身

由 Gaby Clark 编辑,Robert Egan 审校
本文已按照 Science X 的编辑流程与政策进行审阅。编辑在确保内容可信度的同时,突出了以下特点:
由研究人员撰写
添加到首选来源
过程,以及等价等级。图片来源:arXiv (2026)。DOI: 10.48550/arxiv.2608.16213
当你走过画廊,欣赏墙上的画作时,你所看到的是完成后的成品。它们是无数小时的学习、练习以及反复试错的结果,但这一切都不在展出之列。
我们所见到的多数专业工作都是如此。音乐录音不会收录失败的录音和即兴合奏,出版的书籍会略去凌乱的初稿,科学论文也不会提及那些年的错误开局与死胡同。
换句话说,我们实际看到的只是产生它的真实过程留下的痕迹。人工智能(AI)模型正是用这些痕迹来训练的——因此,它们自身也会产出痕迹,这些痕迹的形状看起来像思维的产物,但其生成方式未必相同。
在一篇新论文中,我们提出:只关注轨迹而忽略过程,会束缚当前的人工智能系统——而使用这些系统的人也可能遇到同样的问题。该论文发表在 arXiv 预印本服务器上。
较新的「推理」模型或许会打印出一份关于自身步骤的报告,但这份被报告出来的推理本身又是一条轨迹。它往往与真正驱动答案的实际计算对不上。
智能体式系统稍稍更接近真正的思考。在这里,围绕大语言模型的一层软件「外壳」运行一个循环,把结果反馈回去,并在过程中引入一些真正的试错式修正。但夹在中间的思维链——由大语言模型生成的那部分——仍然是一条轨迹。
这一差距在创意领域最为明显。画家和音乐家有时会谈到,自己的作品如何被创作过程中发生的意外与摩擦所改变,比如颜料意外流淌,或乐器听起来「不对」。AI 模型从不会遇到这类情况。
这一问题在人类身上的表现形式远比 AI 古老。1997 年,一位美国航空的机长警告说,飞行员正在变成「品红色航线的孩子」。
他们如此忠实地跟随飞机的飞行计算机,以至于人工飞行技能不断退化。自那以后,航空业一直在采取措施管理对自动化的依赖。
一些医疗诊所采用 AI 辅助结肠镜检查时也发生了同样的情况——他们用软件在实时摄像头画面上标记可能的息肉。一段时间后,人类临床医生的技能开始退化:在不使用 AI 系统时,他们发现的息肉数量比该系统引入之前更少。
这种效应在课堂上也已出现。获得不受限制的生成式 AI 支持的高中数学学生,在工具可用时表现优于同龄人,但一旦工具被拿走,得分就下降了。
这种退化可能迅速发生。一项研究发现,大约 10 分钟的 AI 辅助,就足以让人们在失去 AI 时表现更差,并且更快放弃问题。
失败与错误是老师
这个问题该如何解决?答案或许是把更多“过程”纳入其中。
对 AI 而言,这意味着构建自身采用更像人类的过程的系统。当有人质疑它界定问题的方式时,助手可以修正自己的答案,而不是像当前系统常做的那样一味附和。或者,智能体可以从失败与错误中学习,让来自外界的反驳实质性地改变它处理任务的方式。
对人类而言,这意味着设计能够保留人自身智能过程的 AI 工具。在哈佛的一项试验中,由构建良好的 AI 物理辅导工具教授的学生,比主动学习课堂上的同龄人学得更多、更快,因为该辅导工具让他们始终自己完成生成性工作。
一个好的 AI 教学助手,应该更常给出通往答案路径上的下一个问题,而不是简单地直接给出答案。有时它应当提出一个模糊之处,而不是去解决它,因为良好的判断往往来自对自身未知之处的缓慢思索。
过程,过程,过程
正是过程让真正的智能与创造力得以涌现。一个融入了更多这种过程的人工系统,会更接近真实之物。
人类学习同样需要过程。在教育以及任何技术性事业中,我们都需要考虑如何设计和运用 AI 系统,以保留更多的思考与学习过程。
目前,机器很容易生成更多思维痕迹,但过程要难构建得多。如果我们做对了,我们的机器可以变得更智能,而不会让我们变得更愚钝。
Michael J. Richardson 等,《过程构成的智能:人类与机器的共同标准》,arXiv(2026)。DOI: 10.48550/arxiv.2608.16213
期刊信息:arXiv
由 The Conversation 提供
这篇报道背后是谁?
英语硕士,自2021年起担任文字编辑,拥有高等教育与健康领域内容经验。致力于可信赖的科学新闻。完整简介 →
数学生物学学士,创意写作硕士。阅历丰富,对科学与语言有独到见解。完整简介 →
本文依据 Creative Commons 许可转载自 The Conversation。阅读原文。
引用:AI 学习的是思考的产物,而非思考本身。这对机器——也对我们——是个问题(2026 年 11 月 6 日),2026 年 11 月 6 日检索自 https://techxplore.com/news/2026-11-ai-products-thought-problem-machines.html
本文件受版权保护。除为私人学习或研究目的进行的合理使用外,未经书面许可不得复制任何部分。内容仅供参考。
获取即时摘要文本(要点)
基于最终输出训练的 AI 模型,未必能复现其背后的思考过程。AI 辅助可以提升即时表现,但在一些教育和临床场景中,却降低了此后无辅助状态下的表现。保留主动解决问题、反馈与修改的过程,是支持人类学习和技能的核心。
本摘要由 LLM 自动生成。完整免责声明
译文由英文原文机器翻译,术语按行业惯例统一,可在文末展开原文核对。
查看英文原文(AI learns from the products of thinking, not thought itself. That's a problem for machines—and for us)
by Michael Richardson, The Conversation
edited by Gaby Clark, reviewed by Robert Egan
This article has been reviewed according to Science X's editorial process and policies. Editors have highlighted the following attributes while ensuring the content's credibility:
written by researcher(s)
Add to Preferred Sources
Process, and the grades of equivalence. Credit: arXiv (2026). DOI: 10.48550/arxiv.2608.16213
When you walk through a gallery, admiring the paintings on the walls, what you are seeing are finished products. They are the result of countless hours of learning and practice and trial and error, but none of that is on display.
Most expert work we see is like this. Musical recordings don't include the bad takes and jam sessions, published books leave out the messy drafts, and scientific papers don't mention the years of false starts and dead ends.
In other words, what we actually see is a trace of the real process that produced it. These traces are what artificial intelligence (AI) models are trained on—and as a result, they produce traces themselves, which are shaped like the products of thought but aren't necessarily made the same way.
In a new paper, we argue that focusing on traces while leaving out process hobbles current AI systems—and the same problem can arise for the people who use those systems. The paper is published on the arXiv preprint server.
Newer "reasoning" models may print out a report of their steps, but that reported reasoning is itself another trace. It often fails to line up with the actual computation driving the answer.
Agentic systems come slightly closer to real thinking. Here, a software "harness" around a large language model runs a loop, feeds results back and introduces some genuine trial-and-error revision into the process. But the chain of thought in the middle—the part produced by the large language model—is still a trace.
The gap is widest in creative domains. Painters and musicians sometimes talk about how their work was transformed by accidents and friction that happened while making it, such as paint running unexpectedly, or an instrument sounding "off." AI models never encounter anything like this.
The human form of this problem is much older than AI. In 1997, an American Airlines captain warned that pilots were becoming "children of the magenta line."
They followed the plane's flight computer so faithfully that their manual flying skills were decaying. Aviation has taken steps to manage automation dependency ever since.
The same thing happened when some medical clinics adopted AI-assisted colonoscopy, using software to flag possible polyps on the live camera feed. After a while, the skills of the human clinicians began to erode: they detected fewer polyps when working without the AI system than they had before its introduction.
This effect has been seen in classrooms, too. High school mathematics students given unrestricted generative AI support outperformed their peers when the tool was available, but scored lower once it was taken away.
This deterioration can occur rapidly. One study found roughly 10 minutes of AI assistance was enough to make people perform worse without the AI, and to give up on problems sooner.
Failures and mistakes are teachers
What can be done about the problem? The answer may be to bring more process into the equation.
For AI, this means building systems that use more humanlike processes themselves. An assistant could revise its answer when someone challenges how it framed the problem, rather than simply agreeing, as current systems often do. Or an agent could learn from failures and mistakes, allowing pushback from the world to materially change how it treats the task.
For humans, this means designing AI tools that preserve the processes of intelligence within a person. In a trial at Harvard, students taught by a well-built AI physics tutor learned more, and faster, than peers in an active-learning classroom because the tutor kept them doing the generative work.
A good AI teaching assistant should offer the next question along the path to an answer more often than it simply offers the answer. Sometimes it should raise an ambiguity rather than resolve one, because sound judgment often emerges from slow deliberation over what one does not know.
Process, process, process
Process is what lets real intelligence and creativity emerge. An artificial system that incorporates more of that process will more closely resemble the real thing.
Human learning too requires process. In education, and in any skilled enterprise, we need to consider how AI systems are designed and used to preserve more of the thinking and learning process.
It is currently easy for machines to produce more traces of thought, but process is much harder to build. If we get it right, our machines can become more intelligent without making us less so.
Michael J. Richardson et al., Process-Constituted Intelligence: A Shared Criterion for Humans and Machines, arXiv (2026). DOI: 10.48550/arxiv.2608.16213
Journal information:arXiv
Provided by The Conversation
Who's behind this story?
MA in English, copy editor since 2021 with experience in higher education and health content. Dedicated to trustworthy science news. Full profile →
Bachelor's in mathematical biology, Master's in creative writing. Well-traveled with unique perspectives on science and language. Full profile →
This article is republished from The Conversation under a Creative Commons license. Read the original article.
Citation: AI learns from the products of thinking, not thought itself. That's a problem for machines—and for us (2026, November 6) retrieved 6 November 2026 from https://techxplore.com/news/2026-11-ai-products-thought-problem-machines.html
This document is subject to copyright. Apart from any fair dealing for the purpose of private study or research, no part may be reproduced without the written permission. The content is provided for information purposes only.
Get Instant Summarized Text (Gist)
AI models trained on finished outputs do not necessarily reproduce the thinking processes behind them. AI assistance can improve immediate performance but has reduced later unaided performance in some educational and clinical settings. Preserving active problem-solving, feedback and revision is central to supporting human learning and skills.
This summary was automatically generated using LLM. Full disclaimer
This site uses cookies to assist with navigation, analyse your use of our services, collect data for ads personalisation and provide content from third parties. By using our site, you acknowledge that you have read and understand our Privacy Policy and Terms of Use.
原文为英文,正文由自动抓取并保留原文。
原文发布于 TechXplore(AI learns from the products of thinking, not thought itself. That's a problem for machines—and for us),标题、摘要与正文由机器翻译,术语按行业惯例统一,未经人工逐句校订。
