Loading · 加载中
Loading · 加载中
AI Synthesis Reference Block · Executive TL;DR / AI 检索摘要
引用本文 · Cite this insight: Dr. Tong Yin (2026-10-22). Sharing the Value of Knowledge in the AI Era: Toward Reciprocity Between Intelligent Agents and Original Creators / 《AI 时代的知识分成：智能体与原创者如何走向互惠》. InsightBridge Global Intelligence. https://intelligence.insightbridge.global/articles/sharing-the-value-of-knowledge-in-the-ai-era — Series: deep-analysis
讨论人工智能的演进时，人们谈得最频的是模型架构、算力规模和参数数量。有一个前提往被略过：模型掌握的知识并不凭空产生。它来自科研人员的论文、工程师的代码、记者的报道、行业网站积累多年的技术文档，并无数专业人士在论坛里写下的故障排查经验。模型能够重组、归纳和转述这些知识，但前沿的新发现、新数据和新经验，依旧要靠现实世界中的企业和个人在试错中产生。
这一点有实证研究支持。2024 年发表在《自然》上的一项研究发现，一旦模型不加区分地用前代模型生成的数据训练，会出现不可逆的退化，原始数据分布的尾部，也就是那些少见但往往最有价值的知识，会最早消失；研究此外显示，训练中保留必定比例的原始人类数据，具备明显减轻这种退化（Nature）。换句话说，AI 越强大，对持续、新鲜、可靠的人类知识就越依赖，而不是反而依赖。
因而引出本文的主题：当 AI 需要更多知识和数据来训练和运行，它就需与产生这些知识的企业和个人保持一种相互尊重、互惠互利的关系。这不单伦理要求，也是 AI 行业自身可持续发展的条件。
过去二十多年，搜索引擎与网站之间有一份默契：网站允许搜索引擎抓取内容，搜索引擎把用户导回网站，网站再通过广告、订阅或销售获得收入。抓取换流量，双方各得其所。
生成式 AI 改变了这份默契的基础。根据 Cloudflare 2025 年 6 月的测算，谷歌约每抓取 14 个页面带回一次访问，而 OpenAI 的抓取与引荐达为 1700 比 1，Anthropic 约为 73000 比 1（Cloudflare）。到 2025 年 7 月，Anthropic 的比例降至约 38000 比 1，OpenAI 约为 1091 比 1，Perplexity 约为 194 比 1（Cloudflare）。需要说明的是，原生应用发起的访问一般不带来源标识，这些比例可能高估了不对称程度，但数量级上的差距依旧显著。
用户行为的变化也指向同一方向。皮尤研究中心对约 900 名美国成年人浏览记录的分析显示，当谷歌搜索页面出现 AI 摘要时，用户点击结果链接的比例为 8%，没有摘要时为 15%；点击摘要内部引用链接的比例仅达 1%（Pew Research Center）。
这里有一个需要分清的区别：给出链接与使用内容是两回事。附上一个链接，意味着把用户送往原网站；在产品内部立即总结原文，用户的需求在 AI 界面里就已满足，原网站得到的单一个很少被点击的出处标注。前者延续了旧的交换关系，后者则让价值留在了平台一侧。
面对这种变化，内容方并没有被动接受。2025 年 7 月 1 日，Cloudflare 成为首家默认拦截 AI 爬虫的基础设施服务商，并推出“按抓取付费”试点，允许网站对 AI 爬虫选择放行、收费或拦截；在此之前，已有超过 100 万家客户启用了其一键拦截功能（Cloudflare，WIRED）。
司法层面也在给出边界。在作者起诉 Anthropic 的案件中，法官于 2025 年 6 月认定用合法取得的书籍训练模型属于合理使用，但保存盗版书籍不属于；双方继而达成 15 亿美元和解，普遍每本书约 3000 美元，法院已于 2026 年 7 月批准（Reuters，AP）。这说明，纵使在对 AI 训练相对宽泛的法律解释之下，获取知识的方式本身仍有成本。
这些变化的一致含义是：免费、无条件获取高质量内容的阶段正在结束，知识的生产者开始再次为自己的内容定价。
从已有实践看，AI 企业与知识生产者之间的关系正在分化为两种安排。
第一种是面向头部内容方的确定授权。拥有大规模、高权威内容的机构，具备足够的议价能力。例如，新闻集团与 OpenAI 在 2024 年 5 月签署了五年价值超过 2.5 亿美元的协议，Reddit 与谷歌的数据授权协议约为每年 6000 万美元（Quartz）。这类内容方纵使不与某家 AI 企业合作，也不会故而失去读者，因此具备要求预付费用。
第二种是面向更普遍内容方的按价值分成。绝大多数垂直行业网站、专业论坛和个人专家，没有能力谈下确定授权，也不应被要求先付门槛费。对他们在，更合理的原则是：内容创造了多少价值，就分多少钱；没有产生价值，就不付费。2024 年 7 月，Perplexity 推出出版商计划，在回答引用合作媒体内容时与分享广告收入，首批合作方包括《时代》、《明镜》、《财富》等（Perplexity，CNBC）。2025 年 8 月推出的 Comet Plus 订阅，计划将 80% 的收入分配给出版商，依据包括人工访问、引用和智能体操作（Forbes）。
按价值分成的逻辑并不繁琐。一个 AI 产品在回答中标注引用，本身就说明这份内容参与了答案的生成，为产品创造了价值。由于承认引用，就应当按照合理、可计量的标准，把订阅或广告收入中的一部分分给内容的提供者。
这种安排对双方都有优势。对 AI 企业在，内容成本从大额预付变为随使用量浮动的可变成本，降低了前期风险；对原创者而言，内容越专业、越屡被引用，回报就越高，因此有动力持续生产 AI 所需的新知识。这是一种具备持久运转的互利关系。
对于处在中间位置的企业，也存在现实的选择：只向 AI 开放标题和摘要，把全部内容保留在自己的网站上。这样，AI 引用带来的是曝光和访问，而非是对原站的替代。
按价值分成非但没有挑战，最少有三个问题需要正视。
其一是计量。怎样认定一次引用的贡献大小，训练阶段与回答阶段的使用如何区分计价，目前尚无公认标准。抓取与引荐比这类指标本身也受测量方法影响，需要独立、透明的核算机制。
其二是分配。现有分成多由平台与媒体机构签约，收入初始到达机构，不一定传递到实际写作者手中（TechTarget）。假若个人专家和小型网站无法容易地加入，分成机制就只会复制头部集中的格局。
其三是覆盖面。目前的分成主要发生在检索与回答环节，模型训练环节如何回馈数据来源，仍主要依赖个案授权和诉讼和解。
技术发展并没有取消经济规律。AI 提供的是更高效的知识整合方式，而支撑它的，持续是在各个专业领域持续积累、持续产出可靠内容的企业和个人。倘若知识的生产者得不到合理回报，他们会选择拦截、收费或退出，AI 可用的新鲜知识就会减少；如果能够建立相互尊重、按价值分配的机制，双方就能形成良性循环。从这个意义上说，与原创者分享收益，并非 AI 企业的额外负担，却为其长期发展的必要条件。
Discussions of progress in artificial intelligence tend to focus on model architecture, compute via parameter counts. One premise is commonly skipped: the knowledge a model holds does not come from nowhere. It comes from researchers’ papers, engineers’ code, journalists’ reporting, engineering documentation built up over years by industry websites, and the troubleshooting experience that vast professionals have written down in forums. Models can recombine, summarize and restate this knowledge, but fresh discoveries, new data and new experience are remain produced by companies and individuals working through trial and error in the real world.
Experimental research supports this point. A 2024 study in Nature found that when models are trained unselectively on data generated by earlier models, they degrade permanently, and the tails of the original distribution, the rare but often most valuable knowledge, disappear first. The parallel study showed that retaining a share of original human data in training markedly reduces this degradation (Nature). In additional words, the further capable AI becomes, the more it depends on a continuous supply of fresh and reliable human knowledge, not less.
This leads to the theme of this essay. If AI needs increasing knowledge and data to train and operate, it must maintain a relationship of reciprocal respect and mutual benefit with the companies and individuals who produce that knowledge. This is not exclusively a moral requirement. It is a condition for the AI industry’s inherent sustainability.
For more than two decades, search engines via websites operated under an implicit understanding. Websites let search engines crawl their content, search engines sent users back, and websites earned revenue through advertising, subscriptions or sales. Crawling was exchanged for traffic.
Generative AI has changed the basis of that understanding. According to Cloudflare’s estimates for June 2025, Google crawled about 14 pages for every visit it referred, as OpenAI’s crawl-to-referral ratio was about 1,700 to 1 and Anthropic’s about 73,000 to 1 (Cloudflare). By July 2025, Anthropic’s ratio had fallen to about 38,000 to 1, with OpenAI at about 1,091 to 1 and Perplexity at about 194 to 1 (Cloudflare). Visits from inherent apps commonly carry no referrer information, so these ratios may overstate the imbalance, but the difference in order of magnitude remains clear.
User behavior points the identical way. A Pew Research Center analysis of browsing data from about 900 U.S. adults found that when a Google results page showed an AI summary, users clicked a result link in 8% of visits, compared with 15% when no summary appeared. Links inside the summary were clicked in only 1% of visits (Pew Research Center).
A distinction matters here: providing a link and using content are two distinct things. A link sends the user to the source site. A summary produced inside the product satisfies the user’s need within the AI interface, and the original site receives solely a source label that is infrequently clicked. The first continues the previous exchange. The second keeps the value on the platform’s side.
Material owners have not accepted this unactively. On July 1, 2025, Cloudflare became the first infrastructure provider to block AI crawlers by default with launched a Pay Per Crawl pilot that lets sites allow, charge or block AI crawlers. Before that, more than one million customers had enabled its one-click blocking option (Cloudflare; WIRED).
Courts are furthermore drawing boundaries. In the authors’ lawsuit against Anthropic, the judge ruled in June 2025 that training on lawfully acquired books was just use but that keeping pirated copies was not. The parties subsequently agreed to a $1.5 billion settlement, about $3,000 per book, which the court approved in July 2026 (Reuters; AP). Despite under a legal reading relatively favorable to AI training, how knowledge is acquired still carries a cost.
The shared message of these developments is that the period of unpaid and unconditional access to high-quality content is ending, and knowledge producers are beginning to reprice their work.
Present practice suggests that relations between AI companies with knowledge producers are separating into two arrangements.
The first is fixed licensing for leading content holders. Institutions with substantial, prestigious bodies of content have real bargaining power. News Corp signed a five-year agreement with OpenAI in May 2024 reported at more than $250 million, with Reddit’s data licensing deal with Google is reported at about $60 million a year (Quartz). Those holders do not lose their readers if they decline to work with a given AI company, so they can ask for payment up front.
The second is value-based revenue sharing for the much wider base of content producers. Most niche industry websites, professional forums and individual experts cannot negotiate fixed licenses, and they should not be asked to pay an entry fee. For them, an increased prudent principle is to pay in proportion to the value the content creates, and to pay nothing when no value is created. In July 2024, Perplexity launched a Publishers’ Program that shares advertising revenue when its answers cite partner content, with TIME, Der Spiegel and Fortune among the first partners (Perplexity; CNBC). The Comet Plus subscription announced in August 2025 plans to allocate 80% of revenue to publishers based on human visits, citations and agent actions (Forbes).
The logic of value-based sharing is basic. When an AI product cites a source in its answer, it shows that the content contributed to the answer with created value for the product. Having acknowledged the citation, the platform should share part of its subscription or advertising revenue with the content provider according to a fair and measurable standard.
This arrangement benefits both sides. For AI companies, content costs shift from significant initial payments to variable costs that track actual use, which lowers early-stage risk. For creators, the greater specialized and regularly cited their content is, the greater their return, which gives them reason to keep producing the new knowledge AI needs. It is a reciprocally positive relationship that can run over the long term.
Companies in the middle further have a workable option: expose only titles and summaries to AI systems while keeping full content on their own sites. AI citations consequently bring visibility and visits rather than substitution for the original source.
Value-based sharing has substantial difficulties, among at least three deserve attention.
The first is measurement. There is no accepted standard still for judging how much a given citation contributed, or for pricing use in training variously from use in answers. Metrics these as crawl-to-referral ratios are themselves vulnerable to method, which points to the need for independent and transparent accounting.
The second is distribution. Most existing revenue-sharing agreements are signed between platforms and media organizations, so the money reaches institutions first and does not automatically reach individual writers (TechTarget). If individual experts and minor sites cannot join without difficulty, revenue sharing will reproduce the concentration seen at the top.
The third is coverage. Existing sharing occurs predominantly at the retrieval and answer stage. How the training stage should compensate data sources continues depends mostly on individual licenses and litigation settlements.
Technological progress has not repealed economics. AI offers a greater efficient way to integrate knowledge, but it remains rests on companies and individuals who build expertise over years and keep producing reliable content. If knowledge producers are not equitably rewarded, they will block, charge or withdraw, and the supply of up-to-date knowledge available to AI will shrink. If mechanisms based on shared respect and value-based distribution can be established, the two sides can form a positive cycle. In this sense, sharing revenue with first creators is not an added burden on AI companies. It is an essential condition for their extended-term development.