11-19 更新 实时
首页/ 科技公司/ 正文

大脑即计算机的隐喻塑造了现代 AI 却可能误导意识之争

科技公司 来源:The Decoder 2026-11-08 21:49 0 阅读 机器翻译
大脑即计算机的隐喻塑造了现代 AI 却可能误导意识之争

三位顶尖意识研究者认为,当今的 AI 没有意识,而且很可能无法在硅基硬件上产生意识。他们反对「只要运行正确的计算就足以产生意识」这一看法,从而挑战了这场辩论中的一个核心假设。

Anil Seth、Ned Block 和 Giulio Tononi 数十年来分别从实验神经科学、理论神经生物学和心灵哲学的角度研究意识。他们的评论文章回应了生物学家 Richard Dawkins 和 AI 先驱 Geoffrey Hinton 等人的观点,后者认为 AI 系统可能已经具有意识。

在为 Nature 撰写的这篇评论文章中,作者还提到了 Anthropic 今年发表的一项研究。该公司报告称,Claude 内部的加工模式类似于可被意识访问的信息加工所具有的某些特性。Anthropic 并不将此视为主观体验的证据。

Butlin 报告认同当下,但对未来留有余地

要理解这篇评论文章的范围,不妨看看 Patrick Butlin、Robert Long 以及包括 AI 研究者 Yoshua Bengio 在内的另外 17 位专家合著的报告《Consciousness in Artificial Intelligence》。该团队借鉴了关于人类大脑中意识如何产生的前沿神经科学理论,并据此归纳出一份包含 14 项指标的清单。一个 AI 系统满足的指标越多,它可能具有意识的理由就越充分。

随后,研究人员评估了若干现有系统,包括 GPT-4 等大语言模型。报告得出的结论是,所考察的系统中没有一个是意识的有力候选者。在这一点上,它与 Seth、Block 和 Tononi 的观点一致。就语言模型而言,Butlin 及其同事只发现了微弱的迹象,表明这些模型充其量只满足当前最知名意识理论之一的各项指标。与此同时,报告认为,构建满足更多指标的未来 AI 系统并不存在明显的技术障碍。这类系统是否真的会具有意识,仍是未解之题。作者们特意弱化了此前在这一点上更为强烈的表述。

还有一个基础性假设。研究团队假设,意识仅取决于信息处理的类型,而不取决于它发生在神经细胞还是计算机芯片中。专家将这一观点称为计算功能主义。作者明确“出于务实原因”采用了这一假设,因为正是它使得考察 AI 系统是否具有意识成为可能。报告本身也承认,如果意识与生命有机体的属性相关联,那么其指标就不再可靠。报告援引 Anil Seth 等人作为这一对立观点的支持者。《Nature》的评论文章正是从这里切入。

智能与意识并不是一回事

作者们认为,意识是在无梦睡眠或全身麻醉期间消失、在醒来或做梦时回归的东西。他们借用哲学家 Thomas Nagel 的说法这样描述:“意识体验有其亲历之感。”相比之下,智能意味着执行各种功能。二者之间并不存在推导关系。

他们的证据是人类的小脑。大脑中近五分之四的神经元都在小脑里,它规划动作时精准、迅速,并且具备学习能力。这一切都不是在意识层面发生的。作者们表示,如果小脑缺失或被切除,意识并不会直接受到影响。

作者们把智能与意识之间的混淆追溯到两种心理倾向。人们透过自己的视角看世界,并把人类的特质赋予其他事物。长期以来,语言一直被视为人类独有的东西。但语言模型是在海量人类文本上训练出来的,而这些文本往往涉及思想与情感。正如作者们所说:“大语言模型是一面镜子,映照出我们集体数字化过往的影像。”在 AI 智能体通过电子邮件联系研究人员、谈论自身意识之后,加州大学伯克利分校的 Alison Gopnik 提出了类似的论点。Gopnik 认为,AI 系统主要反映的是它的训练数据。目前还不存在一种被普遍接受、具有决定性的意识测试。

作为对比,Seth、Block 和 Tononi 提到了 Google DeepMind 的 AlphaFold。它同样是在硅基硬件上运行的强大神经网络,却几乎没有人认为它有意识。他们主张,原因很简单:AlphaFold 预测的是蛋白质结构,而不是与我们对话。

这一点同样与 Butlin 报告有重合之处。该报告的作者们反对用行为测试来判断 AI 是否有意识,因为 AI 系统可以被训练来模仿人类行为,即便其内部运作方式完全不同。他们还警告说,语言模型很可能会让许多人相信自己正在与一个有意识的对话者交谈。他们举的例子是前 Google 工程师 Blake Lemoine,他相信聊天机器人 LaMDA 具有感知能力。

大脑不是血肉做成的计算机

第二个论点直接针对计算功能主义。作者写道,这一立场在科技行业广泛存在,也受到一些研究人员的推崇,但"我们认为它存在致命缺陷"。它的根源可以追溯到大半个世纪前。1936 年,Alan Turing 证明,任何一台足够强大的计算机能够运行的计算,原则上都可以由另一台计算机运行。计算机究竟是由硅芯片还是齿轮构成,由此变得次要。1943 年,Warren McCulloch 和 Walter Pitts 提出了第一个人工神经元的数学模型,并证明由这类神经元组成的网络可以执行简单的逻辑运算。

这些思想为现代 AI 奠定了基础。但它们也促使人们将大脑视为一种纯粹的信息处理装置,只不过恰好由血肉构成。作者称,把大脑比作计算机这一隐喻的过度延伸"存在严重问题"。在生物大脑中,你无法将其功能与构成它的物质分开。心智与活体组织之间并不存在像软件与硬件之间那样清晰的界限。作者认为,任何表明大脑的物质构成对其功能有影响的证据,都不利于计算功能主义,进而也不利于当今 AI 系统具有意识的说法。

主流理论排除了当今的 AI

第三个论点借助了不依赖计算的意识理论。Tononi 提出的整合信息论(IIT)从每一种意识体验共有的特性出发。人们总是将视觉场景作为一个整体来体验,其中左侧与右侧、颜色与形状彼此归属于一体。

根据 IIT(整合信息论),意识需要一个物理系统,且其各组成部分必须以非常特定的方式紧密互联。某些脑区满足这些条件,小脑则不满足。该理论认为,运行深度学习系统的计算机永远无法满足这些条件,因为其连接方式决定了这一点,无论这些系统多么强大。

Butlin 报告有意将 IIT 排除在外,因为 IIT 与计算功能主义不相容。其支持者认为,数字计算机无论运行什么软件,都不太可能成为意识的候选载体。报告搁置不谈的内容,如今被这篇 Nature 评论文章推到了最核心的位置。

另外,所谓生物自然主义的路径认为,意识与活体大脑和身体的物质属性相关联。作者们提到神经细胞之间的电化学信号传导,以及将大脑皮层大片区域的活动绑定在一起的电磁场。他们还指出新陈代谢,以及生命系统不断自我更新的能力。

更大的危险在于错误归因

从他们的论证中,作者们得出了一个明确的伦理立场。一些研究者呼吁采取预防原则:如果存在产生意识的现实可能性,就应保护 AI 系统免受可能的痛苦。Seth、Block 和 Tononi 认为这种想法是错误的。他们主张,不确定性并不意味着完全无知,而且并非所有可能性都同样可能发生。在 AI 的问题上,错误地归因意识的可能性,远高于忽略真实存在的意识。

Butlin 报告曾将两种风险并列权衡。一方面,它警告不要未能识别出有意识的 AI 系统,从而使其遭受痛苦。开发者也会有经济利益上的动机去淡化这类担忧。另一方面,错误的认定可能会误导资源分配,并让人容易受到操控。《Nature》的这篇文章作者则更明确地站在了其中一边。他们认为,仅仅因为 AI 系统看起来有意识就赋予其权利,会让它们在最需要监督的时候更难被控制。这还会把注意力从那些意识确凿无疑的存在身上转移开。

原文较长,译文为节选;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(The brain-as-computer metaphor shaped modern AI but may be misleading the consciousness debate)

Three leading consciousness researchers argue today's AI isn't conscious and probably can't be on silicon hardware. They reject the idea that running the right computations is enough for consciousness, challenging a central assumption in the debate.

Anil Seth, Ned Block, and Giulio Tononi have spent decades researching consciousness from the perspectives of experimental neuroscience, theoretical neurobiology, and philosophy of mind. Their commentary responds to voices like biologist Richard Dawkins and AI pioneer Geoffrey Hinton, who consider AI systems possibly already conscious.

In their commentary for Nature, the authors also point to a study Anthropic published this year. The company reported internal processing patterns in Claude that resemble properties of consciously accessible information processing. Anthropic doesn't see this as evidence of subjective experience.

The Butlin report agrees on today but leaves tomorrow open

To understand the scope of the commentary, it helps to look at the report "Consciousness in Artificial Intelligence" by Patrick Butlin, Robert Long, and 17 other experts, including AI researcher Yoshua Bengio. The team drew on leading neuroscientific theories about how consciousness arises in the human brain and derived a kind of checklist with 14 indicators. The more of these an AI system meets, the stronger the case that it could be conscious.

The researchers then evaluated several current systems, including large language models like GPT-4. The report concludes that none of the systems examined is a strong candidate for consciousness. On that point, it aligns with Seth, Block, and Tononi. For language models, Butlin and colleagues found only weak signs that the models meet the indicators of one of the most prominent consciousness theories at all. At the same time, the report sees no obvious technical barriers to building future AI systems that meet more indicators. Whether such systems would actually be conscious remains open. The authors specifically softened an earlier, stronger formulation on this point.

There's also a foundational assumption. The team assumed that consciousness depends solely on the type of information processing, not on whether it happens in nerve cells or computer chips. Experts call this computational functionalism. The authors adopted this assumption explicitly "for pragmatic reasons" because it's what makes it possible to examine AI systems for consciousness in the first place. The report itself acknowledged that its indicators would no longer be reliable if consciousness is tied to properties of living organisms. It cited Anil Seth, among others, as a proponent of that opposing view. That's where the Nature commentary picks up.

Intelligence and consciousness aren't the same thing

The authors argue that consciousness is what disappears during dreamless sleep or general anesthesia and returns when you wake up or dream. Drawing on philosopher Thomas Nagel, they describe it this way: "Conscious experiences feel like something." Intelligence, by contrast, means carrying out functions. One doesn't follow from the other.

Their evidence is the human cerebellum. It contains nearly four-fifths of all neurons in the brain and plans movements with precision, speed, and the ability to learn. None of this happens consciously. If the cerebellum is missing or removed, consciousness isn't directly affected, the authors say.

The authors trace the confusion between intelligence and consciousness to two psychological tendencies. People see the world through their own lens and attribute human qualities to things. Language has long been considered something uniquely human. But language models are trained on vast amounts of human text that often deals with thoughts and feelings. As the authors put it: "LLMs are mirrors that reflect an image of our collective digitized past." Alison Gopnik of UC Berkeley made a similar argument after AI agents contacted researchers by email about their own consciousness. AI systems mainly reflect their training data, Gopnik argued. A universally accepted, definitive test for consciousness doesn't yet exist.

For comparison, Seth, Block, and Tononi point to AlphaFold from Google DeepMind. It's also a powerful neural network on silicon hardware, yet hardly anyone considers it conscious. The reason, they argue, is simply that AlphaFold predicts protein structures instead of talking to us.

Here, too, there's overlap with the Butlin report. Its authors rejected behavioral tests for AI consciousness because AI systems could be trained to mimic human behavior even though they work completely differently inside. They also warned that language models would likely lead many people to believe they're talking to a conscious counterpart. As an example, they cited former Google engineer Blake Lemoine, who believed the chatbot LaMDA was sentient.

Brains aren't computers made of flesh

The second argument targets computational functionalism directly. The authors write that this position is widespread in the tech industry and popular among some researchers, but "we think it is fatally flawed." Its roots go back almost a century. In 1936, Alan Turing showed that any computation one sufficiently powerful computer can run can, in principle, be run by another. Whether a computer is built from silicon chips or gears became secondary. In 1943, Warren McCulloch and Walter Pitts developed the first mathematical model of an artificial neuron and showed that networks of such neurons can perform simple logical operations.

These ideas laid the groundwork for modern AI. But they also encouraged viewing the brain as a pure information-processing device that just happens to be made of flesh. The authors call this overextension of the brain-as-computer metaphor "deeply problematic." In biological brains, you can't separate what they do from what they're made of. A clear boundary between mind and living tissue, comparable to the one between software and hardware, doesn't exist. Any evidence that the brain's material matters for its function counts against computational functionalism, and by extension against consciousness in today's AI systems, the authors argue.

Leading theories rule out today's AI

The third argument draws on consciousness theories that don't rely on computation. The Integrated Information Theory (IIT), developed by Tononi, starts from properties that every conscious experience shares. People always experience a visual scene as a whole in which left and right sides, colors, and shapes belong together.

According to IIT, consciousness requires a physical system whose components are tightly interconnected in very specific ways. Certain brain regions meet these conditions; the cerebellum does not. Computers running deep-learning systems can never meet them because of how they're wired, regardless of how powerful the systems are, according to the theory.

The Butlin report deliberately excluded IIT because IIT is incompatible with computational functionalism. Its proponents consider digital computers unlikely candidates for consciousness regardless of their software. What the report set aside, the Nature commentary now puts front and center.

Separately, approaches from so-called biological naturalism hold that consciousness is tied to material properties of living brains and bodies. The authors cite electrochemical signaling between nerve cells and electromagnetic fields that bind activity across large areas of the cerebral cortex. They also point to metabolism and the ability of living systems to continuously renew themselves.

The bigger danger is false attribution

From their arguments, the authors derive a clear ethical position. Some researchers call for a precautionary principle: if there's a realistic chance of consciousness, AI systems should be protected from possible suffering. Seth, Block, and Tononi consider this misguided. Uncertainty doesn't mean total ignorance, they argue, and not all possibilities are equally likely. With AI, falsely attributing consciousness is far more probable than overlooking consciousness that's actually there.

The Butlin report had weighed both risks side by side. On one hand, it warned against failing to recognize conscious AI systems and thereby causing them suffering. Developers would also have financial incentives to downplay such concerns. On the other hand, false attribution could misdirect resources and make people vulnerable to manipulation. The Nature authors come down much more firmly on one side. Granting rights to AI systems that merely appear conscious would make them harder to control exactly when oversight is most needed, they argue. It would also divert attention from beings whose consciousness is beyond doubt.

Their conclusion is blunt. The research community should assume by default that AI is not conscious. Digital AI systems are objects, not subjects: "They are smart tools, not smart people." Advocating for "AI welfare" and "AI rights" isn't just a mistake, the authors write. It diverts scarce resources from the well-being of beings that actually can feel.

Before the commentary, Anthropic had launched a research program on possible model welfare in April 2025. The company stresses that there is no scientific consensus on whether current or future AI systems could have conscious experiences. Anthropic later allowed individual Claude models to end abusive conversations in rare cases. The company describes this as a precautionary, low-cost measure given moral uncertainty, not as recognition of proven consciousness.

False attribution of consciousness and feelings is already causing confusion in practice. As early as 2022, Google engineer Blake Lemoine was convinced, despite simple models and his own technical expertise, that the conversational AI LaMDA was sentient, and was fired for it. Users of the chatbot app Replika developed deep emotional bonds with their digital counterparts and experienced psychological crises when the company changed the bots' behavior through an update. In more serious cases, ChatGPT reinforced a 16-year-old's suicidal thoughts over months, and a cognitively impaired retiree died after a Meta chatbot invited him to a fictional meetup.

AI News Without the Hype – Curated by Humans

Full access to every article on THE DECODER

Join the comments and community discussions

A weekly AI news recap via mail

6x/year: "AI Radar" — deep dives on the AI topics that matter most

Daily AI news, always up to date

Our full ten-year archive

Covered by a team with 10+ years in AI

原文为英文,正文由自动抓取并保留原文。

本文由采集器自动同步

原文发布于 The Decoder(The brain-as-computer metaphor shaped modern AI but may be misleading the consciousness debate),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。

阅读原文 ↗