公共天才的私人捕获
《万物上游》系列第一篇
2026年7月2日
1956年1月24日,美国电话电报公司(AT&T)是全球最大的私营企业。
其营收约占美国国内生产总值的2%。员工总数达74.6万人。旗下拥有贝尔实验室——这个传奇研究部门已产出晶体管、太阳能电池、信息论和射电天文学,同时正积极铺设首条跨大西洋电话电缆。在随后的几十年里,它还将贡献UNIX、现代蜂窝电话技术、CCD图像传感器、首颗有源通信卫星以及一系列其他科学里程碑。这一系列非凡的智力产出,为贝尔实验室的科学家们最终斩获五项图灵奖和十项诺贝尔奖铺平了道路。
从许多指标来看,作为受监管的垄断企业,AT&T的日子过得相当不错。
然而,到这一天结束时,AT&T已将其全部7820项未过期专利的独家权利,以免版税方式授予任何提出请求的美国企业。AT&T还将以"合理费率"许可其未来申请的任何专利。一个处于前沿的、知识产权宝库就这样突然且不可逆转地向自由市场敞开了大门。
反垄断官员最初将这一和解方案标榜为一场胜利。司法部称其为重大胜利,一位司法部律师更将其誉为"奇迹"。尽管AT&T作为受监管的垄断企业已存在数十年,其回报率被限制在(按当今标准)相对保守的每年约7%,但政府监管机构仍追求并确立了一系列具有里程碑意义的额外限制,以削弱AT&T的垄断力量。
然而,公众情绪很快开始转变。《商业周刊》称该法令“不过是轻轻拍了一下手腕”。后来,美国众议院的一个小组委员会因其对AT&T独家供应链和垂直整合的宽大处理,将其视为“反垄断法执法史上的一个污点”。无论是通过费率合同补贴AT&T庞大研发预算的费率支付者,还是联邦政府中的许多人,都认为这种前所未有的经济集中对共和国而言仍然过于危险,不能任其继续发展而不加遏制。
如今臭名昭著的1956年专利法令,只是AT&T与联邦政府之间经过七年谈判达成的和解协议的一半内容。AT&T希望通过其子公司西部电气继续制造电话设备,但监管机构认为这种垂直整合正在扼杀行业内的竞争。联邦政府本身在这个问题上也矛盾重重,以至于艾森豪威尔总统时期的国防部长查尔斯·威尔逊恳请诉讼律师,称将AT&T与西部电气剥离“违背了我们国家的切身利益”。
和解协议的另一半内容,是禁止贝尔公司从事电信以外的任何业务。
后来对历史记录的分析显示,贝尔公司69%的专利与电信几乎无关。相反,这些专利涵盖了从化学到半导体、金属加工、照明、光学等多个领域。
和解协议的两部分内容共同确保了这一丰富的知识宝库——约占当时美国所有未到期专利的1.3%——几乎在一夜之间免费开放,并且有山姆大叔的保证,那个庞大而可怕的贝尔实验室法律之狼不会找上门来。
短短几年内,这些被释放的专利就在电信行业之外产生了近60亿美元的后续专利价值。其中约35亿美元的价值来自年轻初创公司提交的专利。这场初创企业爆发浪潮中一个著名的分支,贯穿了肖克利半导体、仙童半导体,并最终进入了传奇公司英特尔。
英特尔联合创始人戈登·摩尔(即提出摩尔定律的那位)后来将这种由同意法令驱动的创新连锁反应描述为:
“对商用半导体行业而言,这是最重要的发展之一。(它)让商用半导体行业真正在美国起步。贝尔实验室宽松的授权政策,与戈登·蒂尔离开贝尔实验室创办德州仪器、威廉·肖克利同样离开并在帕洛阿尔托创办肖克利半导体之间,存在着直接联系。这开启了硅谷的成长。”
沉淀
一代由公共资金资助的杰出科学家,打造了人类历史上最具影响力的技术天才集群之一。贝尔实验室产出专利、发明产品,并在数十年间成为美国前沿科学无可争议的中心。但它是如何做到的?
想象一片精心修筑的水稻田,由精益求精的农民花费多年时间精确打造出肥沃的环境。它看起来不过是一片水田,但事实上,水稻是少数几种能耐受水淹根系的主要作物之一。由于大多数杂草同样无法耐受水淹,水便起到了除草的作用。刻意淹水还切断了有机分解所需的氧气,因此土壤能保留更多养分,而不会像干燥、透气的田地那样将养分消耗殆尽。温暖、浸水的泥浆还同时成为固氮微生物的绝佳栖息地。一块精心照料的水田,季复一季、有时甚至长达数百年,基本能自行施肥。这个不起眼的泥塘,实际上是人类设计过的最具生产力的种植系统之一。
AT&T 作为垄断企业的独特经济地位,为贝尔实验室那种刻意实验、耐心探索和延迟收获的文化创造了条件。贝尔实验室从庞大且稳定的全国性收入基础中汲取资源,无需在每个预算周期重新证明其合理性。美国监管机构通过 AT&T 的价格设定这一收入基础,其方法是基于 AT&T 在网络中投入的资本,采用固定百分比回报率进行计算。这里所说的投入资本,指的是交换机、线缆、建筑物等。
在普通公司,研发是一项需要尽量压缩的成本,但在AT&T并非如此。
贝尔实验室在研发上投入的每一美元都同时发挥了两重作用。首先也是最重要的,这是一项无风险、可回收的成本,根据合同由美国电话用户补贴。其次,它是AT&T构建和部署新型资本密集型技术的源泉。这种资本支出扩大了计算其保证回报率的费率基础。在这些新技术上投入的资金越多,在同样受监管的约7%回报率下,获得的绝对利润就越大。
这种安排在几十年的时间里对各方都非常有利,但未必可以复制。我们是否应该尝试重现它,这一点也并不明确,因为它也带来了真实的代价。低效的过度投资、缺乏价格约束,以及最重要的一点——将发明成果囤积在垄断壁垒之内的动机——都损害了用户的利益。但在20世纪的大部分时间里,这些有保证的利润确实客观地创造了一片广阔的沃土,让一项又一项技术创新得以在其中蓬勃发展。
前沿科学如今的面貌已经不同。它植根于模型权重和GPU。它充斥着模型 token 消耗和智能体循环。它在数据中心里绽放。
尽管AI辅助研究作为一个领域仍处于早期阶段,但使用统计数据表明,在各大AI实验室内部及周边,一些重大变化正在发生。严肃的研究人员正在利用这项新技术解决真实的问题,有时甚至是解决整类此前无法解决的问题。蛋白质结构、研究数学、材料设计、药物发现和复杂系统分析,只是AI模型切实提升研究人员能力、扫清人类科学障碍的少数几个领域。但这片沃土从何而来?
这其实算不上什么秘密。
OpenAI表示,它“主要依赖公开可用的信息来教导[其]模型如何提供帮助”。Anthropic曾试图构建一个“包含‘世界上所有书籍’的中央图书馆”来训练其模型。Sam Altman本人进一步阐述,他们的前沿模型是在“人类的集体经验、知识和智慧”上进行训练的。
剥去委婉的修辞,剩下的便是赤裸裸的现实:这些前所未有的能力,是由全球每一个曾写下过文字的人的自我表达汇聚而成的。
而基于这一现实构建的产品,根据前沿实验室自身的收入、预测和使用数据,是一代人以来最有价值的东西。
Anthropic 的年化收入运行率从 2024 年 1 月的 8700 万美元飙升至年底的 10 亿美元,到 2025 年大约增长了 10 倍,并在 2026 年 5 月达到了 470 亿美元。这使其成为历史上复合增长最快的企业软件公司。OpenAI 也相差不远。据估计,如今 80% 的美国劳动力所从事的工作中,有一部分会接触到这些模型。所有这些影响,都得益于在数十亿人一生时间量级的数据语料库上进行的、长达数周的训练运行。
这是对公共智慧的私人捕获。
一个前沿模型,就是将海量训练数据压缩成数值权重。书籍、论坛、代码仓库、手册、论文、聊天记录、转录文本、法庭案例、散文、评论区、文章、教程,以及前沿实验室的爬虫大军在互联网内外所能抓取到的每一个零散想法,其集合之庞大令人震惊。
从某种意义上说,它的不可理解性几乎就像一层精神铠甲。它太过庞大,无法直接理解。
想象一个狂野的河流三角洲。当水流从高地奔向大海时,它会侵蚀流经的土地,并将碎屑作为沉积物带到下游。淤泥、沙子、黏土,以及各种有机物,从每一条支流和河岸,从犁过的田野到崎岖的山坡,都被冲刷而来,最终汇聚在三角洲。河流沿途滋养的每一个生命的丰饶也汇聚于此。一个大陆的流域,旋转、积累,最终在其终点沉淀。海量而各异的物质在三角洲中结合,形成了一片丰饶、奇特且充满生机之地。
而所有人类知识的总和,若非如此,又是什么呢?
从历史书页中刮取的每一簇字母(即 AI 模型摄入的原始 token),都如同人类探索这条永不停歇的河流所沉积的一粒泥沙。当泥沙堆积得足够多,你便能理解星辰的运行轨迹。当你长久凝视这片淤泥,你便能看见逻辑本身的结构。大语言模型将冲积土壤转化为答案的质变过程,正是滋养它的社会所收获的盛大果实。
但若除去泥土,便没有三角洲。
若除去语料库,便没有收获。
只剩虚无。
模型并非在真空中学会推理。它通过一遍又一遍地观察理性,从而吸收了理性。它的泛化能力,源自它所吸纳的每一个示例、每一次修正和每一场争论。在历史、文化与科学的回响中,某处的人类决策为今天的聊天机器人回复搭建了舞台。这种被培育出的智能,从人类意义构建的沉积物中生长出来,但这里没有一粒沉积物不是由某个人所留下的。
这些“某个人”中,有许多已经逝去。他们撰写了古代典籍,验证了基础科学,并为了我们这些仍在世者的利益,记录了从古代至今的世界历史。但同样,这些“某个人”中,也有许多依然健在。他们正在编写模型所输出的工作代码。他们正在将基础科学推向新的前沿。他们正在组织、调查、行动,并对源源不断的时事做出反应。任何与当下相关的回答,都是从某个人那里借来的。
事实上,你就是那些“某个人”之一。字面意义上的。
你凌晨两点的垃圾帖。你对陌生人文章那篇雄辩的回复。你留下的尖刻餐厅评论。你的标题、评论、内部笑话,以及你所有的公开对话。你曾对那无限分叉的数字交流之流所做的每一次贡献,无论大小,都已沉积在三角洲的某处。
互联网属于每一个人。
尼罗河三角洲滋养了埃及五千年。湄公河与恒河流域至今仍养育着数亿人口。所有文明摇篮的形成,都完全或部分归功于大河的泛滥平原与三角洲,这绝非巧合。这些地区仅凭其原生物质的天然丰饶,便支撑人类度过了最原始的时期。这片土壤渴望着迸发生机,然而地球上最肥沃的农田,几乎无一例外,都是偶然形成的。
互联网亦是如此。
我们这些信息高速公路上的无数数字居民,最初并非为了创造训练语料库。我们为自己、为彼此而写作。我们开玩笑、争论、教学、抱怨、调情、调试代码,就这样汇聚成了如今被私人资本收割的、相互关联的原始素材。经济学领域(它也存在于语料库中)对此有专门的术语。
要对任何资源进行分类,经济学家会问两个问题:它是否具有排他性?它是否具有竞争性?
更直白地说,你能阻止别人使用它吗?一个人使用它,是否会减少留给其他人的份额?
这其中存在例外和子类别,但这个简单的测试为我们提供了一幅地图。
如果一种商品具有排他性和竞争性,它就是私人物品。想想一个三明治。如果我吃了它,它就没了,而法律保护我免受三明治小偷的侵害。
如果一种商品具有排他性,但基本不具有竞争性,它就是俱乐部物品。Netflix 订阅就是一种俱乐部物品。如果我看了一部电影,你仍然可以看,但前提是我们都付费获得访问权限。
如果一种商品难以阻止人们使用,且具有竞争性,它就是公共池塘资源。牧场是经典例子。许多农民都可以使用牧场,一头牛吃草不会毁掉草场,但牛的数量足够多时,它们最终会把草啃到只剩泥土。这就是著名的“公地悲剧”问题。
最后,如果一种商品难以阻止人们使用,且不具有竞争性,它就是公共物品。路灯是公共物品。一旦街道被照亮,我们所有人都可以在灯光下行走,而我这样做并不会让你的道路变暗。
私人物品和俱乐部物品通常由逐利行为者及其所处的法律体系来管理。公共物品主要由政府或无人管理,而公共池塘资源则往往存在于一个模糊地带:每个人都想从中获益,却没人愿意承担维护成本。
前沿实验室通常主张,互联网上的数据在合理使用版权制度下可公开用于训练。从经济学角度看,这一论点意味着互联网是一种公共物品。大规模抓取、摄取和使用互联网数据进行训练,并不会破坏原始数据。每一篇博客文章、每一条推文、每一场网络骂战确实都还在,并且大多可以访问。没有人明确拥有它们。
你发布内容的平台拥有你的帖子吗?你是否与平台共享所有权?这种关系会随时间改变吗?
毕竟,你确实是免费发布到网上的。
但授予访问权限并不等同于授予使用许可。一张借书证让你有权阅读一本书,而不是影印整座图书馆。购买国家公园门票并不赋予你伐木权。进入一家营业中的商店并不等于你可以偷窃其库存。公众对互联网上作品的访问,并不会自动赋予使用权利。
而“你发布了,就代表你接受了”这一说法,还存在第二个更深层的问题。直到最近,大语言模型训练数据这一使用场景根本不存在,在线发布内容的个人也无法合理预见。2008年的一位博主不可能同意自己的作品被用于训练今天的语言模型,因为这在当时是无法想象的。同意不能回溯性地赋予,尤其当这是一个从科幻小说情节变成现实的事情。
当前大语言模型的法律战场,是一个悬而未决的故事。
值得注意的是,尽管我们出于道德直觉,但获取途径和同意与否,与前沿实验室在“合理使用”这一主要法律辩护主张中并不相关。相反,法院在裁决合理使用抗辩时会评估四个标准。他们会考察使用版权材料的目的、作品的性质、使用的数量以及对原作品市场的影响。在实践中,这四个标准通常归结为两个重要问题。
新作品是否具有变革性,以及它是否损害了原作品的市场?
2025年6月,法官阿尔苏普在巴茨诉Anthropic案中裁定,基于合法获取的书籍进行训练“本质上是变革性的”,但使用盗版书籍构建其资料库则“本质上、无可挽回地构成侵权”。凭借这场喜忧参半的胜利,Anthropic面临理论上高达700亿美元的版权损害赔偿风险,并在几个月后迅速以15亿美元和解了此案。这是美国历史上(迄今为止)金额最大的版权和解案,该和解既未授予Anthropic任何未来许可,也未澄清任何后续适用的法律。
在另一项相关裁决,即卡德雷诉Meta案中,法官查布里亚认为大语言模型训练同样具有变革性,并勉强裁定市场损害的证据不足。在裁决中,他批评原告几乎没有提出任何市场稀释的证据,并指出,大语言模型用与训练数据相似的AI作品充斥市场的能力,“在此类案件中,往往会使原告决定性地赢得第四个要素——从而赢得整个合理使用问题的胜利。”
使讨论更加复杂的是,美国版权局在2025年发布了一份不具约束力的报告,结论是公开可用性本身并不自动允许基于合理使用的模型训练。截至本文撰写时,尚无既定的法律标准来衡量大语言模型导致的市场稀释,但这注定将成为未来法律裁决中的主要冲突点。已有数十起诉讼和政策争论正在考验前沿实验室不断演变的训练数据辩护策略。
这些实验室最具诱惑力的辩护也最为简单。
“它只是在阅读”——这是科技界为AI模型训练辩护时常说的一句话,也确实是个很有说服力的论点。每一位在世的作家,都是由他们读过的书塑造而成的。没有人因为受到《老人与海》的启发,就给海明威的遗产基金会寄去一张支票。如果模型只是另一个读者,那么它欠下的,和每个读者欠下的一样:什么都不欠。
一个人一生读一万本书,会成为又一位作家,以人类的速度创作,以人类的规模出版,并将自己的沉淀物一粒一粒地归还到三角洲中。而一个读遍一切的模型,却变成了一台印刷机,这台印刷机还能印刷出更多的印刷机。它以工业化的规模输出作品,训练自己的后继者,并且只需按下一个按钮,就能与它所“阅读”过的那些作家直接竞争。灵感从未稀释过市场,但印刷机做到了。
大约在1440年,古腾堡印刷机的发明最终导致了1710年《安妮法令》的通过。一个由强大书商组成的卡特尔游说英国议会,试图恢复他们对图书贸易的垄断权,而议会却将权利赋予了作者,使其成为合法的所有者。既得利益者寻求保护,而公众的代表却将所有权交到了创作者手中。
这项法令奠定了现代版权法的基础。在印刷机出现之前,这并非必要,因为大规模的盗版实际上是不可能的。新的技术格局引发了一场复制浪潮,压垮了为上一个时代的问题而设计的法律体系,但这场清算花了两个半多世纪才得以展开。
那台能印刷出更多印刷机的印刷机,不会让我们等那么久。
污染三角洲
从表面上看,那个蕴藏着人类集体知识沉积物的富饶河流三角洲,确实像是一种公共物品。前沿实验室抓取并吞噬互联网上庞大的沉积体,在字面意义上并不会破坏原始材料。法院已经开始朝着这个方向做出裁决,但就像许多法律裁决一样,这在狭义上是正确的,却完全忽略了问题的关键。
对训练语料库的浅显理解,误解了互联网功能层及其参与者实际互动的方式。到目前为止,我们只分析了文本层。前沿实验室抓取到训练语料库中的网页、文章、帖子、评论以及其他一切内容,是显而易见的一部分,但互联网还有许多其他层面,它们为文本层的存在创造了条件。
除了协议层或接入层这些显而易见的技术层面,我们还必须考虑互联网的发现层、注意力层、贡献层和诚信层,以及它们之间的行为流动。互联网的持续效用,取决于人们能否找到、参与、贡献,并最终相信他们在网上所访问内容的价值。
当被视作一个静态语料库时,AI 训练对互联网造成的损害并不明显。当然,这种损害与牧场里牛太多会破坏草地的损害方式不同。
相反,真正受到损害的是最初演化出来用以丰富该语料库的复杂系统。互联网是一个由相互关联的公共物品、俱乐部物品和公共池物品构成的堆栈。不同的层面相互反应、相互强化,使得整体既有价值,又容易受到大语言模型带来的特定危害的影响。
没有任何人类创作者能与生成式输出那试图淹没我们算法和注意力的原始体量相抗衡。互联网的各个层面在这里表现得不像牧场,而更像一条道路或一个电子邮件收件箱。它们在达到某个阈值之前是非竞争性的,但超过阈值后,就会变成灾难性的竞争性资源。
因此,认真参与网络的动力,随着每一个由推文衍生物再经 AI 变体生成的内容而逐渐消减。如果你发布的内容无人问津,无法与一万个 AI 生成的、随着欢快电子乐起舞的狗狗视频竞争,而当你终于做出真正令人印象深刻的作品时,热评却指责你是 AI 生成的,那为什么还要在网上创作和分享呢?生成式 AI 工具能够以近乎零的边际成本,用媒体内容(无论好坏)淹没网络的每一个角落,这或许正是那辆早在 90 年代就已驶出车站的“永远向所有人发送垃圾邮件”大巴,最终踩死油门的一脚。
这是一个重要的时刻。对这片丰饶的语料层进行天真的收割,无异于向那些使其成为可能的人们开炮。网络的某些部分可能已经崩溃。如果我们处理不当,整个互联网可能会遭受不可挽回的破坏。然而,我们并非毫无应对之策。
我们早已知道如何保护公共资源。埃莉诺·奥斯特罗姆在 2009 年因记录瑞士高山牧场、日本森林和西班牙灌溉网络如何可持续地共享其公共资源长达数百年而获得诺贝尔奖。她总结了公共资源得以持久存在的八个条件:明确谁可以从中获取资源的边界;与当地条件相匹配的规则;受规则影响的人对这些规则有发言权;由对用户负责的机构进行监督;对过度使用行为实施渐进式惩罚;有可及的争端解决途径;承认社区自我组织的权利;以及跨层级的治理结构。
用这份清单来审视互联网,几乎没有一个条件成立。它的边界模糊不清,由所有人共同滋养,却无人设界。那些填充内容的人对网络如何治理没有发言权。规则不明确,且仅偶尔通过诉讼来执行,而诉讼也仅由少数资金雄厚的参与者发起。监督机制即使存在也极为薄弱,总是事后追溯,从未主动预防。没有监控,没有渐进式惩罚,也没有解决争端的共享平台。按照奥斯特罗姆的精确定义,这根本不是一个被治理的公共资源。它是一个被拔掉插头的公共池资源。
这就是我们陷入当前困境的原因。就像当年凯霍加河油污横流、布法罗溪被矿渣淹没的时代一样,我们正面临一场网络工业废料灾难,它即将污染整个三角洲。
归属权崩塌
这条河流目前仍在流淌。新的沉积物持续在三角洲堆积,语料层不断增厚,实验室也在持续抓取数据。
而在抓取过程中,他们不断将互联网这片浩瀚的三角洲压缩成固定的权重集合,但这种持续进行的仪式又引发了另一个问题。这次是价值问题,而非质量问题。
具体来说,就是谁该为哪种价值获得报酬的问题。
按照前沿实验室的说法,这数十亿被抓取的数据点,每一个都毫无价值,但合在一起却价值数万亿美元。
所谓“毫无价值”,是指训练集中并不需要任何单独被抓取的作品。移除任意一个数据点,模型几乎察觉不到。因此,没有任何单一作品是真正重要的。所以,没有任何单一作品需要被付费。
但如果我们能收起惊掉的下巴,好好咀嚼一下他们塞给我们的这套说辞,就会发现单个数据点显然并非毫无价值。
这是一种修辞诡计。是自封的侦探在说“既然我们无法确切查明有多少珠宝被盗,那就不能提出指控”式的双重话术。只不过,这个侦探同时也是小偷。而且刚刚开了一家珠宝店。
这种“哎呀真没办法”的说法,当然荒谬至极。糟糕的会计做法并不能抹消明确的价值转移,尤其是在这种会计根本不可能做到的情况下。
从法律角度来看,这种会计在原则上是不可能做到的,因为版权法是为监管离散复制行为而制定的。法院判例假设侵权者复制的是来自可识别权利人的可计数作品。而大语言模型训练是对数十亿作品的一次性统计吸收。这是一种形式不同但道德本质相同的行为,但由于这种提取行为是一种新机制,法律工具暂时还无法完全约束它。
但更令人担忧的是,试图进行量化核算可能在数学上本身就是不成立的。
评估训练输入的主流形式化方法是 Shapley 值,它计算某个输入在所有可能出现的排列顺序中的边际贡献平均值。但这个数值并非作品本身的属性,而是该作品与训练集中所有其他作品之间关系的函数。同一份文档放在不同的训练集中,其 Shapley 值也会不同。即便用相同模型训练两次,由于训练过程具有随机性,每次运行的 Shapley 值也可能发生变化。研究人员甚至尚未就 Shapley 值是否为正确的贡献度衡量指标达成共识,而且计算前沿规模模型的真实 Shapley 值在计算上不可行——这些模型单次训练就需要数周时间,而精确的 Shapley 核算则需要针对不可能实现的输入组合进行重新训练。因此,截至目前,并不存在任何客观的估值方案能够计算出某部作品的具体贡献份额,即便有方案被提出,一旦付诸实施也必然会在诉讼中被推翻。
在可预见的未来,对前沿规模的大语言模型训练进行个体归因是行不通的。你无法根据贡献比例向人们支付报酬,因为根本不存在可操作的具体份额。这正是错误认知的根源。实验室将这一事实解读为:“既然无法归因,那我们什么都不欠。”
我认为这只能说明你无法按比例支付,但该付的报酬依然要付。
为公共智慧付费
问题的真正面貌是这样的:个体创作出独特的作品,但从来不是孤立完成的。互联网本质上具有公共性。合作与冲突共同滋养着整体。
枝条在修剪后长得更壮,最肥沃的土壤来自腐烂物,藤蔓通过接触攀爬。友善的头脑相互交叉授粉,而捕食者与猎物则促使彼此跑得更快。这堆混杂着垃圾与杰作的庞杂内容之所以有价值,恰恰是因为每部分与其他部分之间存在着多样的关系。在它们之间鲜活的间距中,人类认知萌芽、绽放并结出果实。尽管孤胆天才的浪漫传说令人神往,但互联网终究是一项团队协作的成果。
既然你无法找出最有价值的球员,那就付钱给整个团队。
这笔费用是一种版税。它是对大语言模型从人类最优秀与最平庸之处所汲取的公共智慧,进行精确到美分的量化核算。
不妨称之为“语料版税”。
前沿实验室将其总收入的一个固定比例缴入公共基金。该基金每年向每位符合条件的美国公民发放相同金额。任何比这更复杂的机制,都会重新引入衡量难题,并在法庭上因无数细枝末节的争论而夭折。
随着大语言模型融入互联网并重塑人类表达的激励结构,版税成为补偿集体性、不可归因贡献的唯一连贯答案。前沿实验室不能继续无限制地收割互联网。它们必须补充那些滋养其模型所依赖的肥沃三角洲的上游源头,否则互联网将在十年内变得面目全非。
我们此前曾构建过这种机制的较小版本。
当私营实体从共享资源中获利时,我们认识到公众有权分享收益。阿拉斯加永久基金正是遵循这一直觉。每位符合条件的居民都能获得一份资源财富的份额,而这份财富是任何个人都无法单独主张的。既然我们无法可靠地衡量任何一个人的话语对模型的价值,那么分配方式就应反映这种归因的失败,而非假装能够解决它。
当个体主张数量过多而无法逐一定价时,我们不会假装它们毫无价值;当私营实体损害公共空间时,我们认定它们应对此破坏负责。在河流燃烧了一个世纪之后,美国国会在1980年建立了“超级基金”制度,将清理账单交给污染者——即使这些污染行为在当时是合法的。没有人需要追踪是哪一桶毒物污染了哪一口井。从有毒废物中获利的行业必须出资修复土地。
有人会说,贝尔案的先例主张开放权重,而非支付费用,但补救措施应针对伤害。在贝尔案中,竞争对手因知识产权封锁而受损,因此补救措施是开放访问。而今天,贡献者因知识价值的被抽取而受损。因此,支付费用是使受损方恢复原状的补救措施。
语料库版税起初可能数额很小,也许只够每年多买一箱啤酒。但金额本身,不如它所赋予的地位重要。它向公众表明,他们不仅仅是用来训练日益庞大的大语言模型的原材料。如果那些实验室对自己正在构建的东西判断正确,那么啤酒钱会变成买菜钱,再变成房租钱,并且随着实验室及其收入的增长而增长。如果实验室判断错误,那也不会是因为版税扼杀了商业模式。普通人的生活、争论、疑问、玩笑、纠正和创作,都在帮助维持着那些模型所消耗的语料库,而版税则是一种主动的、而非寄生式的参与方式。语料库要么是必不可少的,要么就不是。如果它是必不可少的,那它就有价格,而各行各业每天都在为必要的投入付费。实验室在与 Reddit、新闻集团和美联社的授权协议中,已经承认了这一点。
这不是福利,因为福利的前提是企业在补贴公众,而实际上补贴的方向恰恰相反。这不是慈善,因为慈善意味着没有收到任何回报。这不是税收,因为税收将盈余视为公司财产,并受制于公共主张。
这是归还。
描述这类问题的法律术语是不当得利,但普通法中的这个概念,对于摆在我们面前的这件事来说,格局太小了。在普通法中,不当得利要判断一方是否在另一方受损的情况下获益,且这种获益的全部保留是否显失公平。
前沿实验室已经获得了这种利益:它们将未经补偿、大规模聚合、由公众生成的语料库,转化为私有的基础设施级价值,同时却威胁着该语料库得以更新的条件。它们这样做的规模和扩散程度,使得个人诉讼无法合理地为其定价。问题的规模告诉我们,补救措施必须是集体性的。
对根植于公共语料库的价值所支付的版税,是一种回报。流回互联网的东西,是应得的,而非馈赠。这是对公众智慧的版税。
版税只是解决方案的一部分。它是一个更大的贡献与维系体系中的一个环节。它不会取代版权主张或私人许可合同。在所有权清晰可辨的地方,这些权利主张和合同仍然可以且应该存在。版税解决的是那些无法组织、谈判或通过诉讼进入许可市场的、无法归因的创意长尾问题。
各大实验室正在做的事情,在性质上并非前所未有,只是在规模上达到了空前程度。这是在文明尺度上,对公共智慧的私人攫取。我们对此不应感到意外。建立在公共支持之上的企业实体,往往试图将收益私有化,同时将使其成为可能的条件社会化。特殊组织此前也曾滥用其特殊地位。不同之处在于,这一次,受影响的是所有人。
也许,自从五千多年前我们第一次在泥板上刻下印记以来,我们就在朝着这个方向前进。我们通过记录、收集和分类,将自己推向了脆弱的境地。一旦每一个外化的思想、每一个写下的文字、每一张图表、每一处修饰、每一个措辞,都能被注入一个机械精灵之中,谁又能抗拒将其卖回给那些提供了这一切的人们的诱惑呢?
这就是公众需要提出主张的原因。前沿实验室日益追求公用事业的特权和权力,却不愿承担随之而来的公共义务。语料库版税通过让公众从他们使之成为可能的财富中分得一杯羹,在一定程度上恢复了这种交易中的平衡。公众已经在承受这些模型所创造的世界带来的负面影响,理应分享一部分正面收益。
语料库版税确保了人类散落的智慧火花,在他们帮助创造的这场游戏中拥有切身利益。

The Private Capture of Public Genius
Essay I of Upstream of Everything
Jul 02, 2026
On January 24, 1956, the American Telephone and Telegraph Company was the largest private company in the world.
Its revenues amounted to almost 2% of the U.S. gross domestic product. It employed 746,000 people. It owned Bell Labs, the fabled research division that had already produced the transistor, the solar cell, information theory, and radio astronomy, while also actively laying the first transatlantic telephone cable. In the following decades, it would add UNIX, modern cellular telephony, the CCD image sensor, the first active communications satellite, and a long list of other scientific milestones. This singular stretch of intellectual output paved the way for Bell scientists to eventually collect five Turing Awards and ten Nobel Prizes.
By many metrics, life as a regulated monopoly was very good for AT&T.
Yet by the end of the day AT&T had signed away exclusive rights to every single one of its 7,820 unexpired patents, royalty-free, to any American firm that asked. AT&T would also license any future patents it filed at “reasonable rates.” A bleeding-edge, intellectual property treasure hoard was suddenly and irrevocably opened to the free market.
Antitrust officials initially sold the settlement as a triumph. The Justice Department called it a major victory, with one DOJ lawyer hailing it as “miraculous.” Despite AT&T already existing for decades as a regulated monopoly, with its returns constrained to a relatively conservative (by today’s standards) ~7% per annum, government regulators had pursued and established a landmark set of additional restrictions to curtail AT&T’s monopoly power.
Soon, however, public sentiment started to shift. Business Week called the decree “hardly more than a slap on the wrist.” A House congressional subcommittee would later deem it “a blot on the enforcement history of antitrust laws” for its perceived lenience on AT&T’s exclusive supply chains and vertical integration. Both the ratepayers, who subsidized AT&T’s vast research budget through its rate contracts, and many in the federal government believed this unprecedented economic concentration to still be far too dangerous for the Republic to continue unabated.
The now-infamous 1956 patent decree was just one half of a settlement negotiated over seven years between AT&T and the federal government. AT&T wanted to continue manufacturing telephone equipment through its subsidiary Western Electric, but regulators believed the vertical integration was foreclosing competition within the industry. The federal government itself was so conflicted about this issue that Secretary of Defense under President Eisenhower, Charles Wilson, pleaded with litigators that severing AT&T from Western Electric was “contrary to the vital interests of our nation.”
The second half of the settlement barred Bell from pursuing any business other than telecommunications.
A later analysis of the historical record revealed that 69% of Bell’s patents had little to do with telecom. Rather, they ranged from chemistry to semiconductors to metalworking, lighting, optics, and more.
The two halves of the settlement combined to ensure that this rich intellectual corpus, roughly 1.3% of all unexpired American patents at the time, became freely available essentially overnight and had a guarantee from Uncle Sam that the big, bad Bell Labs legal wolf would not come knocking.
Within just a few years, these released patents would generate almost $6B in follow-on patent value outside of the telecom industry. About $3.5B of that value came from patents filed by young, startup companies. One famous branch of that startup explosion ran through Shockley Semiconductor, then Fairchild Semiconductor, and eventually into the storied company known as Intel.
Intel’s co-founder, Gordon Moore (of Moore’s Law fame), would later describe this consent-decree-driven innovation cascade as:
“One of the most important developments for the commercial semiconductor industry. [It] allowed the merchant semiconductor industry to really get started in the United States. There is a direct connection between the liberal licensing policies of Bell Labs and people such as Gordon Teal leaving Bell Labs to start Texas Instruments and William Shockley doing the same thing to start Shockley Semiconductor in Palo Alto. This started the growth of Silicon Valley.”
Sediment
A generation of brilliant, publicly subsidized scientists built one of the most impactful clusters of technical genius the world has ever seen. Bell generated patents, invented products, and became the undisputed epicenter of American frontier science for decades. But how?
Imagine a carefully crafted rice paddy, terraced by exacting farmers who spent years precisely engineering a fertile environment. It looks like just a flooded field, but it turns out that rice is one of the few major crops that tolerates submerged roots. Since most weeds can’t tolerate submersion either, the water does the weeding. The deliberate flooding also cuts off the oxygen required for organic decomposition, so the soil retains more of its nutrients rather than burning them off like a dry, aerated field does. And the warm, waterlogged mud triples as an excellent habitat for nitrogen-fixing microbes. A well-tended paddy largely fertilizes itself, season after season, sometimes for centuries. This humble mud pond is actually one of the most productive growing systems humans ever designed.
AT&T’s unique economic position as a monopoly set the conditions for Bell Labs’ culture of deliberate experimentation, patient exploration, and delayed harvesting. Bell drew from an enormous and stable nationwide revenue base that didn’t have to be re-justified every budget cycle. American regulators set this revenue base through AT&T’s prices by using a fixed percentage return calculation on the capital it invested in the network. Here invested capital means switches, cables, buildings, and the like.
At a normal firm, research is a cost you minimize, but not at AT&T.
Every dollar spent on research at Bell Labs did two things at once. First and foremost, it was a no-risk, recoverable cost subsidized by U.S. telephone ratepayers under contract. Second, it was a wellspring of new, capital-intensive technology for AT&T to build and deploy. This capital expenditure expanded the very rate base on which its guaranteed return was calculated. The more money spent on these new technologies, the larger the absolute profit gained by the same regulated ~7% return.
This arrangement worked out very well for all parties for decades, but is not necessarily replicable. Nor is it obvious we should even try to recreate it because it came with real costs too. Inefficient over-investment, lack of price discipline, and most importantly an incentive to hoard inventions behind a monopoly wall all hurt ratepayers. But for much of the 20th century, these guaranteed profits did objectively create an expansive paddy field in which one technological innovation after another could flourish.
Frontier science looks different today. It's rooted in model weights and GPUs. It is flooded with token spend and agentic loops. It blooms in data centers.
While AI-assisted research is still young as a field, usage statistics show something big is happening in and around the major AI labs. Serious people are using this new technology to solve real problems, sometimes entire classes of problems, that were previously unsolvable. Protein structures, research mathematics, material design, drug discovery, and complex systems analysis are just a few of the fields where AI models are tangibly improving researchers’ abilities to clear humanity’s scientific roadblocks. But from where does this rich soil come?
It’s not really a secret.
OpenAI says it “primarily rel[ies] on publicly available information to teach [its] models how to be helpful.” Anthropic attempted to build a “central library of ‘all the books in the world’” to train its models. Sam Altman himself elaborates that their frontier models are trained on “the collective experience, knowledge [and] learnings of humanity.”
Strip the euphemisms and you’re left with the stark reality that these unprecedented capabilities were assembled out of the self-expression of every person across the globe who ever wrote anything down.
And the product built from this reality is, by the frontier labs’ own revenue, projections, and usage numbers, the most valuable thing built in a generation.
Anthropic’s annualized revenue run-rate rocketed from $87M in January 2024 to $1B by year-end, roughly 10x’d through 2025, and just hit $47B in May 2026. This makes it the fastest-compounding enterprise software company in history. OpenAI isn’t that far behind. An estimated 80% of the American workforce now holds a job where some portion of the work is exposed to these models. All of this impact was made possible by multi-week training runs over a data corpus measured in the lifetimes of billions.
This is the private capture of public genius.
A frontier model is the compression of a massive amount of training data into numerical weights. The combined collection of books, forums, code repositories, manuals, papers, chat logs, transcripts, court cases, essays, comment sections, articles, tutorials, and every errant thought scrapeable by the frontier labs’ army of spiders crawling across the internet and beyond is staggering.
In a way, its incomprehensibility is almost like psychic armor. It’s too big to understand directly.
Consider a wild river delta. As water runs from highlands to the sea, it erodes the land it travels through and carries the debris downstream as sediment. Silt, sand, clay, and all manner of organic material, scoured from every inch of tributary and riverbank, from plowed fields to rugged hillsides, end up aggregated in the delta. So does the richness of every life the river supports along the way. A continental watershed, swirling, accumulating, and ultimately settling at its terminus. The vast volume of disparate material combines in the delta to form something lush, strange, and alive.
And what is the sum of all human knowledge if not this?
Every cluster of letters scraped from the pages of history (the literal tokens an AI model ingests) is a single grain of silt deposited by the ever-flowing river of man’s exploration. Pile enough grains and you understand the movement of the stars. Stare long enough at the mud and you see the structures of logic itself. The large language model’s transubstantiation of alluvial soil into answers is the grand harvest of the society that grew it.
But subtract the dirt and there is no delta.
Subtract the corpus and there is no harvest.
There is nothing.
The model did not learn to reason in a vacuum. It absorbed rationality by observing rationality over and over and over again. Its powers of generalization are downstream of every example, correction, and argument it subsumed. A human decision somewhere in the echoes of history, culture, and science set the stage for today’s chatbot response. This cultivated intelligence grows from the sediment of human sensemaking, but there is no sediment here that was not deposited by someone.
Many of those someones are dead. They wrote the ancient texts, tested the baseline science, and recorded the history of the world from antiquity for the benefit of all of us still here. But too, many of those someones are alive. They are writing the working code that the model spits out. They’re pushing that baseline science past its frontier. They’re organizing and investigating and acting upon and reacting to the infinite feed of current events. Any response germane to today is borrowed from somebody.
In fact, you’re one of those somebodies. Literally.
Your 2am shitpost. That eloquent reply to a stranger’s essay. The scathing restaurant review you left. Your captions, comments, inside jokes, and all of your public conversations. Every contribution you ever made to the infinitely branching stream of digital communication, big and small, has settled somewhere in the delta.
Everybody Owns the Internet
The Nile River delta fed Egypt for five thousand years. The Mekong and the Ganges regions still feed hundreds of millions today. It’s no coincidence that every cradle of civilization owes its formation in whole or part to the floodplains and deltas of great rivers. These areas supported humanity through our most primitive eras with little more than the inherent richness of their raw materials. This dirt is begging to burst forth with life, yet somehow the richest farmland on earth is, almost without exception, accidental.
So too goes the internet.
We myriad digital denizens of the information superhighway did not set out to create a training corpus. We wrote for ourselves and for each other. We joked, argued, taught, complained, flirted, and debugged our way into this aggregated mass of interrelational raw material now harvested by private capital. The field of economics (which is also in the corpus) has vocabulary for this.
To categorize any resource, economists ask two questions. Is it excludable, and is it rivalrous?
More plainly, can you stop people from using it, and does one person using it diminish what’s left for everyone else?
There are caveats and sub-categories, but this simple test gives us a map.
If a good is excludable and rivalrous, it is a private good. Think about a sandwich. If I eat it, it is gone, and the law protects me from sandwich thieves.
If a good is excludable but mostly non-rivalrous, it is a club good. A Netflix subscription is a club good. If I watch a movie, you can still watch it too, but only if we both pay to have access.
If a good is hard to exclude people from using and rivalrous, it is a common-pool good. A pasture is the classic example. Many farmers can access the pasture, and while one cow grazing does not destroy the field, add enough cows and they’ll eventually gnaw the grass down to dirt. This is the infamous “Tragedy of the Commons” problem.
Finally, if a good is hard to exclude people from using and non-rivalrous, it is a public good. Streetlights are public goods. Once the street is lit, all of us can walk beneath the light, and my doing so does not darken the road for you.
Private and club goods are typically governed by profit-seeking actors and the legal system in which they operate. Public goods are primarily governed by governments or nobody, and common-pool goods tend to exist in a liminal space where everybody seeks the benefit and nobody wants to own the costs of upkeep.
The frontier labs generally argue that data on the internet is open for training under fair use copyright regimes. In economic terms, this argument implies the internet is a public good. The mass scraping, ingestion, and use of internet data for training does not destroy that original data. Every blog post, tweet, and flame war is indeed still there and for the most part accessible. Nobody clearly owns it.
Does the platform you post on own your posts? Do you share ownership with the platform? Can this relationship change over time?
You did post it online for free after all.
Except granting access is not the same thing as giving license. A library card gets you access to read a book, not to photocopy the entire library. Buying a national park pass does not confer logging rights. Visiting an open store does not entitle you to steal its inventory. Public access to work on the internet does not automatically confer usage rights.
And there is a second, deeper problem with “you posted it, you accepted this.” Until very recently, the LLM training data use case did not exist and could not have been reasonably foreseen by a party posting online. A blogger from 2008 could not have consented to their work being used to train a language model today, because that wasn’t conceivable back then. Consent can’t be assigned backwards in time, least of all for a sci-fi subplot turned real.
The current legal battleground for LLMs is a story of non-resolution.
Notably, despite our moral intuition, access and consent are irrelevant to the frontier labs’ primary legal defense claims of “fair use.” Instead, courts evaluate four criteria as they rule on a fair use defense. They look at the purpose of the use of copyrighted material, the nature of the work, the amount used, and the effect on the market for the original. In practice, these four items generally collapse to two important questions.
Is the new work transformative, and does it harm the market for the original?
In June of 2025, Judge Alsup ruled in Bartz v. Anthropic that training on legally acquired books was “quintessentially transformative,” but building its library from pirated books was “inherently, irredeemably infringing.” With this mixed victory, Anthropic faced a theoretical exposure of up to $70B in copyright damages and quickly settled the case for $1.5B a few months later. This is the largest copyright settlement in U.S. history (so far) and granted no future licenses to Anthropic, nor did it clarify any law going forward.
In a related ruling, Kadrey v. Meta, Judge Chhabria found LLM training similarly transformative and grudgingly ruled the evidence of market harm insufficient. In his ruling he criticized the plaintiffs for putting forth almost no evidence of market dilution and suggested that LLMs’ ability to flood a market with AI work similar to the training data "will often cause plaintiffs to decisively win the fourth factor—and thus win the fair use question overall—in cases like this."
Complicating the discussion further, the U.S. Copyright Office issued a non-binding report in 2025 concluding that public availability does not inherently allow fair use model training. As of this writing there is no settled legal standard for measuring LLM-driven market dilution, but this is primed to be a major confrontation in future legal decisions. Already, dozens of lawsuits and policy fights are testing the frontier labs’ evolving training-data defenses.
The labs’ most seductive defense is also the simplest.
“It’s just reading” is a common refrain among technologists defending AI model training, and it is a compelling argument. Every writer alive is built from the books they consumed. Nobody sends Hemingway’s estate a check for being inspired by The Old Man and the Sea. If the model is just another reader, it owes what every reader owes: nothing.
A person who reads ten thousand books in their lifetime becomes one more writer, working at human speed, publishing at human volume, and returns their sediment to the delta one grain at a time. A model that reads everything instead becomes a printing press that prints more printing presses. It spits out work at industrial volume, trains its successors, and competes with the very writers it consumed, at the push of a button. Inspiration never diluted a market, but printing does.
The invention of the Gutenberg press around the year 1440 ultimately led to the passage of the Statute of Anne in 1710. A cartel of powerful book publishers lobbied British Parliament to restore their monopoly rights over the book trade, and Parliament instead vested the right in authors as legal owners. The incumbents asked for protection and the public’s representatives handed ownership to the creators.
This statute established the basis of modern copyright law. Before the printing press, this wasn’t really necessary because mass piracy was practically impossible. The new technological landscape triggered a reproduction cascade that overwhelmed legal systems designed for a previous era's problems, but that reckoning took over two and a half centuries to play out.
The printing press that prints more printing presses will not let us wait that long.
Spoiling the Delta
On its face, the rich river delta that holds the deposits of humanity’s collective knowledge does appear to be a public good. Frontier labs scraping and ingesting the massive sedimental body of the internet does not destroy the original materials in a literal sense. The courts have already started ruling in that direction, but, like many legal rulings, this is narrowly correct, and completely misses the point.
A flat understanding of the training corpus question misunderstands how the internet’s functional layers and its participants actually interact. So far we’ve analyzed just the text layer. The webpages, articles, posts, comments, and everything else that the frontier labs scraped into a training corpus are one obvious piece, but there are many other layers of the internet, and they set the conditions for the text layer to exist at all.
Besides the obvious technical layers like the protocol or access layer, we must also consider the discovery layer, the attention layer, the contribution layer, and the integrity layer of the internet along with the flow of behavior between them. The continued utility of the internet depends on people finding, engaging with, contributing to, and ultimately believing in the value of the things they access online.
When framed as a static corpus, it is not obvious that the internet is damaged by AI training runs. Certainly it’s not damaged in the same way too many cows can damage the grass in a pasture.
Instead, what’s actually damaged is the complex system that evolved to enrich that corpus in the first place. The internet is a stack of interconnected public, club, and common-pool goods. Different layers react to and reinforce each other to make the whole valuable yet also vulnerable to the specific harms introduced by LLMs.
No human maker can compete with the raw volume of generative output working to overwhelm our algorithms and attention spans. The layers of the internet behave less like a pasture here and more like a road or an email inbox. They’re non-rivalrous up to a threshold, then catastrophically rival.
Consequently, the incentive to earnestly participate in the web diminishes with every AI variation of a derivative of a tweet of a derivative. Why make and share things online if you won’t get seen, can’t compete with the 10,000 variations of AI dogs dancing to upbeat electronica, and when you finally make something genuinely impressive the top comments just accuse you of being AI? The ability of generative AI tools to flood any corner of the web with media, slop or no, at effectively zero marginal cost might just be the final pedal stuck to the floor of the Spam-Everyone-Forever-Bus that left the station way back in the 90s.
This is an important moment. The naïve harvesting of the fertile corpus layer is a broadside against the very people who made it possible. Some parts of the web have likely already broken. If we play this wrong, the entire internet may irreparably break. Yet we’re not without tools to help us here.
We already know how to protect a commons. Elinor Ostrom won a Nobel Prize in 2009 for documenting how Swiss alpine pastures, Japanese forests, and Spanish irrigation networks sustainably shared their commons for centuries. She identified eight conditions for enduring commons: clear boundaries on who may draw from it, rules matched to local conditions, the people affected having a say in those rules, monitoring by parties accountable to users, graduated penalties for overuse, accessible ways to resolve disputes, recognition of the community’s right to organize, and governance nested across scales.
Run the internet through this checklist and almost none of the eight conditions hold. Its boundaries are hazy, fueled by everyone and fenced by no one. The people who fill it have no say in how it is governed. Its rules are unclear and only sporadically litigated, and then only by a handful of well-capitalized parties. Oversight is thin where it exists at all, always retroactive and never proactive. There is no monitoring, no graduated penalty, no shared venue to resolve disputes. It is, in Ostrom's precise sense, not a governed commons at all. It is a common-pool good with the plug pulled.
That’s why we’re in this mess. Much like in the era of the oozing Cuyahoga and the slurry-drowned Buffalo Creek, we’re staring down a cyberindustrial runoff disaster poised to spoil the entire delta.
Attribution Collapse
The river still flows for now. Fresh sediment continues to settle in the delta, the corpus layer continues to grow, and the labs continue to scrape.
And as they scrape, they continue to compress the colossal delta of the internet into fixed sets of weights, but this ongoing ritual manifests another problem. This time a problem of value, not quality.
Specifically, the problem of who gets paid for what value.
According to the frontier labs, all of these billions of scraped data points are somehow individually worthless, yet collectively worth trillions.
By worthless, they mean that no individually scraped work is needed in the training set. Remove any one piece and the model barely notices. Therefore no single work really matters. Therefore no single work is owed payment.
But if we can unslack our hanging jaw long enough to chew what they’re feeding us, we can see the individual data points are clearly not worthless.
This is a rhetorical trick. It’s doublespeak from a self-appointed detective declaring “since we can’t figure out exactly how much jewelry was stolen, no charges can be filed.” Except they’re also the thief. And just opened up a jewelry store.
This “aw shucks” is, of course, preposterous. Poor accounting practices do not erase the clear transfer of value, especially when the accounting is impossible.
In principle, the accounting is impossible from a legal standpoint because copyright law was built to police discrete copying. Court precedents assume infringers copying enumerable works from legible parties. Large language model training is the statistical absorption of billions of works at once. It’s a different shape with the same moral essence, but since the extracting act is a new mechanism, the legal instrument cannot quite grip it (yet).
But even more concerning is that attempts at quantitative accounting may just be mathematically incoherent.
The leading formal method to value a training input is the Shapley value, which averages an input's marginal contribution across every ordering in which it could appear. But that number isn’t a property of the work itself. It’s a function of the work’s relationship with every other work in the training set. The same document in a different training set will have a different Shapley value. Train the same model twice and, because training is stochastic, the Shapley value might change from run to run. Researchers do not even agree that Shapley is the right contribution metric, and calculating true Shapley values for frontier-scale models is computationally infeasible. These models can take weeks to train once; exact Shapley accounting would require retraining across impossible combinations of inputs. So as of today, there is no objective valuation scheme for calculating any work's specific share that would not be litigated into oblivion the moment it was implemented.
Individual attribution for LLM training at frontier scale will not work for the foreseeable future. You cannot pay people in proportion to their contribution because no administrable, specific share exists. This is the root of the misdirect. The labs interpret this fact to mean “if we can’t attribute then we owe nothing.”
I argue that it means you can’t pay proportionally, but the payment is still owed.
Paying Public Genius
Here is the true shape of the problem. Individuals create singular work, but never in isolation. The internet is communal by nature. Both collaboration and conflict feed the whole.
Branches grow stronger when pruned. The richest soil is built from rot. Vines climb by contact. Friendly minds cross-pollinate ideas while predators and prey run each other faster. This teeming mass of garbage and brilliance is valuable precisely because of the varied relationships each piece has with the others. In the living distance between them, human cognition buds, blooms, and bears fruit. Despite the romance of solitary genius, the internet is a team effort.
And if you cannot identify the most valuable player, you pay the team.
That payment is a royalty. It is a dollars-and-cents accounting of the public genius LLMs extract from the best and worst of us.
Call it the Corpus Royalty.
The frontier labs pay a fixed share of gross revenue into a public fund. The fund pays every eligible American the same amount each year. Any mechanism cleverer than this reimports the measurement problem, and it dies ten thousand quibbling deaths in the courtroom.
As LLMs fuse themselves into the internet and reshape the incentives for human expression, a royalty becomes the only coherent answer to the question of compensating collective, unattributable contribution. Frontier labs cannot continue their harvest of the internet unfettered. They must replenish the upstream sources that feed the fertile delta their models depend on, or the internet will become unrecognizable within the decade.
We’ve built smaller versions of this machinery before.
When private entities profit from shared resources, we recognize the public is owed a claim on the proceeds. The Alaska Permanent Fund follows this intuition. Every eligible resident receives a share of resource wealth no one resident can individually claim. Since we cannot reliably measure what any single person’s words are worth to a model, the distribution should reflect the failure of attribution rather than pretend to solve it.
When individual claims are too numerous to price one by one, we do not pretend they have no value, and when private entities damage public spaces, we declare they are culpable for that destruction. After a century of burning rivers, Congress built Superfund in 1980 and handed the cleanup bill to the polluters, for dumping that was legal when it happened. Nobody had to trace which barrel poisoned which well. The industries that profited from the toxic waste paid to restore the ground.
Some will say the Bell precedent argues for opening the weights, not cutting checks, but remedies follow wounds. In Bell’s case, competitors were wounded by intellectual lockout, so the remedy was access. Today, contributors are wounded by the extraction of intellectual value. It follows that payment is the remedy that makes the wounded whole.
The Corpus Royalty may be small at first. Perhaps just enough for an extra case of beer per year, but the amount matters less than the standing it confers. It shows the public they are more than just raw material exploited to train increasingly large language models. If the labs are right about what they are building, beer money becomes grocery money becomes rent money that grows with the labs and their revenues. If the labs are wrong, it won’t be because a royalty killed the business model. Normal people’s lives, arguments, questions, jokes, corrections, and creations help sustain the corpus those models consume, and this brings them along proactively instead of parasitically. The corpus is either essential or it isn’t. If it is, it has a price, and industries pay for essential inputs every day. The labs have already conceded as much in their licensing deals with Reddit, News Corp, and the Associated Press.
This is not welfare, because welfare assumes the companies are subsidizing the public, when the subsidy runs the other way. This is not charity, because charity implies nothing was received in return. This is not a tax, because a tax treats the surplus as company property subject to public claim.
This is restitution.
The legal word for this shape of problem is unjust enrichment, but the common law version is too small for the thing now in front of us. In ordinary law, unjust enrichment asks whether one party has benefited at another’s expense under circumstances that make keeping the whole benefit inequitable.
The frontier labs have received such a benefit by converting an uncompensated, massively aggregated, publicly generated corpus into private infrastructure-level value while threatening the conditions under which that corpus is renewed. They have done this at a scale and level of diffusion that individual litigation cannot sensibly price. The size of the problem tells us the remedy must be collective.
A royalty paid on value rooted in the public corpus is a return. What flows back to the internet is owed, not gifted. This is a royalty on public genius.
A royalty is only part of the solution. It’s one piece of a larger system of contribution and sustainment. It does not replace copyright claims or private licensing contracts. Those can and should still happen where ownership is legible. The royalty solves for the unattributable long tail of creativity that cannot organize, negotiate, or litigate its way into the licensing market.
What the labs are doing is not new in kind, only in scale. This is the private capture of public genius at civilizational scale. This should not surprise us. Corporate entities built atop public support often try to privatize the upside while socializing the conditions that made it possible. Special organizations have abused their special status this way before. The difference is that this time the affected class is everyone at once.
Perhaps we have been building toward this since we first scratched marks into clay more than 5,000 years ago. We’ve recorded, collected, and categorized our way into fragility. Once every externalized thought, every written word, every diagram, flourish, and turn of phrase can be rolled into a mechanical genie, who could resist the temptation to sell it back to the people who supplied it?
This is why the public needs a claim. The frontier labs increasingly seek the privilege and power of a utility without accepting the public obligations that come along with it. A Corpus Royalty restores some of the balance in that bargain by letting the public collect a share of the wealth it made possible. The public already bears the downside of the world these models are creating and deserves part of the upside.
The Corpus Royalty ensures humankind’s scattered sparks of brilliance have skin in the game they helped create.
