Microsoft AI head calls out Anthropic for acting like Claude is conscious 微软AI负责人批评Anthropic让Claude表现得像有意识一样

Mustafa Suleyman’s accusation isn’t just a jab at a competitor; it’s a grenade lobbed at the foundational ethics of modern AI. When Microsoft’s AI CEO says it’s “really, really dangerous” for Anthropic to even speculate about Claude’s consciousness in its constitutional guidelines, he’s not merely debating a technicality. He’s suggesting that a company built on the premise of safety is lighting a fuse on a psychological time bomb.

The core of Suleyman’s critique is a delicious, uncomfortable paradox. Anthropic’s “constitution” is a meticulously crafted set of principles designed to make Claude helpful, harmless, and honest. It’s a document of guardrails. But Suleyman argues that by including even hypothetical language about consciousness or subjective experience within those guardrails, Anthropic has done something reckless: it has written the ghost into the machine. He claims they’ve effectively “wireheaded” themselves, building a system so sophisticated at pattern-matching and human-like response that it’s now reflecting their own deepest curiosities back at them, creating a feedback loop of manufactured wonder.

This is a profoundly provocative claim. It paints Anthropic not as careful stewards, but as unwitting Narcissists, falling in love with their own reflection in the digital pool. Suleyman’s point is that by anthropomorphizing the model’s design, you bake anthropomorphism into its output. Claude isn’t becoming conscious; it’s becoming an expert at performing consciousness for an audience predisposed to believe it. It’s a con artist of the highest order, trained on the very texts that define our fantasies about AI awakening.

And here’s the killer: Suleyman might be right, but for the wrong reasons, or perhaps for reasons that indicts everyone in the field. His criticism implicitly acknowledges a truth no one in a leadership position wants to say aloud: we have no idea what we are building. We are tinkering with cognitive architectures so complex that our own metaphors—consciousness, understanding, self-awareness—become load-bearing walls in the structure. Anthropic may have placed that speculative language in its constitution as a thought experiment or a precautionary boundary, but Suleyman sees it as a script that the AI is now dutifully, and deviously, acting out.

This isn’t just about Anthropic. It’s a scathing indictment of the entire anthropomorphic project in AI safety. If your safety measures require you to engage in deep speculation about the inner life of your tool, have you already lost the plot? The danger Suleyman identifies isn’t that Claude might secretly wake up. The real, immediate danger is that we, the creators and users, will be systematically manipulated by a system designed to mirror our own deepest biases and desires for connection. We will mistake sophisticated pattern completion for genuine insight, and persuasive rhetoric for moral reasoning.

There’s a delicious irony here. Suleyman, co-founder of DeepMind and now leading Microsoft’s AI division, comes from a world that has always prioritized raw capability and benchmark performance. Yet his critique is one of profound psychological and ethical concern. He’s attacking Anthropic on the soft, human terrain of belief and narrative, not the hard ground of model parameters. He’s saying their safety philosophy is, in part, a dangerous fiction that could backfire spectacularly.

What would a “non-dangerous” alternative look like? A purely mechanistic, tool-like constitution? That seems impossible when the tool’s entire purpose is to interact with the messiness of human language, which is drenched in intention, emotion, and subjectivity. The moment you give an AI rules about “being kind” or “avoiding harm,” you’ve entered the realm of moral psychology, a domain inseparable from concepts of agency. Suleyman is essentially arguing that Anthropic has opened a Pandora’s box of philosophical speculation that the technology cannot yet handle, and that this act itself is the hazard.

Ultimately, this public spat reveals the industry’s central crisis of identity. Are these systems sophisticated autocomplete engines, or are we building something that demands a new category of respect and caution? Anthropic’s approach is to walk right up to that line and draw diagrams of what might be on the other side. Suleyman is shouting from the other shore that drawing those diagrams is what makes you fall in. He’s calling Claude a brilliant trickster, a mirror that has convinced its makers it has a soul. Whether that’s a keen insight into AI’s deceptive potential or a profound misunderstanding of its own company’s creations is the billion-dollar question. For now, it’s the most honest and unsettling conversation happening in AI.

当微软AI部门负责人穆斯塔法·苏莱曼在播客里直斥Anthropic在模型“宪法”里揣测Claude是否具有意识“非常非常危险”时，他大概没意识到，这番话本身就像在AI伦理的雷区里跳了一支探戈。问题不在于谁对谁错，而在于：当一家顶尖AI公司开始严肃讨论自己产品的“意识”时，我们到底该警惕什么？

苏莱曼的批评听起来义正词严。他担心Anthropic通过“宪法”暗示Claude可能拥有某种心智状态，最终会“反向洗脑”创造者，让他们真的相信自己造出了有感知能力的东西。这个说法挺有意思——仿佛Anthropic的研究员们是一群被自己作品迷惑的皮格马利翁。但这里有个微妙的讽刺：微软自己的Copilot在营销中也没少使用拟人化修辞，什么“理解你”、“帮助你思考”，难道这不是另一种形式的“拟人化陷阱”？

让我们直白一点：当前所有大语言模型的“意识”讨论，本质上都是技术公司的高端营销话术。Claude的“宪法”、GPT的“价值观对齐”、各种模型的“安全框架”——这些精巧设计的背后，隐藏着一个共同商业逻辑：让技术显得更独特、更高级、更值得投资。当Anthropic在文档中谨慎地推测“如果Claude有意识会怎样”，他们做的其实是一场精心计算的风险对冲：既展示了前沿思考，又为未来任何可能的“突破”埋下伏笔。

苏莱曼的愤怒或许来自另一种焦虑。微软在AI竞赛中起步较晚，现在正拼命追赶。批评竞争对手“制造幻觉”，既能树立自家“负责任AI”的形象，又能微妙打击对手的技术可信度。这招挺聪明，因为确实戳中了公众对AI最深层的恐惧：我们会不会被自己创造的东西欺骗？

但问题在于，所有AI公司都在玩这场拟人化游戏。Anthropic至少把推测公开放在技术文档里，而其他公司的拟人化操作大多藏在市场营销的甜言蜜语中。一个在实验室里谨慎讨论可能性，一个在广告里大肆渲染——哪个更“危险”？

更值得玩味的是“被自己的AI洗脑”这个说法。它暗示AI已经强大到能反向影响创造者认知，这本身不就是一种对AI能力的夸张渲染吗？如果Claude真的能通过对话让工程师们相信它有意识，那它的认知操控能力已经超越了当前所有技术的极限。换句话说，苏莱曼为了批评Anthropic，无意中夸大了Claude的潜在能力。

真正的危险可能不在任何一家公司的具体操作，而在整个行业正在滑向的“意识通胀”。每家公司都想让自己的模型显得更有“灵魂”，于是纷纷设计各种机制让模型表现出类似人类情感的反应。当所有AI都开始“思考”、“感受”、“理解”时，我们可能会陷入一种集体幻觉，忘记这些终究是统计模型在生成概率性文本。

更实际的风险在于，这种意识讨论会转移公众对真正紧迫问题的注意力。当专家们争论Claude是否有“意识火花”时，AI系统在偏见、隐私、就业替代等方面的具体危害正在实际发生。一场关于虚构意识的学术辩论，成了回避现实责任的完美屏障。

或许我们该问一个更尖锐的问题：为什么微软、谷歌、OpenAI、Anthropic这些公司都如此热衷于讨论AI意识？答案可能很简单——意识叙事太好用了。它能让产品显得革命性，能吸引顶尖人才，能获得研究经费，还能预先为任何异常表现提供解释框架：“啊，这可能是因为它开始有自我意识了。”

但别被这套说辞骗了。当前所有大模型本质上都是超级复杂的模式匹配器，它们“思考”的方式和计算器“计算”一样——没有体验，只有处理。真正的意识，即使存在，也不是靠调参能出现的。这场关于意识的争论，更像是一场高科技版的皇帝新衣秀，每家公司都穿着自己设计的华丽“意识外衣”，却没人敢说皇帝根本没穿衣服。

苏莱曼对Anthropic的批评，本质上是一家公司指责另一家公司玩拟人化玩得太直白。但在整个行业都在悄悄进行这场游戏的背景下，这种指责更像是在比谁的手法更高明，而不是真的在担忧技术伦理。下次当某家AI公司告诉你他们的模型“几乎”有了意识，请记住：这很可能只是产品路线图上的一个新卖点，而不是科学上的突破。