为什么 AI 无法像人类一样「思考」

作者:Didier Camilo Olarte Gonzalez、Pier-Luc de Chantal,The Conversation
编辑:Lisa Lock,审阅:Andrew Zinin
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,强调了以下属性:
由研究人员撰写
添加到首选来源
当人类拿起一个咖啡杯时,我们会基于对世界的理解进行预测性处理。对机器人而言,拿起一个杯子则需要大量训练与仿真。图片来源:Pavel Danilyuk / Pexels
如今的人工智能技术在生成看似极具人类特征的文本和想法方面已变得非常出色。由 ChatGPT、Claude、Google Gemini 和 Perplexity AI 等大语言模型驱动的聊天机器人,能够写诗、头脑风暴,并为复杂问题找到解决方案。
这些产品的开发者大量借鉴了人脑的灵感,基于人类神经元的工作方式构建了人工神经网络。他们的系统如今产生的输出,往往与人类思维的输出相似。
然而,AI 与人类心智的运作方式并不相同。结果相似,未必意味着过程相似。
“世界模型”是一项新兴技术,它可以帮助 AI 系统学习预判可能发生的情况,并利用这种预测来采取行动,从而更接近人类的“思考”方式。尽管如此,AI 仍不会像人类那样去感知、学习或理解世界。
一只掉落的杯子,一次复杂的预测
想象你伸手去拿放在台面边缘的一只杯子。如果你误判了距离、握得太松或动作太快,杯子可能掉落摔碎。
对人类而言,这样抓取杯子是家常便饭。但这一简单动作背后,却依赖着预判下一步可能发生什么的复杂能力。我们的大脑不断通过理解事件及其成因,努力做到先人一步。我们能预见到移动一只装得过满的杯子会弄成什么烂摊子,也能在买沙发之前就想象出它卡在门口的样子。
这些预测依托于我们在与世界互动中形成的预期。反过来,我们所预判的内容也会影响我们感知世界的方式以及我们的行动。
有一个颇具影响力的理论流派叫作预测处理(predictive processing),它提出感知在一定程度上正是以这种方式运作的:大脑生成预期,将其与接收到的信息进行比较,并在两者不匹配时作出调整。
让我们回到杯子的例子。想象你伸手去拿一只你以为是满的、实际却是空的杯子。你的手预期这是一个更重的物体,于是用了超出必要的力气去抬它。杯子猛地升起,感觉几乎轻若无物。这种不匹配提供了反馈,帮助你修正动作,并更新你对它重量的估计。
我们的预期塑造着我们的行动,而接下来发生的事又会更新我们的预期。这一切的发生,只是为了拿起一个杯子。
AI 的预测方式是否不同?
即便对今天的机器人而言,要可靠地操控杯子也可能需要大量训练和仿真。这一挑战涉及学习物理世界的行为方式,并预测某个动作会带来怎样的后果。
使用大语言模型的聊天机器人同样在做预测,但属于另一类。它们学习预测下一个词或文本片段。这种看似简单的预测方式,却仍能产生出能够解决复杂问题、能力惊人的系统。
以这句话为例:“一只手碰倒了桌子边缘附近的一个杯子。杯子……”由于从海量数据中学到了模式,聊天机器人可能会接着写:“……从边缘翻落,掉到地上。”它能做到这一点,靠的是从数据集中预测模式,而从未真正用物理方式碰过杯子。
然而,能够描述看似合理的结果,并不一定意味着聊天机器人已经学会了这些结果为何会发生。
教机器人做出预判
这正是世界模型发挥作用的地方——世界模型是一大类 AI 方法,旨在预测情境如何演变,在某些情况下还预测行动可能如何改变情境。这些模型可以从视频、仿真,或机器人所见所为之记录中学习。
对于家用机器人来说,这可能意味着预判抓握过紧会捏碎杯子,或者动作失误会把杯子碰倒。
初步研究表明,通过纳入动作、运动和感官信息,这些系统有助于把预测与物理世界中实际发生的情况联系起来。
世界模型最终或许能对杯子做出与我们相同的预测。而随着它从动作和感官反馈中学习,它与我们的相似之处可能越来越难以忽视。不过,得出相同的预测,并不能告诉我们是什么知识在支撑这一预测,也无法说明这些知识是如何获得的。
相同的预测,相同的认知?
对我们来说,杯子不只是桌上的一个圆柱体。它可能可以被握住、可能被打碎,可能是你的、也可能是我的,可能太烫,或者快要倒了。我们的理解把杯子的物理属性与一个由可能动作和后果构成的网络联系起来。
预测杯子会掉下来,只揭示了这种理解的一小部分。
而这种理解很早就发展起来了。研究表明,仅几个月大的婴儿就已经对物理对象的运动方式有了基本预期。儿童发展研究者 Jean Piaget 提出,儿童通过作用于世界来建构知识。随着这些知识的发展,他们能够在不实际执行每个动作的情况下,在头脑中预判后果。
对人类而言,学习不只是积累关于杯子的信息。它涉及把伸手去拿杯子、感受它的重量、看着它洒出来等经历联系起来,并最终甚至无需亲自尝试就能想象可能发生什么。
借助世界模型,AI 除了可以从视频、模拟和其他数据中学习之外,还能从动作和感官反馈中学习。和人类一样,它可以学会与世界互动。问题在于,它的这些经验如何促成它所构建和使用的知识。
AI 无疑可以更接近预判可能发生什么,并利用这种预测来行动。然而,仅凭这一点,并不能说明 AI 像我们一样感知、学习或理解世界。
随着 AI 在复现人类认知结果方面越来越强,区分表现与过程只会变得更加重要。
由 The Conversation 提供
这篇报道背后是谁?
艺术史学士,物质文化硕士。曾任博物馆编辑、急救员和移植协调员。自 2021 年起为 Science X 从事编辑工作。完整简介 →
物理学硕士,具备科研经历。资深的科学新闻爱好者。在 Science X 的编辑成功中发挥关键作用。完整简介 →
本文依据知识共享(Creative Commons)许可,转载自 The Conversation。阅读原文。
引用:为什么 AI 无法像人类那样“思考”(2026 年 11 月 5 日),2026 年 11 月 6 日检索自 https://techxplore.com/news/2026-11-ai-humans.html
本文档受版权保护。除为私人学习或研究目的进行的任何合理使用外,未经书面许可,不得复制本文档的任何部分。内容仅供参考。
获取即时摘要文本(要点)
大语言模型预测的是文本模式,而人类认知则整合了预期、行动与感官反馈。AI 世界模型可以学会预测物理事件和行动结果。但类似的预测并不能证明 AI 会以与人类相同的过程进行学习、感知或理解。
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(Why AI cannot 'think' like humans do)
Why AI cannot 'think' like humans do
Why AI cannot 'think' like humans do
by Didier Camilo Olarte Gonzalez, Pier-Luc de Chantal, The Conversation
edited by Lisa Lock, reviewed by Andrew Zinin
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
When a human picks up a coffee cup, we use predictive processing based on our understanding of the world. For a robot, picking up a cup requires extensive training and simulation. Credit: Pavel Danilyuk from Pexels
Today's artificial intelligence technologies have become remarkably good at generating text and ideas that seem profoundly human. Chatbots powered by large language models such as ChatGPT, Claude, Google Gemini and Perplexity AI can write poems, brainstorm and find solutions to complex problems.
The developers of these products drew heavily on the human brain for inspiration, creating artificial neural networks based on the workings of human neurons. Their systems now produce outputs that often resemble the outputs of human thinking.
AI and the human mind, however, do not work the same way. Similar results do not necessarily mean similar processes.
"World models" are an emerging technology that could help AI systems move closer to human "thinking" by helping them learn to anticipate what might happen and use that prediction to act. Still, AI will not perceive, learn or understand the world as humanity does.
A falling cup, a complex prediction
Imagine reaching for a cup at the edge of a counter. The cup may fall and smash if you misjudge the distance, grip it too loosely or move too quickly.
For humans, grabbing a cup like this is routine. Yet this simple action relies on the complex ability to anticipate what might happen next. Our minds constantly try to stay one step ahead by understanding events and their causes. We can foresee the mess that would result from moving an overfilled cup or imagine a sofa getting stuck in a doorway even before we buy it.
These predictions draw on expectations that form when we interact with the world. What we anticipate can, in turn, impact how we perceive the world and how we act.
One influential family of theories, known as predictive processing, proposes that perception works partly in this way. The brain generates expectations, compares them to the information it receives and makes adjustments when the two don't match.
Let's return to the example of the cup. Imagine reaching for an empty cup that you believe to be full. Your hand, expecting a heavier object, lifts it with more force than necessary. The cup rises abruptly and feels almost weightless. That mismatch provides feedback that helps correct the movement and update your estimate of its weight.
What we anticipate shapes our actions, and what happens next updates our expectations. All this happens just to pick up a cup.
Does AI predict differently?
Even for today's robots, reliably manipulating cups can require extensive training and simulation. The challenge involves learning how the physical world behaves and predicting the consequences of an action.
Chatbots using large language models also make predictions, but of a different kind. They learn to predict the next word or fragment of text. This apparently simple way of predicting can nevertheless produce remarkably capable systems able to solve complex problems.
Take the sentence: "A hand bumps a cup near the edge of a table. The cup…" Having learned patterns from enormous amounts of data, a chatbot might continue with "tips over the edge and falls to the floor." It can do so by predicting patterns from a data set and without ever physically handling a cup.
However, being able to describe plausible outcomes does not necessarily mean a chatbot has learned why they happen.
Teaching a robot to anticipate
This is where world models kick in—a broad family of AI approaches that aim to predict how situations evolve and, in some cases, how actions may change them. These models can learn from videos, simulations or recordings of what robots see and do.
For a domestic robot, this could mean anticipating that gripping too tightly might break the cup, or that a wrong movement could tip it over.
Preliminary research shows that, by incorporating actions, motion and sensory information, these systems can help connect predictions with what happens in the physical world.
A world model could eventually make the same predictions we do about a cup. And as it learns from actions and sensory feedback, the resemblance to us may become increasingly hard to ignore. Still, arriving at the same prediction does not tell us what knowledge supports it or how that knowledge was acquired.
Same prediction, same cognition?
To us, a cup is not only a cylinder on a table. It can be graspable, breakable, yours, mine, too hot or almost falling. Our understanding connects the cup's physical properties to a network of possible actions and consequences.
Predicting that the cup will fall reveals only a small part of that understanding.
And this understanding develops quite early. Studies suggest that infants just a few months old already have basic expectations about how physical objects behave. Child-development researcher Jean Piaget proposed that children build knowledge through acting on the world. As this knowledge develops, they become able to mentally anticipate consequences without having to carry out each action.
For humans, learning involves more than accumulating information about cups. It involves connecting experiences when reaching for one, feeling its weight, seeing it spill and eventually imagining what might happen without even having to try it.
Using world models, AI can also learn from actions and sensory feedback, in addition to videos, simulations and other data. Like humans, it can learn to interact with the world. The question is how its experiences contribute to the knowledge it builds and uses.
AI could definitely come closer to anticipating what might happen and using that prediction to act. This alone, however, does not show that AI perceives, learns or understands the world as we do.
As AI gets better at replicating the results of human cognition, distinguishing performance from process will only become more important.
Provided by The Conversation
Who's behind this story?
BA art history, MA material culture. Former museum editor, paramedic, and transplant coordinator. Editing for Science X since 2021. Full profile →
Master's in physics with research experience. Long-time science news enthusiast. Plays key role in Science X's editorial success. Full profile →
This article is republished from The Conversation under a Creative Commons license. Read the original article.
Citation: Why AI cannot 'think' like humans do (2026, November 5) retrieved 6 November 2026 from https://techxplore.com/news/2026-11-ai-humans.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)
Large language models predict text patterns, whereas human cognition integrates expectations, actions and sensory feedback. AI world models can learn to predict physical events and action outcomes. Similar predictions do not establish that AI learns, perceives or understands through the same processes as humans.
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(Why AI cannot 'think' like humans do),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。
