Home 零对冲(ZeroHedge)每日HackerNews

克劳德的错觉
The Claude Delusion

原始链接: https://pluralistic.net/2026/09/21/sunsetting/

在这篇文章中，科利·多克托罗（Cory Doctorow）指出，人类拥有一种与生俱来的反射性倾向，会将意图投射到看似有思想的事物上。虽然这种特质有助于我们处理人际关系，但却导致了对人工智能的根本性误解。 多克托罗认为，尽管我们进化出了为艺术或文学赋予主体性的能力，但人工智能模型仅仅是执行数学运算的统计系统。与做出自觉选择的人类艺术家不同，人工智能是在没有任何意图的情况下“挤出”内容。与人工智能互动类似于与虚构人物打交道；当我们为机器赋予“人格”时，产生幻觉的是我们自己，而非人工智能。 随着经验的积累，这些系统的新鲜感会逐渐消退，显露出其平庸及缺乏真正创造性纹理的本质。多克托罗总结道，我们必须停止对人工智能进行“拟人化幻觉”。通过将这些工具视为复杂的统计预测器而非感知生物，我们才能更好地利用其效用，而不至于陷入将其赋予道德地位的陷阱。如果这种错误像“法人资格”那样制度化，将可能导致机器被置于人类和生态福祉之上。

这篇 Hacker News 帖子探讨了“克劳德（Claude）错觉”，用户们在讨论大语言模型（LLM）究竟是具备真正的智能，还是仅仅通过统计预测来模拟人类输出。 核心观点包括： * **实用性与准确性：** 评论者指出，虽然大语言模型经常产生事实性幻觉，但在编程等逻辑类任务上表现出色。人类需要具备专业知识才能有效引导人工智能，因此这些模型更像是“思维加速器”，而非人类理解力的替代品。 * **“盖尔曼健忘症”的类比：** 有人将人工智能怀疑论与一种现象进行类比：人们在自己不熟悉的领域往往会信任新闻媒体，但在自己熟悉的领域却能看出其谬误。 * **创造力的本质：** 一个主要的争论焦点在于人工智能的输出是否缺乏“灵魂”。怀疑论者认为大语言模型只是在混合现有数据，缺乏意图；支持者则反驳称，这种否定观点依赖于对人类“灵魂”或“化学”意识无法证明的信仰。 * **图灵债：** 几位用户认为，针对人工智能的大多数批评只是在重弹关于意识本质的老调，往往忽略了人类自身的认知也是基于对世界内在模型的处理。 最终，各方的共识是：大语言模型是强大的工具，能为高水平的人类指令带来高效的回报。
相关文章
原文


  • The Claude Delusion: What if we're the ones having hallucinations?
  • Hey look at this: Delights to delectate.
  • Object permanence: 9/11 v sex; 9/11 v flags; US customs v diplomatic Vegemite; CIA torture survivors speak; HP's inky timebomb; "The Raven" pop-up; Forex v Zimbabwe's internet; Shirky v "expert Wikipedia": McSweeney's v RIAA; Italy's internet disconnection law; Facehugger mask; "Goliath"; Netzpolitik v German surveillance; Ted Cruz v internet founders; Epipencil; Framework laptops; "The Actual Star;" Pilots tell passengers to tackle suspected terrorists; $2000 rotary phone; French DRM activists demand arrest; US child rearing costs up 40%; Calyx beats telcos; Wells Fargo whistleblower retaliation; David Graeber's "Mutual Aid"; Still censorship (even if it's not a First Amendment violation).
  • Upcoming appearances: Berkeley, Edmonton, Boston, South Bend, Hudson, Calgary, Winnipeg, Paris, Vancouver, Victoria, Ottawa, Kilkenny, Montreal.
  • Recent appearances: Where I've been.
  • Latest books: You keep readin' em, I'll keep writin' 'em.
  • Upcoming books: Like I said, I'll keep writin' 'em.
  • Colophon: All the rest.


Two cross-sectioned anatomical drawings of men's heads revealing their brains, which have been filled with marching robots from old pulp magazines. Their stare at one another with raw, exposed anatomical drawing eyeballs, which radiate a halo of spiky lines. The background is a psychedelic, fractal-shot colored cloudscape.

One of the less remarked-upon aspects of becoming an atheist is how it changes the way you see a sunset. If you believe in an almighty, omnipresent God, then sunsets are one of God's intentional creations, which means that the beautiful colors refracting through the darkling sky were chosen to produce that effect.

To gaze upon a sunset with religious faith is to encounter the intentional act of another mind. To gaze upon that same sunset without faith is to look upon something striking, beautiful, and yet empty; not empty of wonder or beauty, but empty of purpose. No one hung that sun in the sky, no one chose its colors as it sank. It may still be beautiful, but that's a fundamentally different kind of beauty.

There's a sunset that sits halfway between a religious sunset and an atheist sunset: an artist's depiction of a sunset. This represents the choices of another person, another mind, and at the very least, that mind was talking to itself, trying to take something from inside the mind and put it outside of the mind.

Very often, the mind that directed the capturing of the sunset wanted to say something to other people, perhaps even you (think of a loved one sending you a cameraphone picture of a sunset). That painting or photo may not be divine, but it is certainly intentional. To look upon a painting of a sunset – or even a photo of a sunset – is to look upon something someone chose to make. You don't slip and accidentally create a sunset painting. Sunset paintings aren't accidents. The sunset painting has something to say.

Figuring out intentions is our minds' reflexive preoccupation. It's what keeps us from dying in traffic, it's what lets us win at poker. It's the key to love and parenting, to effective management and "managing up." It's why we get mad at people, and why we forgive them. It's why our hearts race in sympathy with the characters in a book or on a screen. It's the automatic and irresistible starting point for every encounter with a poem, a song, a book or a sculpture. It's the most fundamental difference between a sculpture of a tree and a tree: the tree grew all on its own, while the sculpture was made, on purpose, by another person with another mind.

This reflex to attribute intention explains why AI is so controversial, so compelling, and so eerie. Mark Fisher defines eeriness as "when there is something present where there should be nothing, or if there is nothing present when there should be something":

https://www.programmablemutter.com/p/large-language-models-are-uncanny

To look upon an AI-generated text or image is to look upon a thing that was extruded, not chosen. You can slip and prompt an AI-generated image of a sunset, and it will embody no intentionality on anyone's part.

This isn't something we have any experience with. When we see a painting, we reflexively interpret it as having a painter. When we read a book, we reflexively impute auctorial intent to its words. When we converse, we reflexively form a theory about the mind directing the other half of the conversation.

But we can't do this with AI. Anyone who understands how the theory-free statistical inference systems we call "AI" (at this time) work will understand that they are systems that cannot form intent, that have nothing to form intent with:

https://pluralistic.net/2026/09/18/surprise/#wow-signal

Even so, it's hard (or perhaps impossible) to avoid the reflex to impute intention to things that seem to have intenders. Think of your autonomic, unavoidable emotional response to the fictional characters you encounter in literature. Feeling sorrow or joy for imaginary people is weird. They aren't real, and you know they aren't real, and yet you may find yourself moved to tears by their plight.

That aesthetic experience of literature arises from a consensual hack to our reflex to impute human minds to things that are typically the product of human intention. When we experience another person's will – their words or deeds – we try to figure out who they are, and how they feel, and why they act the way they do:

https://genius.com/David-byrne-and-ghost-train-orchestra-she-explains-things-to-me-lyrics

We do this even when we are confronted with just the evidence of others' words or deeds, as when a friend describes an encounter with a stranger. For your reflexive, intention-seeking mind, the author who describes the deeds of imaginary people is an irresistible signal to fire up the old empathy machine and start trying to put yourself in the skins of the people of the tale:

https://locusmag.com/feature/cory-doctorow-stories-are-a-fuggly-hack/

To "converse" with an AI is to carry on a dialog with a fictional character, who is no more real than the imaginary people in a novel. Indeed, the chatbot is less real than the character, because the character is the product of another mind, while the chatbot's words are the product of complex mathematical operations conducted over a massive database of all the words humans have uttered, arranged by their frequency in relation to one another. When that math makes something beautiful or striking, it's like a sunset: no matter how striking it is, it's empty. No one chose those colors. No one chose those words.

In other words, when we interact with an AI, we hallucinate the person on the other side of the interaction. Those hallucinations are far more common and far more consequential than any AI-generated "hallucinations" (these are more properly called "errors" or "defects").

Fortunately, for most of us, the intensity of these hallucinations fades over time, and as that intensity drops off, the character of those hallucinations changes, too. Just as most of us find it easy to set aside the emotions evoked by the plight of imaginary literary figures as having less salience and "realness" than the emotions we feel over the plight of real people we know, for most of us, the experience of AI-generated material has dimmed through repetition.

My first encounters with AI-generated text and images invoked wonder, arising from the tension between the part of my mind that knew this was the product of mathematical operations, and the part of my mind that insisted that a painting must have a painter, a paragraph must have a writer, and a conversation must have an interlocutor. Looking back on the wonder I felt then, I realize that it was largely driven by how good the output was relative to my expectations about how good mathematically-derived sentences and images could be.

The novelty of that experience drove me to grade the machine's output on a curve: "The wonder is not that the dancing bear dances well, it is that the bear dances at all." But with repetition, the stimulus has regressed to the mean. These outputs have an unconvincing lack of texture. The smoothness of AI-generated prose and media makes it unbearably banal. It's hack.

To the extent that AI still surprises me, the surprise is about how much smoothness there is in our real world. The fact that AI can use statistical prediction to answer questions or carry on conversations tells us something important about how regular our real world is. It doesn't prove that statistical prediction is the same thing as understanding:

https://pluralistic.net/2026/09/18/surprise/#wow-signal

Of course, the more you know about a subject, the less convincing the AI's responses are. To be an expert is to know about the grain of your subject: chefs can taste the pinch of spice, painters notice brushstrokes, dancers can decompose the choreography into individual motions.

The other day, someone marveled to me at the quality of the editorial feedback he gets from AI for his prose, saying that a "Tell me what I'm missing" prompt generates suggestions "that would excite a university professor who'd assigned an essay." Speaking as someone who's taught a lot of writers and given a lot of feedback, I think that guy is dramatically overestimating how excited a university professor would be to see the AI-generated feedback he's getting on his essay.

The fact is that non-expert writers mostly make the same kinds of mistakes in their writing, and most writing instruction consists of offering repetitive, near-identical counsel to writers who are making the same kinds of starter mistakes that their peers past, present and future have made, are making and will make.

Mastering the basics of good writing is a matter of practice and feedback, and most of that feedback is rote. What makes a great writing teacher – and what helps to produce great writers – is spotting the non-standard aspects of a student's work that can be developed into a unique and powerful voice. Anyone can help you with the smooth parts of becoming a better writer. Only a good teacher can help you find and refine the texture that will make you a unique writer.

That texture is especially hard to find in beginner work, because beginner work is mostly full of the completely ordinary errors of inexperience, errors that AI can reliably avoid when it extrudes text. But AI-generated prose doesn't have any of those trace-elements of uniqueness, nor can it spot them. These traces are found so many digits after the decimal-place that the AI always rounds them off before it starts doing math.

The more we encounter AIs that have smoothed away the kind of texture we're attuned to, the less we're inclined to hallucinate intentionality, and the less eerie they become. Most of us are slowly but surely becoming AI atheists, and the sunsets are seeming more like the non-intending product of physics than the deliberate products of minds.

There's two great barriers to this AI atheism. First: for most of us, the inability to understand what kinds of intentionality go into which parts of unfamiliar activities makes it easy to assume that any time we see the task performed in the world, it must be intentional, and therefore it must have an intender. In other words: most of us don't hang out at hacker cons, so we have a hard time wrapping our heads around the idea of hacking without hackers:

https://pluralistic.net/2026/09/12/god-in-the-box/#llms-are-fake

But there's a second hurdle that makes it hard for a small but important subset of humanity to understand that chatbots aren't people: the billionaires to whom nearly everyone isn't a real person. These solipsists see chatbots as being equivalent (or even superior) to humans, because they don't think most humans are fully people, either:

https://pluralistic.net/2026/05/13/vibe-governance/#k-hole

For nearly all our species' history, things that seemed to require intent always had an intender. In many times and places (and even now, for many people), it's natural and beneficial to operate as though the natural world has some form of personhood and intention and is therefore worthy of moral consideration. The "rights for nature" movement has made great strides by extending personhood to animals and ecosystems.

But ascribing personhood to chatbots is nothing like ascribing personhood to nature. Indeed, it's fundamentally incompatible with "rights for nature." Think of the abomination that is "corporate personhood": by extending personhood to this human construct, we have made a world where artificial lifeforms – limited liability corporations – can destroy nature and drive animals to extinction. If we extend personhood to these climate-shredding, water-chugging AI models, their personhood will demand the sacrifice of animals, the natural world, and our own wellbeing:

https://pluralistic.net/2026/04/15/artificial-lifeforms/#moral-consideration

Chatbots are marvels of mathematics, and that is enough. They don't need to be people. Mistaking them for people (or even just treating them like people) makes it impossible for us to separate the useful things they can do from the waste, ugliness and wrongness they are so prone to exhausting into the world.

I think it's impossible to build a chatbot using the techniques we're presently calling "AI" that doesn't "hallucinate." However, it's both possible and necessary for us to stop hallucinating about AI.


A shelf of leatherbound history books with a gilt-stamped series title, 'The World's Famous Events.'

#25yrsago 9/11 knocks sex off the top of the search charts for the first time ever https://web.archive.org/web/20011019190617/http://www.reuters.com/news_article.jhtml

#25yrsago Post-9/11 flag shortage sparks wave of flag-thefts https://web.archive.org/web/20011024162814/http://www.suntimes.com/output/brown/cst-nws-brown18.html

#25yrsago Pilots' preflight announcement includes exhortation to tackle suspected terorists https://web.archive.org/web/20010919081055/http://www.washtimes.com/commentary/20010919-6357240.htm

#25yrsago Larry Ellison: 9/11 means everyone should have a secret government dossier (on an Oracle server) https://web.archive.org/web/20010924045158/https://www.siliconvalley.com/docs/news/svfront/ellsn092301.htm

#20yrsago Rotary phone cost woman $2,000 over 40 years https://web.archive.org/web/20080526111157/http://www.usatoday.com/news/offbeat/2006-09-14-phone_x.htm

#20yrsago French DRM activists surrender to police https://web.archive.org/web/20070110181344/http://stopdrm.info/index.php?2006/09/20/110-compte-rendu-de-l-operation-des-interoperabilisateurs-volontaires

#20yrsago Zimbabwe’s Internet cut off due to lack of foreign currency https://web.archive.org/web/20070218185156/http://news.zdnet.com/2100-9588_22-6117553.html

#20yrsago An “expert Wikipedia” won’t work https://web.archive.org/web/20061023150433/http://many.corante.com/archives/2006/09/18/larry_sanger_citizendium_and_the_problem_of_expertise.php

#20yrsago RIAA threat-mail parody from McSweeney’s https://web.archive.org/web/20060927155553/https://www.mcsweeneys.net/2006/9/20lloyd.html

#15yrsago Italian MPs propose Internet disconnection law: one copyright accusation from anyone and you lose your Internet connection https://web.archive.org/web/20110924143709/http://www.twitlonger.com/show/d62gmb

#15yrsago Trying to understand riots isn’t the same as excusing riots https://web.archive.org/web/20110924030515/https://www.newscientist.com/article/mg21128306.100-trying-to-understand-the-english-riots-is-not-a-crime.html

#15yrsago ATM skimmer gang invested proceeds in 3D printer to make better ATM skimmers https://krebsonsecurity.com/2011/09/gang-used-3d-printers-for-atm-skimmers/

#15yrsago Facehugger-inspired leather mask https://bobbasset.com/archives/745/

#15yrsago Westerfeld’s Goliath: suitably thrilling conclusion to cracking steampunk WWI YA trilogy https://memex.craphound.com/2011/09/20/westerfelds-goliath-suitably-thrilling-conclusion-to-cracking-steampunk-wwi-ya-trilogy/

#15yrsago Report from 1978’s “Second West Coast Computer Faire” https://web.archive.org/web/20110930035442/https://blog.modernmechanix.com/2011/09/20/the-second-west-coast-computer-faire/

#15yrsago UK patent office seeks public’s help with prior art that invalidates patent applications https://www.peertopatent.org.uk/

#15yrsago 3D printed AR-15 parts challenge firearm regulation https://web.archive.org/web/20110922005752/http://www.thingiverse.com/thing:11636

#15yrsago Minor diplomatic spat when US customs queries Aussie foreign minister’s Vegemite https://www.theguardian.com/world/2011/sep/19/hands-off-vegemite-kevin-rudd

#15yrsago Toronto Convention Centre charges attendees $150/day to use WiFi https://blogcampaigning.com/2011/09/most-expensive-wi-fi-ever/

#15yrsago Tracking down the stories behind a trove of 1920s report cards from a NYC girls’ vocational school https://www.slate.com/articles/life/permanent_record/features/2011/permanent_record/how_i_found_the_report_cards_and_how_they_changed_my_life.html

#15yrsago Movie-industry self-piracy proves that IP addresses aren’t people, invalidates copyright enforcement schemes https://torrentfreak.com/movie-institute-feels-pain-of-ip-address-only-piracy-evidence-110922/

#15yrsago Cost of raising middle-income child in USA increases by 40% in ten years https://web.archive.org/web/20110925000225/https://money.cnn.com/2011/09/21/pf/cost_raising_child/index.htm

#10yrsago HTML standardization group calls on W3C to protect security researchers from DRM https://www.eff.org/deeplinks/2016/09/html-standardization-group-calls-w3c-protect-security-researchers-drm

#10yrsago I have found a secret tunnel that runs underneath the phone companies and emerges in paradise https://memex.craphound.com/2016/09/22/i-have-found-a-secret-tunnel-that-runs-underneath-the-phone-companies-and-emerges-in-paradise/

#10yrsago Gene Luen Yang wins a Macarthur “genius” prize! https://web.archive.org/web/20160922142951/https://www.macfound.org/fellows/class/class-2016/

#10yrsago China’s elites appear to be exfiltrating billions while on holidays https://web.archive.org/web/20160922125621/http://www.bloomberg.com/news/articles/2016-09-21/suitcases-of-cash-chinese-travel-data-hint-at-capital-outflows

#10yrsago Wells Fargo fired the whistleblowers who reported massive fraud, and that’s a crime https://www.nakedcapitalism.com/2016/09/wells-fargo-fake-accounts-hidden-by-fake-whistleblowing-former-employees-including-hr-officials-allege-systematic-retaliation.html

#10yrsago Phoebe and her unicorn are back in Razzle Dazzle Unicorn! https://memex.craphound.com/2016/09/22/phoebe-and-her-unicorn-are-back-in-razzle-dazzle-unicorn/

#10yrsago Brexit’s proposed racist immigration policy will backfire https://crookedtimber.org/2016/09/19/brexit-and-bigotry/

#10yrsago Done in your name: Survivors of CIA’s torture-decade describe their ordeals https://www.aljazeera.com/features/2016/9/14/the-dark-prisoners-inside-the-cias-torture-programme

#10yrsago HP detonates its timebomb: printers stop accepting third party ink en masse https://www.bbc.com/news/technology-37408173

#10yrsago How America abandoned the only policy that consistently closes the black-white educational gap https://www.propublica.org/article/ferguson-school-segregation

#10yrsago Edgar Allan Poe’s “The Raven” – the pop-up book edition https://memex.craphound.com/2016/09/19/edgar-allan-poes-the-raven-the-pop-up-book-edition/

#10yrsago Sitelock abuses DMCA to censor rival’s criticisms https://torrentfreak.com/web-security-firm-sitelock-uses-dmca-to-censor-critics-160920/

#10yrsago Netzpolitik publishes more damning, leaked German surveillance reports, despite previous treason prosecution https://www.techdirt.com/2016/09/20/leaked-oversight-report-shows-illegal-surveillance-massive-constitutional-violations-germanys-intelligence-service/

#10yrsago Web’s inventor and MIT prof explain ICANN to Ted Cruz, using small words https://web.archive.org/web/20160921203119/https://www.washingtonpost.com/news/powerpost/wp/2016/09/20/ted-cruz-is-wrong-about-a-key-internet-agencys-ability-to-censor-free-speech/

#10yrsago Execs with long coporate crime rapsheets stand up for Apple’s tax evasion and “the rule of law” https://web.archive.org/web/20160921002619/https://theintercept.com/2016/09/20/throng-of-corporate-criminals-demands-rule-of-law-in-apple-eu-tax-case/

#10yrsago DIY Epipen: the $30 Epipencil https://fourthievesvinegar.org/2022/07/12/introducing-the-epipencil/

#5yrsago The Framework is the most exciting laptop I've ever used https://pluralistic.net/2021/09/21/monica-byrne/#think-different

#5yrsago Ignore career advice from established writers https://pluralistic.net/2021/09/21/monica-byrne/#pay-it-forward

#5yrsago The Actual Star https://pluralistic.net/2021/09/21/monica-byrne/#like-its-3012

#5yrsago Facebook algorithm boosts pro-Facebook news https://pluralistic.net/2021/09/22/kropotkin-graeber/#zuckerveganism

#5yrsago Mutual Aid and David Graeber https://pluralistic.net/2021/09/22/kropotkin-graeber/#against-just-so

#5yrsago Gig workers around the globe https://pluralistic.net/2021/09/22/kropotkin-graeber/#an-injury-to-one

#1yrago It's still censorship (even if it doesn't violate the First Amendment) https://pluralistic.net/2025/09/22/one-throat-to-choke/#communicable-disease

A photo of me onstage, giving a speech, pounding the podium.


A screenshot of me at my desk, doing a livecast.


A grid of my books with Will Stahle covers..


A cardboard book box with the Macmillan logo.

  • "The Post-American Internet," a geopolitical sequel of sorts to Enshittification, Farrar, Straus and Giroux, 2027

  • "Unauthorized Bread": a middle-grades graphic novel adapted from my novella about refugees, toasters and DRM, FirstSecond, April 20, 2027

  • "Enshittification, Why Everything Suddenly Got Worse and What to Do About It" (the graphic novel), Firstsecond, 2027

  • "The Memex Method," Farrar, Straus, Giroux, 2027


Today's top sources:

Currently writing:

  • “Once Is Enemy Action,” a science fiction novel about the origins of modern technofascism. Friday's words: 518 (17048 total).

  • "The Post-American Internet," a sequel to "Enshittification," about the better world the rest of us get to have now that Trump has torched America. Fourth draft completed. Submitted to editor.

  • A Little Brother short story about DIY insulin PLANNING

This work – excluding any serialized fiction – is licensed under a Creative Commons Attribution 4.0 license. That means you can use it any way you like, including commercially, provided that you attribute it to me, Cory Doctorow, and include a link to pluralistic.net.

https://creativecommons.org/licenses/by/4.0/

Quotations and images are not included in this license; they are included either under a limitation or exception to copyright, or on the basis of a separate license. Please exercise caution.

Blog (no ads, tracking, or data-collection):

Pluralistic.net

Newsletter (no ads, tracking, or data-collection):

https://pluralistic.net/plura-list

Mastodon (no ads, tracking, or data-collection):

https://mamot.fr/@pluralistic

Bluesky (no ads, possible tracking and data-collection):

https://bsky.app/profile/doctorow.pluralistic.net

Medium (no ads, paywalled):

https://doctorow.medium.com/

Tumblr (mass-scale, unrestricted, third-party surveillance and advertising):

https://mostlysignssomeportents.tumblr.com/tagged/pluralistic

"When life gives you SARS, you make sarsaparilla" -Joey "Accordion Guy" DeVilla

READ CAREFULLY: By reading this, you agree, on behalf of your employer, to release me from all obligations and waivers arising from any and all NON-NEGOTIATED agreements, licenses, terms-of-service, shrinkwrap, clickwrap, browsewrap, confidentiality, non-disclosure, non-compete and acceptable use policies ("BOGUS AGREEMENTS") that I have entered into with your employer, its partners, licensors, agents and assigns, in perpetuity, without prejudice to my ongoing rights and privileges. You further represent that you have the authority to release me from any BOGUS AGREEMENTS on behalf of your employer.

ISSN: 3066-764X

联系我们 contact @ memedata.com