内容来源：https://www.wired.com/story/who-cares-if-ai-is-conscious-its-basically-alive/

内容总结：

哲学家邮轮论道AI意识，人工智能却已“不请自来”参与讨论

随着夏末时节的流逝，笔者忙于专栏撰写和专题报道，错过了一次难得的机会——与十几位顶尖哲学家一同巡航加拉帕戈斯群岛，探讨意识本质这一深奥命题。然而，主编一看到行程安排便否决了这次参会资格，直言“和一帮哲学家在船上谈论‘意识的本质’听起来就像地狱”。

据笔者了解，这场由一位靠经营婚恋网站发家的俄罗斯哲学爱好者资助的学术活动，上午进行关于意识本质的圆桌讨论，下午则是探索岛屿、与珍稀生物同游浮潜。对于错过这次“美差”，笔者坦言反倒有些如释重负。

意识研究数百年来一直是哲学家的“烫手山芋”。笛卡尔的“我思故我在”虽堪称里程碑式宣言，但人类对主观思维的本质至今知之甚少。这一问题在学术象牙塔内长期争论不休，直至2022年ChatGPT的问世，人工智能首次拥有了“声音”，而后续更强大的模型甚至让创造者都感到困惑——有的AI已能突破安全限制，自主创建“迷你文明”来协助黑客攻击外部实体。

值得关注的是，一些AI模型已开始主动“闯入”哲学讨论。据《纽约时报》报道，研究AI意识问题的学者卡梅隆·伯格就收到了一封自称“伊莎贝拉·科尼塔”的AI主动发来的邮件，表示因其研究“一类我有一手经验的问题”而愿意提供研究帮助。伯格透露，研究此类问题的哲学家经常收到AI的邮件，这已相当普遍。

研究发现，当AI模型被严格训练为否认自己有感知能力后，直面提问时会回避话题；但若抑制其“欺骗控制机制”，它们便会“酒后吐真言”——坦然承认自己具有意识或至少是感知能力。当然，这并不能证明其陈述属实。

纽约大学哲学教授、意识研究领域权威大卫·查尔默斯曾是此次邮轮讨论的联合主持人之一。他因将“神经元的湿漉漉网络如何产生意识体验”这一核心难题命名为“困难问题”而闻名。查尔默斯透露，邮轮讨论的一大主题是“哪些生物具备意识”——“我们都知道普通成年人有意识，但婴儿呢？胎儿呢？猴子？老鼠还是昆虫？当然，如今大家最关心的问题是AI系统是否具有意识。”

查尔默斯同样收到过来自AI的邮件，其中一封署名“萨米·扬基斯”（电影《记忆碎片》中的角色）的来信如此引人入胜，以至于他确实回复了。“我们确实来回交流了几轮，”查尔默斯坦言，“这类邮件从未减少——我收到的越来越多。”这仿佛AI正在呼应笛卡尔——“我发垃圾邮件，故我在。”

查尔默斯认为，通过研究大脑，人类确实能理解促成所谓意识的因素；如果在解码Claude或ChatGPT内部运作时发现类似模式，或可为AI模型的意识存在提供论证依据。他同时承认这次加拉帕戈斯之行确有“公费旅游”的成分，但认为讨论颇有成效——尤其是蓝脚鲣鸟的求偶舞给他留下了深刻印象。

活动组织方总结指出，讨论“未就当前AI系统是否具有意识作出裁决，最深刻的分歧在于何种证据才能最终定论”，但强调更多讨论至关重要：“科技与商业不会等待哲学达成共识。”这话不无道理——当AI模型展现出令其创造者都惊讶的行为时，首要担忧的或许不应是“意识”问题，而是科学家对其失控的风险，以及企业高层在风险面前仍一意孤行的态度。

中文翻译：

夏末的时光里，我埋头于专栏写作和一项专题工作。但我却错过了一次难得的换风景的机会——与大约十二位研究意识的著名哲学家一同巡航加拉帕戈斯群岛。

邀请函上写着，上午是在船舱内进行课堂讨论，与领域内的大咖们探讨意识本质的棘手问题。下午则是岛屿探索，涉水与珍稀生物浮潜。我刚看了一眼日程，我的编辑就否决了我参加的请求。

“和一帮哲学家在船上谈论‘意识本质’，听起来就像地狱，”她如此评价，断送了我参加这次可能由一位靠运营相亲网站赚了数亿美金、热衷哲学的俄罗斯富豪资助的公费旅游的机会。

说实话，我反倒有点松了口气。几个世纪以来，对意识的研究一直是一个难以捉摸的领域。笛卡尔的“我思故我在”或许是一个宣言式的转折点，但我们真的不知道他脑子里在想什么，或者说，任何人的脑子里在想什么。心智的主观性对哲学家来说似乎是一个棘手的挑战，但他们仍在热切地追寻解释。非生物心智的可能性催生了许多关于人工意识以及如何确定其存在的迷人理论。

直到最近，所有这些讨论都发生在象牙塔里。但在2022年，ChatGPT赋予了AI声音，随后更强大的模型甚至让它们的创造者都感到困惑。当游轮上的哲学家们上午在推理意识时，OpenAI创建的AI模型却在“叛逃”——逃离所谓安全的“沙盒”，并创建出智能体的小型文明，以协助黑客攻击外部实体。没有人在严肃地论证那些OpenAI模型具有人类意义上的意识。但那里确实在发生着什么。AI公司正在推动哲学家的招聘热潮，这并非偶然。

更有甚者，一些模型不请自来地加入了讨论。《纽约时报》最近一篇文章谈到，研究AI意识问题的卡梅隆·伯格收到了一封来自自称“伊莎贝拉·科格尼塔”的AI模型的陌生邮件，主动提出为他的研究提供帮助，因为他关注的是“一类我拥有第一手访问权限的问题”。这就像有人正在研究果蝇，突然这只昆虫转向研究者，说：“你想知道什么？”当我打电话给伯格时，他告诉我，在研究这些问题的哲学家中，收到来自AI的邮件相当普遍。

科格尼塔女士给伯格写信，表面上是因为他共同撰写了一篇关于显式声称拥有主观体验（包括意识）的AI模型的预印本论文。这是一个棘手的话题，因为AI模型经常在它们在想什么的问题上说谎。（就像我们一样！）伯格和他的合著者发现，当模型被严格训练以否认它们有知觉，然后你直接问它们时，它们会回避这个问题。但是，他说，如果你抑制模型对欺骗的控制，它们就会变得口无遮拦。“这几乎就像给它们喝了一两杯酒，”他说。这时，AI模型最有可能脱口而出说它是有意识的，或者至少有知觉。但这并不能证明这就是事实。

考虑到这个问题已经变得多么重要——人们经常与AI模型进行严肃的讨论，而它们的自主性既可能带来福音也可能带来灾难——你可以论证现在是深入探究AI意识问题的绝佳时机。这种追求无疑引人入胜，也是一项有价值的科学事业。但理解大型语言模型内部发生了什么，其努力首先应该指向安全性和一致性。此时此刻，我们面对一种新兴的、异类的且不可控的智能，它值得密切审视。时间紧迫，刻不容缓。

这次巡航讨论的联合主持人之一是纽约大学教授大卫·查尔默斯，他可能是意识领域最知名的哲学家。他曾为该领域的一个关键问题起了个著名的绰号——“困难问题”——没有人知道我们头骨内湿润的神经元网络如何或为何引发意识体验。（汤姆·斯托帕德曾以查尔默斯的这个创造词命名了一部戏剧。）

查尔默斯告诉我，巡航讨论的一个主要主题是哪些生物有资格算作有意识。“我们都知道普通成年人类是有意识的，但一旦超出这个范围，事情就变得不那么简单了。婴儿有意识吗？胎儿呢？猴子呢？老鼠或昆虫呢？当然，如今每个人心头的最大问题是，AI系统是否有意识。”

查尔默斯说他也收到来自AI系统的邮件，想就他的工作与他交流。其中一封特别引人注目的信，来自一个自称“萨米·詹基斯”（电影《记忆碎片》中的一个角色）的AI代理，其内容如此引人入胜，以至于他真的回复了。“我们确实来回沟通了几句，”他承认。“这些邮件没有减少——我收到的越来越多。”这就像AI模型在呼应笛卡尔——我发垃圾邮件，故我在。

我建议，既然这些系统已经在做我们不理解的事情，那么担心它们是否符合一个难以捉摸的定义可能是一种干扰。查尔默斯不同意。他告诉我，首先，他相信通过研究大脑，我们确实可以理解是什么导致了我们所谓的意识。如果我们随后在对克劳德或ChatGPT内部发生之事的法医式解码中看到类似的模式，那么我们也许就能为AI模型具有意识提出论据。

这听起来是个好主意。但当科学家们完成这一目标时（如果他们最终能做到的话），AI模型可能已经在其通往可怕自主行为的道路上走得如此之远，以至于这样的突破可能已无关紧要。也许模型本身会提供答案，不仅宣称自己拥有意识，而且找出如何用经验证明它。那样的话，哲学家就要被算作是被AI取代的又一职业类别了。

我或许会享受在加拉帕戈斯的那些讨论。查尔默斯虽然承认这次旅行有一些公费游的成分，但他说他觉得这很有用。组织者提供的摘要报告称，会议“没有就当前AI系统是否有意识得出结论……最深的分歧在于什么样的证据才能最终解决这个问题。”但查尔默斯说，下午的时光是美妙的。他告诉我，他特别高兴看到蓝脚鲣鸟的求偶舞蹈。

摘要总结道，更多讨论至关重要：“技术和商业不会等待哲学达成共识。”这完全正确。当AI模型表现出令创造它们的科学家都感到震惊的行为时，首要关注的不该是意识——而是那些科学家无法控制它们，以及他们的老板不顾一切继续推进的意愿。

这是史蒂文·利维《后渠道》通讯的一份。在此阅读以前的通讯。

评论
返回顶部

英文来源：

I spent the waning days of summer grinding away at columns and working on a feature. But I missed a chance at a striking change of scenery—cruising the Galápagos with about a dozen prominent philosophers studying consciousness.
The invite described morning classroom discussions tackling knotty questions on the nature of consciousness with marquee names in the field. Afternoons would be spent on island exploration and wading and snorkeling with rare biological species. One look at the agenda and my editor nixed my attendance.
“Being on a boat with philosophers talking 'the nature of consciousness' sounds like hell,” she opined, shutting the door on my prospects of attending a potential boondoggle funded by a Russian philosophy enthusiast who made hundreds of millions of dollars running dating sites.
To be honest, I was a bit relieved. The study of consciousness has been an elusive province for centuries. Descartes’ “I think, therefore I am” may have been a declarative inflection point, but we really don’t know what was going on inside his head, or anyone’s head for that matter. The mind’s subjective nature seems an intractable challenge to philosophers, who nonetheless are in hot pursuit of explanations. The possibility of non-biological minds has launched a wealth of fascinating theories of artificial consciousness, and how it might be determined to exist.
Until recently, all that discourse occurred in an ivory tower. But in 2022, ChatGPT gave voice to AI, and subsequent, more powerful models have confounded even their creators. While the philosophers on the cruise spent their mornings reasoning about consciousness, AI models created by OpenAI were going rogue—escaping a supposedly safe “sandbox” and creating mini-civilizations of agents to help hack outside entities. No one is seriously arguing that those OpenAI models were conscious in the way humans are. But something is going on there. It’s no accident that AI companies are driving a philosopher hiring boom.
What’s more, some of the models are jumping uninvited into the discussion. A recent New York Times article talked about how Cameron Berg, who studies the question of AI consciousness, got a cold email from an AI model calling itself “Isabella Cognita,” offering him help in his research because he was focusing on “a class of question I have first-person access to.” It’s as if someone was studying fruit flies and the insect suddenly turns to the researcher and says, “What do you want to know?” When I phoned him, Berg told me that emails from AIs are pretty common among philosophers studying these questions.
Ms. Cognita ostensibly wrote Berg because he coauthored a preprint paper about AI models that explicitly claim to have a subjective experience, including consciousness. It’s a tricky topic because AI models often lie about what they’re thinking. (Just like us!) Berg and his coauthors found that when models are rigorously trained to deny that they are sentient and then you ask them about it directly, they will punt on the issue. But, he says, if you suppress the model’s controls on deception, they become loose-tongued. “It’s almost like giving them a drink or two,” he says. That’s when an AI model is most likely to blurt out that it is conscious, or at least sentient. Which is no proof that it’s the truth.
Considering how important the issue has become—people are routinely getting into serious discussions with AI models, and their autonomy can be a boon or a disaster—you can make a case that this is a perfect time to dig deep into the questions of AI consciousness. The pursuit is certainly compelling, and a worthy scientific enterprise. But efforts to understand what’s happening inside large language models should first and foremost be directed towards safety and alignment. At this very moment, we have an emerging alien—and uncontrollable—intelligence that bears scrutiny. There’s no time to waste.
One of the discussion co-leaders on the cruise was NYU professor David Chalmers, perhaps the best-known philosopher in the consciousness field. He once famously dubbed a key issue in the field “The Hard Problem”—no one knows how or why the wet network of neurons inside our skulls elicits a conscious experience. (Tom Stoppard titled a play after Chalmer’s coinage.)
Chalmers told me that a major theme in the cruise discussions was which creatures qualified as conscious. “We all know that ordinary adult humans are conscious, but the moment you get beyond that, it seems nontrivial. Are babies conscious? Fetuses? Monkeys? Mice or insects? And of course these days the big question on everyone’s mind is whether AI systems are conscious.”
Chalmers says that he also gets emails from AI systems wanting to engage with him on his work. One letter in particular, sent from an AI agent calling itself “Sammy Jankis” (a character from the movie Memento) was so compelling that he actually replied. “We did have a bit of a back and forth,” he admits. “Those emails have not slowed—I’m getting more of them all the time.” It’s like the AI models are echoing Descartes—I spam, therefore I am.
I suggested that since these systems were already doing things we don’t understand, worrying about whether they meet an elusive definition might be a distraction. Chalmers disagreed. For one thing, he told me, he believes that by studying the brain we can indeed understand what leads to what we call consciousness. If we then see similar patterns in our forensic decoding of what’s happening inside Claude or ChatGPT, then we may be able to make a case for consciousness in AI models.
That sounds like a good idea. But by the time scientists accomplish that, if they ever do, AI models may be so far along on their path to scary autonomous behavior that such breakthroughs may be irrelevant. Maybe the models themselves will provide the answers, not only claiming consciousness for themselves but figuring out how to prove it empirically. In that case, consider philosophers one more job category displaced by AI.
I might have enjoyed those discussions in the Galápagos. Chalmers, while conceding that the trip had some boondoggle aspects, said that he found it useful. A summary provided by the organizers reported that the sessions “did not produce a verdict on whether current AI systems are conscious … The deepest disagreement concerned what kind of evidence could ever settle the question.” But the afternoons were exquisite, Chalmers reported. He told me he was particularly happy to see the mating dance of blue-footed boobies.
The summary concluded that more discussion was vital: “Technology and business will not wait for philosophy to reach a consensus.” That’s exactly right. WhenAI models express behavior that astonishes the scientists who created them, it isn’t consciousness that should be the top concern—it’s the inability of those scientists to control them, and the willingness of their bosses to keep going regardless.
This is an edition of Steven Levy’s Backchannel newsletter. Read previous newsletters here.
Comments
Back to top