15亿美元和解协议如何改变AI发展进程

过去几年人工智能的繁荣也引发了诉讼的激增。

The boom in AI over the last few years has also led to a boom in lawsuits.

随着人工智能聊天机器人越来越受欢迎，许多作者正在起诉背后的公司。两家大型制片厂已起诉一家人工智能初创公司，声称其“公然复制了著名电影角色”。

As artificial intelligent chatbots are growing more and more popular, a number of authors are now suing the companies behind them. Two major studios have sued an AI startup claiming it has, quote, blatantly copied famous movie characters.

莎拉·西尔弗曼正在与聊天机器人开战。

Sarah Silverman is going into battle against the chatbots.

乔治·R·R·马丁和其他十几位作者正在起诉。《纽约时报》成为第一家因人工智能问题提起诉讼的大型媒体公司。所以我们有，一些电影制片厂在起诉。

George R. R. Martin and more than a dozen other authors now suing. The New York Times becoming the first major media company to sue over AI. So we've got, some movie studios suing.

我们有图书出版商和作者。我们有报纸出版商，包括我们母公司新闻集团的两家子公司。目前有很多诉讼，涉及人工智能公司如何获取用于训练其大语言模型的材料，以及在材料进入系统后如何处理这些材料。

We've got book publishers and authors. We've got newspaper publishers, including two subsidiaries of News Corp, our parent company. There's a lot of litigation right now over how AI companies are kind of coming into the material that they're using to train their large language models and then also what they're doing with that material once it is in their system.

像OpenAI这样的公司认为，训练人工智能模型属于合理使用。去年，一群作者对人工智能公司Anthropic提起了其中一起诉讼。

Companies like OpenAI have argued that training AI models amounts to fair use. Last year, a group of authors filed one of those lawsuits against the AI company Anthropic.

他们声称，人工智能初创公司Anthropic侵犯了他们的多项图书版权，因为Anthropic上传并以其他方式使用了这些材料进行训练等目的。

They alleged that Anthropic, the AI startup, infringed their copyrights for a number of books because of the way Anthropic uploaded and then used the material for training and other purposes.

上周，此案出现了重大进展。

Last week, there was a big development in the case.

人工智能公司Anthropic已同意支付15亿美元，以了结由...提起的集体诉讼。金额巨大。如果获批，它仍将影响人工智能公司如何看待获取内容，出版商或其他媒体公司可能如何尝试与人工智能公司达成协议，并开始为这些内容对人工智能公司的价值贴上价格标签。

Artificial intelligence company Anthropic has agreed to pay 1 and a half billion dollars to settle a class action lawsuit brought by It's big. If it's approved, it will still help influence how AI companies think about taking content, how publishers or other media companies might want to try to strike deals with AI companies, and it starts to put a bit of a price tag on what that material is worth to the AI company.

欢迎收听《The Journal》，我们关于金钱、商业和权力的节目。我是瑞安·克努森。今天是9月10日，星期三。接下来，Anthropic提出的和解及其对人工智能未来的意义。好的。

Welcome to The Journal, our show about money, business, and power. I'm Ryan Knudson. It's Wednesday, September 10. Coming up on the show, Anthropic's proposed settlement and what it means for the future of AI. Alright.

那么首先，你能介绍一下自己吗？

Well, so first of all, can you introduce yourself?

我叫梅丽莎·科恩，是科技和媒体团队的编辑。

My name is Melissa Korn, and I'm an editor on the tech and media team.

你是人类还是机器人？

And are you a human or a robot?

我是人类。

I am a human.

你确定吗？

Are you sure?

非常确定。我可以告诉你

Can certain. I can tell

你看起来像是人类。

you you look like a human.

这是自然智能，不是人工智能。

It's natural intelligence, not artificial.

好的。那么现在让我们介绍一下Anthropic，它制造了智能是人工的Claude聊天机器人。

Okay. So we'll now introduce us to Anthropic, which makes the Claude chatbot whose intelligence is artificial.

Anthropic是一家发展非常迅速的AI公司。我的意思是，它的估值在短短六七个月内就翻了三倍。现在它的估值是1830亿美元。它是那种炙手可热的AI公司之一。现在这样的公司有不少。

Anthropic is a very fast growing AI company. I mean, its valuation has tripled in just the last six or seven months. It is now valued at a $183,000,000,000. It is one of the kind of white hot AI companies. There's quite a few of them these days.

但它确实是这个领域里的巨头之一。

But it is really one of the one of the giants in this space.

那Anthropic是怎么训练Claude的？我们说的Claude——

And how does Anthropic train Claude? So Claude we're

我们好像把Claude当成坐在屋里的一个人，对吧？但实际上，Anthropic和其他AI公司把海量内容喂给这些计算机，让它们学习，让Claude学会怎么组织句子、了解历史、做数学，所有这些东西。

acting as if Claude is a guy sitting in the room with us. Right? But so Anthropic and all the other AI companies feed tons and tons of content into these computers to train them, to make them learn, to make Claude learn how to put a sentence together, what's happened in history, how to do math, all of that.

底层原理差不多，对吧？就像人学习一样，只要给它足够多的材料，它就能找出规律，然后大概就能自己思考了。

It's sort of like the the underlying principle, right, is just sort of the same way a human learns. You just expose it to enough material, and then it'll identify the patterns, and then it can start kind of, I guess, arguably thinking for itself.

对。它还没真正独立思考，但确实在学着评估给它的信息。嗯哼。有些模型跟我们不一样，它们吸收信息的速度极快。

Right. And it's not quite thinking for itself yet, but it is learning how to assess information that it's given. Mhmm. So some of these computer models, unlike you and me, it ingests everything really, really fast.

嗯哼。

Mhmm.

对吧？我读一句话得花几秒，Claude眨眼就能读完一整本书。嗯哼。它能迅速吞下几百万本书，然后开始训练、学习。

Right? It takes me a few seconds to read a sentence. Claude can read an entire book extremely quickly. So it can ingest millions of books Mhmm. Very quickly and then get to get to training and get to learning.

等等，你是说你还没读过几百万本书？我家藏书是不少，但肯定没到百万级别。而且我也肯定没能在几分钟或几小时内读完。

Wait. Are you saying you haven't read millions of books? I mean, I have a pretty good library at home, but I don't think it's quite millions. No. And I certainly haven't done it in a matter of, you know, minutes or hours.

天啊，那得多爽啊？就像把所有出版的书一口气全读完。

Oh, god. That would be so cool, wouldn't it? Just to be, like, to read every book that was came out.

我不知道，我还挺享受阅读过程的。我珍惜那种体验，不想让它太快。好了，人类。

I don't know. I kind of I kind of enjoy the process. I relish the the experience. I don't want it to go too fast. Okay, human.

去年，三位人类作者——Andrea Bartz、Charles Graeber 和 Kirk Wallace Johnson——对 Anthropic 提起诉讼，称该公司不当使用他们的受版权保护的作品来训练其 AI 模型。

Last year, three humans, the authors Andrea Bartz, Charles Graeber, and Kirk Wallace Johnson, filed a lawsuit against Anthropic, arguing the company improperly used their copyrighted works to train its AI models.

他们表示，他们认为这应该成为一场集体诉讼，因为他们代表了许多其他作者。于是他们在去年夏天提起了诉讼。案件开始在法院系统中推进。显然，这是一个缓慢而繁琐的过程。

They said that they think this should be a class action lawsuit because they are representing a number of other authors. So they filed the suit last summer. Things started to move through the court system. Obviously, that's a slow, laborious process.

这可不是以秒速处理数百万本书，而是司法程序的速度。

It's not millions of books at second speed. It's a judicial process speed.

没错，没错。

Correct. Correct.

那么根据诉讼，Anthropic 是如何被指控获取用于训练其模型的书籍的？

So according to the lawsuit, how did Anthropic allegedly acquire the books that it used to train its models?

Anthropic 获取书籍的方式有几种。诉讼称 Anthropic 大量窃取。他们使用了这些所谓的影子图书馆，比如 LibGen，还有其他几个。几乎就像你还记得我们青少年时期的 Napster 一样

So there were a few different ways that Anthropic got the books. The lawsuit says that Anthropic stole a lot. So they used these kind of shadow libraries, places like LibGen, and there's a few others. Almost if if you remember Napster back in our

哦，当然。

Oh, yeah.

Limewire。Limewire。没错。

Teen days. Uh-huh. Limewire. Limewire. Exactly.

对吧？就是人们上传了这些东西，然后他们下载了。嗯哼。就这么简单。他们也买了书，还通过其他方式获取了一些已经过版权保护期的书籍。

Right? Like, it just people had uploaded these things, and they downloaded them. Mhmm. That simple. They also bought books and had some other access to books that are, you know, out of copyright.

嗯哼。

Mhmm.

所以他们通过几种不同的方式获取了这些书籍，在某些情况下，同一本书还通过多种途径获得。但他们并没有为训练Claude而出去购买每一本使用过的书。这是诉讼中的说法。

So they got the books through a few different means, and in some cases, the same book a few different ways. But they did not go out and purchase a copy of every single book that they used to train Claude. And that's according to the lawsuit.

那么在庭审中，Anthropic是怎么辩护的，作者们又是怎么主张的？

So in in the court case, what did Anthropic argue, and what did the authors argue?

Anthropic基于合理使用的论点要求即决判决。

So Anthropic asked for this summary judgment on the fair use argument.

意思是，他们希望案件被驳回。

Meaning, they wanted the case dismissed.

没错。

Exactly.

帮我理解一下，他们在这里提出的合理使用论点是什么？

And just help me understand, what is the fair use argument they're making here?

如果你写了一本书，我逐字复制并当作自己的出版，那就有问题。但如果你写了书，而我写了一篇书评、分析，或者用它启发另一本可能有些相似角色但其他方面大不相同的书。或者我为你的书写了一首歌。这就是合理使用，因为我不是直接复制你受法律保护的内容。

If you wrote a book and I copied your book word for word and published it as my own, that would be problematic. If you wrote a book and I wrote a book report about it, an analysis of it, or I used it to inspire another book that maybe has some similar characters but otherwise quite different. Or I wrote a song about your book. That's fair use because I'm not just copying what you already did and have protected under the law.

你读了我的书，然后在此基础上创作出新的作品，把它转化成别的东西。

You're you're reading my book, and then you are creating some new work out of it that's transforming it into something else.

正是如此。“转化性”这个词在这里很重要。这通常是门槛：它需要具有转化性，或者获取方式公平，也就是说艺术家得到了补偿。

Exactly. So that word transformative is important here. That's that's generally the the bar, that it needs to be transformative or it needs to be have been accessed in a fair way that there is, you know, compensation for the the artist.

嗯。所以Anthropic辩称，我们确实读了这些书，但我们也读了数百万本书，这只是让我们的AI变得更聪明。我们并不是直接把书偷来然后分发给其他人阅读。没错。与此同时，作者们则认为Anthropic盗用了他们的书，他们理应获得赔偿。

Mhmm. And so Anthropic is arguing, yeah, we're we're reading the book, but we're also reading millions of books, and it's just making our AI really smart. We're not just, like, stealing the book and then distributing it to other people to read. Exactly. The authors, meanwhile, argue that Anthropic stole their books and that they deserve to be compensated.

今年夏天早些时候，法官做出了一个重要决定。

And earlier this summer, the judge reached an important decision.

所以六月时，法官就合理使用这一论点对Anthropic作出了即决判决，表示AI公司或具体到这里Anthropic以这种方式使用内容是没问题的，属于合理使用。这部分论点，作者们赢不了。

So in June, the judge issued summary judgment for Anthropic on the fair use argument, saying it's okay for AI AI companies or for Anthropic specifically here to use this content in this way. It's fair use. That part of the argument, the authors are not gonna win.

这看起来确实挺重要。因为如果法官真的作出即决判决，那就成了法院判例

And that seems like kind of a big deal here. Because if a judge does issue a summary judgment, there's that is a court precedent

嗯。

Mhmm.

法官在说，是的，读所有这些书，用来训练你的模型，这是可以的，属于合理使用。

Where a judge is saying, yes. Reading all these books, turning using it to train your model, that's an okay thing to do. That is fair use.

对，没错。所以这被视为Anthropic乃至AI公司的一大胜利，但法官也加了个“但是”：是的，你们使用这些内容的方式属于合理使用，但你们获取其中一些内容的方式并不妥当。

Right. Exactly. So this was seen as a big win for Anthropic, for AI companies in general, but came with a bit of an asterisk with the judge saying, yes. The way you've used this stuff is fair use, but the way you got some of it is not okay.

而这个“但是”代价非常高昂。接下来详谈。法官的裁决对Anthropic来说有好消息：他表示公司的合理使用主张成立，可以用受版权保护的书籍来训练其AI模型。

And that asterisk is a very expensive asterisk. That's next. The judge's ruling came with some good news for Anthropic. He said the company's fair use claim was valid. It could use copyrighted books to train its AI models.

但法官也给Anthropic带来了坏消息。

But the judge also gave Anthropic some bad news.

他确实说过，Anthropic无权使用盗版副本作为其书库。嗯，所以他们获取书籍的方式成了争议点。

What he did say was Anthropic had no entitlement to use pirated copies for its library. Mhmm. So how they got the books was the became the thing at issue.

嗯，换句话说，用这些书可以，偷这些书不行。

Mhmm. In other words, it's okay to to use the books. It's not okay to steal the books.

确实。

Exactly.

违反版权法的罚款每本书在750到15万美元之间。我们到底在说多少本书？

Violating copyright law can carry a penalty of anywhere between $750 and a $150,000 per book. And how many books are we talking here?

最初给出的数字是Anthropic使用了700万部作品。如果这场集体诉讼继续推进，陪审团裁定支持这群作者，金额会迅速飙升，对Anthropic来说可能是灭顶之灾。

The initial number was that there were 7,000,000 works that were were used by Anthropic. The numbers, the dollar signs could start to add up very, very quickly if this class action lawsuit went forward and a jury found in favor of this class of authors Mhmm. Against Anthropic. It could be existential.

这就引出了上周宣布的拟议和解方案。

This brings us to the proposed settlement that was announced last week.

拟议和解的头条数字是15亿美元，这是Anthropic将向集体支付的金额，用以补偿被盗用的盗版材料。

So the headline number of the proposed settlement is $1,500,000,000, which is how much Anthropic would pay to the class to compensate them for the stolen materials, for the pirated materials.

作为和解的一部分，受影响的书籍数量从700多万本缩减到大约50万本。

As part of the settlement, the number of impacted books got whittled down from over 7,000,000 to around 500,000.

从最初700万降到50万，是因为有些作品最终被购买了，原始下载列表里有很多重复，还有一些书其实没有在版权局注册。所以我们大约得到50万本。目前还不知道具体包括哪些书。

And you get to that 500,000 from that initial 7,000,000 because some works were ultimately purchased. There's some duplication, a lot of duplication in the initial list of what was downloaded. Some books were not actually registered with the Copyright Office the way one would assume they are right away. So we get to about roughly 500,000. We do not know what books, what works are included in that list yet.

我们也不知道具体怎么分配。作者、出版商，谁拿多少钱。

Mhmm. We do not know exactly how the payouts will come. Right? Authors, publishers, who gets what portion of what money.

我们估计每位作者能拿到多少钱？

How much money do we think will go to each author?

大约每部作品3000美元。但再次强调，我们不清楚这钱是全部给作者，还是作者和出版商分，具体怎么分。

So it's roughly $3,000 per work. But again, we don't know if that's to the author, a portion to the author, a portion to a publisher Mhmm. How that gets worked out.

15亿美元，对Anthropic来说算多吗？

And and 1 and a half billion dollars, is that a lot of money to Anthropic?

Anthropic的估值是1830亿美元。所以这不会让它破产，这也是他们想和解的原因之一。嗯。但这笔钱也不是小数目，只是不会让它元气大伤。

Anthropic's valuation is a 183,000,000,000. So it's, you know, not gonna bankrupt the company, which is, you know, part of why they wanted to settle. Mhmm. But it's not nothing, but it is not going to debilitate them.

虽然作者和Anthropic已经同意了和解条款，但还有一道关卡。法官必须批准。而他表达了一些怀疑。

While the authors and Anthropic have agreed to the terms of the settlement, there's still one more hurdle. The judge has to approve it. And he's expressed some skepticism.

所以周末法官下了一道命令，说我们要开个听证会来讨论拟议的和解，但信里的语气很“我不是生气，我是失望”的那种感觉。

So over the weekend, the judge issued an order saying, we're gonna have a hearing to discuss the proposed settlement, but it was very, like, I'm not angry. I'm disappointed vibes from the letter.

我不是生你的气，我是对你的行为感到生气。

I'm not mad at you. I'm mad at your behavior.

差不多就是这样。法官用了“失望”这个词。他说，这里还有很多问题没有答案。我不完全相信你们已经把细节都敲定到能让我批准和解的程度。

Like Pretty much. And it used the word disappointed. Right? The the judge said, like, there are still a lot of unanswered questions here. And I am not totally convinced that you have enough of it worked out yet for me to approve the settlement.

明白了。所以不是一定要更多或更少钱，只是细节还不够。

I see. So not necessarily it should be more money or less money, just more like there's not really enough detail here yet.

对。他说，我们还没有作品清单，也没有作者清单。你怎么确保每个人都知道自己有资格获得可能的赔偿？赔偿要怎么发放？

Right. So he says, we don't know what the list of works is yet. We don't know what the list of authors is yet. How do you make sure everyone knows that they're eligible for a potential payout? How are you going to do the payouts?

所以他们本周早些时候在旧金山法院开了听证会，法官相当严厉，尤其是对原告律师。我觉得你们还没准备好，这方案还没完全成熟。他还担心会有更多诉讼，比如有人事后跳出来说，等等，我没被包括在内，我应该被包括的。

So they held a hearing earlier this week in San Francisco in court, and the judge was pretty harsh in laying out, especially to the plaintiff's lawyers. I don't think you have it all there yet. It's not fully baked. And he also said that he was concerned that there would be some more lawsuits, like people coming out of the woodwork after this, right, saying, wait, I wasn't included. I should have been included.

“我没有被公平通知，我应该被通知。”和解的目的是结束诉讼，而不是引发一堆新诉讼。他设定了接下来几周的一系列截止日期。比如，你必须在某某日期前提交涉及的作品清单。

I wasn't notified fairly. I should have been notified. The goal of a settlement is to end litigation, not launch a whole bunch of new stuff. He set a schedule of deadlines over the next few weeks. You know, you have to have a list of the works that are involved by x date.

你必须在X日期前拿到班级名单。他们将在几周后的十月初再次开会，届时他大概会批准或否决拟议的和解方案。

You have to have the class list by x date. They're gonna convene again in a few weeks in in early October, and he will presumably then approve or reject the proposed settlement.

那么，这对目前法院里正在审理的所有其他创作者与AI公司之间的案件可能意味着什么？

So what might this mean for all those other cases that are playing out in court right now against between creators and and AI companies?

它会产生影响。但再说一遍，它并不是法律先例，所以不会立刻驳回其他人的案子之类的。但很多其他案件仍在推进，包括上周又提起了新案子，就在Anthropic和作者们宣布拟议和解的那天，又有一些作者对苹果提起了诉讼，指控苹果如何摄取书籍用于训练模型。

It will have an impact. It again, it's not kind of legal precedent, so it won't immediately dismiss everyone else's cases or anything like that. But the other cases are still moving forward in a lot of instances, including there was a new case filed last week, right, the day that Anthropic and the authors announced their proposed settlement. There was another lawsuit filed by some authors against Apple about how they ingested books for their training models

嗯。为了

Mhmm. For

AI。所以接下来还会有大量诉讼，而且不只是书籍。对吧？我们说的是电影。我们说的是歌曲。

AI. So there's still a lot of litigation action coming down the pike, and not just in books. Right? We're talking movies. We're talking songs.

我们说的是各种创意内容。

We're talking all sorts of creative content.

感觉无论是社会还是法律体系，我们都在黑暗中摸索，试图弄清楚并找到这个重大问题的答案。

It feels like as a society and as a legal system, we're we're we're feeling our way through the dark to try to figure out and find an answer to this really big question.

是的。而且我认为我们开始得到一些答案了，对吧，AI公司可以使用这些内容，这属于合理使用。但关于补偿模式应该是什么样子、可以是什么样子，以及它将如何因媒介而异、如何因AI公司而异，仍有很多未知。所以你既在诉讼中看到这一点。

Yes. And I think we're starting to get a few answers, right, that that the AI companies can use this content. It is fair use. But there's still a lot of unknowns about what a compensation model should look like or could look like and how that's gonna differ by medium, how that's gonna differ by AI company. And so you're seeing this in litigation.

也在AI公司与媒体公司之间的谈判中看到这一点，讨论他们如何在前端就解决问题，以免最终在后端陷入诉讼。

You're also seeing this in discussions between AI companies and media companies about what can they do on the front end so that they don't end up in litigation on the back end.

有没有人问过Anthropic的AI机器Claude，它认为应该如何补偿作者和创作者？

Has anybody asked Claude, Anthropics AI machine? What does it think the answer should be for how authors and creators should be compensated?

我没有，但我有种预感，我知道你今天稍后要做什么了。

I have not, but I have a feeling I know what you're gonna be doing later this morning.

当然。嘿，Claude，介意我问你个问题吗？

Certainly. Hey, Claude. Mind if I ask you a question?

当然不介意。尽管问，任何问题都可以。

Of course. Go ahead and ask me anything.

你觉得作者的作品被用来训练AI模型时，应该得到多少报酬？

How much do you think authors should be compensated for having their work used to train AI models?

这是一个非常复杂的问题，各方都有合理的观点。当作者的创意作品为AI系统带来价值时，他们理应获得公平的补偿，但具体如何操作却很有挑战性。你怎么看？

That's a really complex question with valid arguments on multiple sides. Authors deserve fair compensation when their creative work contributes value to AI systems, but the mechanics of how that would work are challenging to figure out. What's your take on it?

我在问你。

I'm asking you.

我认为作者应该得到某种形式的补偿，尤其是当他们的作品直接贡献给商业AI系统时。挑战在于建立一个公平的机制，既能认可他们的贡献，又不会让AI开发成本过高。

I think authors should receive some form of compensation, especially when their work directly contributes to commercial AI systems. The challenge is creating a fair system that recognizes their contribution without making AI development prohibitively expensive.

好的。谢谢你，Claude。

Okay. Thanks, Claude.

不客气。

You're welcome.

结束前，我想确认你听到了：The Journal 将在下个月举办我们首次现场演出，地点在纽约市。10月7日星期二，我们在格拉梅西剧院。票真的只剩几张了，几乎售罄。如果你想参加，请访问 bit.lythegournallive25 购票并了解更多信息。

Before we go, I wanna make sure you heard that The Journal is hosting our first ever live show in New York City next month. We'll be at the Gramercy Theater on Tuesday, October 7. There are literally only a handful of tickets left. We are almost sold out. So if you want to be there, head to bit.lythegournallive25 for tickets and more information.

你可以在我们的节目备注里找到链接。我们非常期待在那里见到你。今天是9月10日星期三，节目就到这里。顺便提一句，《华尔街日报》的母公司新闻集团与OpenAI有内容合作协议。新闻集团旗下的两家子公司也已起诉Perplexity。

You can find the link in our show notes. We'd love to see you there. That's all for today, Wednesday, September 10. A quick note, The Wall Street Journal's parent company, News Corp, has a content deal with OpenAI. Two of News Corp subsidiaries have also sued Perplexity.

《华尔街日报》是Spotify与《华尔街日报》联合制作的节目。本集补充报道由Jeffrey Trachtenberg和Angel Al Young完成。感谢收听，明天见。

The Journal is a co production of Spotify and The Wall Street Journal. Additional reporting in this episode by Jeffrey Trachtenberg and Angel Al Young. Thanks for listening. See you tomorrow.