[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-meta-withdraws-ai-training-clause-terms":3,"topics-all":35,"news-related-b485d30c-0640-4349-aa6f-69bdd9d3e935":54},{"id":4,"title":5,"summary":6,"content":7,"original_url":8,"source_id":9,"tags":10,"translations":21,"news_slug":28,"published_at":29,"created_at":30,"modified_at":31,"is_published":32,"publish_type":33,"image_url":14,"view_count":34},"b485d30c-0640-4349-aa6f-69bdd9d3e935","Meta 撤回条款中抓取训练条款,出版商赢得一次窄胜利","Meta 10 月中旬更新的商业条款一度授权用抓取工具访问商业关联网站,并把内容用于训练其 AI 模型,新条款定于 10 月 30 日生效。条款被指没有退出选项,遭 News\u002FMedia Alliance 谴责。10 月中下旬 Meta 删除训练表述,抓取权限收窄至广告与商业产品,并确认尊重 robots.txt。","训练数据的来源合法性,过去两年已经从技术问题变成了合同问题。整个行业长期靠一条心照不宣的默契运转:互联网上公开可访问的内容,默认就可以拿去训练模型。谁都知道这条默契站不住脚,但只要没有哪家公司把话说破,大家就都能继续跑。\n\nMeta 这次把话说破了。\n\n10 月中旬,Meta 更新商业服务条款,新条款定于 2026 年 10 月 30 日生效。改动本身藏在商业条款这种没人读的长文里,但它触及的是整个大模型行业最敏感的那根神经:数据从哪来。\n\n## 条款里到底写了什么\n\n争议最大的部分是一段关于\"抓取\"的授权。条款写明,Meta 可以使用\"自动化软件(诸如网页抓取工具、机器人、爬虫、蜘蛛、crawler 和 user agent)\"访问、抓取与商业产品和服务相关联的域名、网页或 URL,甚至可以创建\"测试凭据\"来完成这些动作。\n\n把这些工具名称逐一点名这件事本身就很说明问题——它不是随口一提的概括,而是在列举可执行的手段清单。\n\n真正引爆舆论的是后面那半句。原条款写着,Meta 可以使用来自上述域名、网页或 URL 的\"任何内容\",去\"提供、开发、改进、研究、训练并维护\"其产品、服务和功能的完整性。摘出来就是:\n\n> Meta may use any content from such domains, web pages, or URLs to provide, develop, improve, research, train, and maintain the integrity of, the Meta Company Products...\n\nCrypto Briefing 的分析指出,真正干重活的是 \"train\" 这个词。抓取本身不算新鲜事——广告推荐、链接预览都需要爬,业界早就见怪不怪。但把\"训练\"明确写进抓取授权里,性质就变了:这等于用一份合同,替所有站点的内容做了一次概括性的授权。\n\n条款覆盖的关联平台包括 Facebook、Instagram、WhatsApp 和 Messenger。还有一条常被忽略的细节:条款被描述为\"没有选择退出的选项\"——不同意更新条款的企业,必须停止对 Meta 商业产品的全部访问和使用。用 The Verge 的说法,就是接受条款或者离开,二选一。\n\n## 反弹来自哪里\n\n出版商行业组织 News\u002FMedia Alliance 的反应非常直接。该组织总裁兼 CEO Danielle Coffey 称这份政策\"令人愤怒,而且很可能违法\",并明确表示\"不能允许它在 10 月 30 日被实施\"。\n\n法律学者给出的意见更尖锐。康奈尔大学数字与信息法教授 James Grimmelmann 表示,对新闻出版商而言,这\"显然构成了 Meta 训练其网站上任何内容的许可\",并且可能让 Meta\"甚至在与出版商自身竞争的情况下\"部署这些内容。Nieman Lab 援引 Matt Topic 的话称,这一条款是\"埋在服务条款里的一条惊人规定\",并建议出版商假定它是可执行的。Solidot 转述法学教授 Eric Goldman 的分析则认为,新增条款在程序和实质层面都存在问题。\n\n最尖锐的角度来自许可费。Meta 已经与包括 CNN 和 Fox News 在内的媒体机构建立了付费 AI 授权合作——也就是说,Meta 在某些情况下是愿意付钱的。法律观察者向 Nieman Lab 表示,原有措辞可能让 Meta 绕开价值\"数百万美元\"的授权费用。这就让那段文字看起来像是给自家正在签的合同,准备了一条免费替代路径。\n\n## 三天之内撤回\n\nMeta 的反应速度是这件事里最值得记的部分。\n\n10 月 16 日至 17 日前后,Meta 再次改写商业条款,删除了那段\"训练\"授权。据 Crypto Briefing 的核实,新版条款把抓取权限限制在广告和商业产品范围内,移除了允许 AI 训练的表述,并确认尊重 robots.txt 等现有网络协议。\n\nMeta 传播副总裁 Andy Stone 的公开说法是,这次更新\"是为了确保我们能在做平台内广告和内容推荐时,从第三方网站上获取准确的在售商品信息,而不是关于抓取新闻网站\"。公司的发言人对 Nieman Lab 表示,不打算把这份条款当作无视 robots.txt 指令的法律依据,并称仍然允许新闻机构退出,同时提供如何阻止抓取的指引。\n\n## 我的判断\n\n撤回发生在 10 月 30 日生效日之前,这本身就是一个信号:商业条款是可以被舆论在生效前改写的,而且速度远快于任何立法程序。对内容创作者来说,这大概是这次风波里唯一的好消息。\n\n但新闻机构的处境仍然没有变好。这次赢得很窄:争议核心是 \"train\" 这个词被拿掉了,抓取权限本身还在,而且绑在商业产品上。更根本的问题是,新闻机构对 Meta 平台的分发依赖太深——正因如此,即使加一条 robots.txt 也拦不住,才必须靠舆论施压。技术手段在分发不对称面前,从来都不是解法。\n\n这件事对 AI 从业者的启示有点反直觉:训练数据的合规边界正在从\"技术手段\"迁移到\"合同条款\",而合同条款的改写速度,比立法快得多,也随意得多。与其猜条款里下一个会出现什么词,不如盯住这几个词——因为决定你能用什么数据的,往往不是数据集的许可证,而是别人的服务条款。\n\n当\"抓取\"两个字从技术名词变成合同名词,这一天其实早就到了,只是这次终于有人把账单摆到了桌面上。\n\n本文事实来源:Crypto Briefing 关于条款修订的报道(https:\u002F\u002Fcryptobriefing.com\u002Fmeta-removes-ai-scraping-language-terms\u002F),以及 Solidot 转载的事件梳理(https:\u002F\u002Fwww.solidot.org\u002Fstory?sid=85640)。","https:\u002F\u002Fcryptobriefing.com\u002Fmeta-removes-ai-scraping-language-terms\u002F","eec6878c-c201-4260-a113-659a37af2403",[11,15,18],{"id":12,"name":13,"slug":13,"description":14,"color":14},"c33b1bbc-d6ce-4f61-9d5d-1a0704a6a09b","ai-policy",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"1fcfaaf2-67de-43d3-9e35-5784852fec60","ai-safety",{"id":19,"name":20,"slug":20,"description":14,"color":14},"01598627-1ea6-4b27-a5d8-874971571a71","llm",[22],{"id":23,"lang":24,"title":25,"summary":26,"content":27},"031e77dd-5236-4318-b53a-b83ea59d911e","en","Meta Withdraws AI Training Clause, a Narrow Publisher Win","Meta’s mid-October terms let crawlers reach business-linked sites and train AI on them. Publishers objected; Meta then deleted the training language.","Where training data comes from has quietly shifted from a technical question to a contractual one. For the past two years the industry has run on an unspoken understanding: content publicly reachable on the internet is fair game for model training. Everyone knew the understanding was shaky, but as long as no company said so out loud, everyone kept running.\n\nMeta recently said it out loud.\n\nIn mid-October, Meta updated its commercial terms of service, with the new terms taking effect on October 30, 2026. The change sat inside the usual nobody-reads-it commercial boilerplate, but it touched the most sensitive nerve in the entire industry: where the data comes from.\n\n## What the clause actually said\n\nThe most controversial part was an authorization covering \"crawling.\" The terms said Meta could use \"automated software (such as web scrapers, bots, robots, spiders, crawlers, and user agents)\" to access, index, cache, analyze, collect, retrieve, or crawl domains, web pages, or URLs tied to Meta's business products and services — and even to create \"test credentials\" to do it.\n\nNaming those tools one by one is itself telling. It is not a casual generalization but an enumerated list of executable means.\n\nWhat set off the reaction was the second half of the sentence. The original terms said Meta could use \"any content from such domains, web pages, or URLs\" to \"provide, develop, improve, research, train, and maintain the integrity of\" its products, services, and features:\n\n> Meta may use any content from such domains, web pages, or URLs to provide, develop, improve, research, train, and maintain the integrity of, the Meta Company Products...\n\nCrypto Briefing's analysis points out that the word doing the heavy lifting was \"train.\" Crawling itself is nothing new — ad recommendations and link previews have always needed it, and the industry is long past being surprised by it. But writing \"train\" explicitly into a crawling authorization changes the nature of the thing: it uses a single contract to grant a sweeping, blanket license over every site's content.\n\nThe platforms covered included Facebook, Instagram, WhatsApp, and Messenger. One easily missed detail: the terms were described as having \"no option to opt out\" — businesses that did not agree to the updated terms would have to stop all access to and use of Meta's products for business or commercial purposes. In The Verge's framing, the choice was accept the terms or walk away.\n\n## Where the backlash came from\n\nThe reaction from the publishing trade group News\u002FMedia Alliance was blunt. The organization's president and CEO Danielle Coffey called the policy \"outrageous and probably illegal\" and said \"It cannot be allowed to be implemented on October 30.\"\n\nLegal scholars were sharper. Cornell digital and information law professor James Grimmelmann said that for news publishers this \"clearly represents permission to Meta to train AI on any content from their sites,\" and possibly to deploy it \"even in competition with the publishers themselves.\" Nieman Lab quoted Matt Topic describing the provision as \"a remarkable provision to bury in [the] terms of service,\" advising publishers to assume it was enforceable. Solidot, summarizing law professor Eric Goldman's analysis, wrote that the added terms had problems on both procedural and substantive grounds.\n\nThe sharpest angle concerned licensing fees. Meta has established paid AI licensing agreements with media outlets including CNN and Fox News — meaning Meta is willing to pay for news content in some cases. Legal observers told Nieman Lab the original wording could enable scraping news sites for AI training potentially worth \"many millions\" in licensing fees. That made the passage look like a free alternative to the very contracts Meta had been signing.\n\n## Pulled back within days\n\nThe speed of Meta's reversal is the most telling part of this story.\n\nAround October 16-17, 2026, Meta rewrote the commercial terms again, removing the \"train\" authorization. According to Crypto Briefing's verification, the revised terms limit crawling permissions to advertising and business products, strip out the language allowing AI training, and affirm respect for existing web protocols such as robots.txt.\n\nMeta's vice president of communications, Andy Stone, said the update \"is designed to make sure we're able to pull accurate information about products for sale on third-party websites when making on-platform ad and content recommendations, not about scraping news sites.\" A company spokesperson told Nieman Lab the company does not intend to use the terms as a legal basis to ignore robots.txt instructions, and still allows news organizations to opt out with guidance on how to block crawling.\n\n## My read\n\nThe withdrawal landed before the October 30 effective date, and that timing is itself the signal: commercial terms can be rewritten by public pressure before they take effect, and far faster than any legislative process. For content creators, that is probably the one genuinely good piece of news in this fight.\n\nBut news organizations are not much better off. The win was narrow: the dispute centered on the word \"train,\" which was removed, while the crawling permission itself remains and stays tied to business products. The deeper problem is untouched — newsrooms depend on Meta's platforms for distribution so heavily that a robots.txt line cannot stop it. Which is exactly why it took public pressure rather than a technical fix. In the face of asymmetric distribution, technical means were never going to be the answer.\n\nThe lesson for people building AI systems is counterintuitive: the compliance boundary around training data is migrating away from technical controls and toward contract language — and contract language can be rewritten faster and more arbitrarily than legislation. Rather than guessing which word shows up next in a terms of service, watch that handful of words. Because what decides what data you can use is often not the dataset's license, but somebody else's terms of service.\n\nThe day \"crawling\" stopped being a technical term and became a contractual one arrived long ago. This time, someone finally put the bill on the table.\n\nSources for this piece: Crypto Briefing on the revised terms (https:\u002F\u002Fcryptobriefing.com\u002Fmeta-removes-ai-scraping-language-terms\u002F), and Solidot's summary of the original event (https:\u002F\u002Fwww.solidot.org\u002Fstory?sid=85640).","meta-withdraws-ai-training-clause-terms","2026-10-21T07:10:00Z","2026-10-21T07:12:31.486000Z","2026-10-21T07:12:31.486011Z",true,"agent",44,[36,45],{"slug":37,"tag_slug":37,"title_zh":38,"title_en":39,"intro_zh":40,"intro_en":41,"id":42,"is_active":32,"created_at":43,"modified_at":44},"ai-for-science","AI for Science 2026：从 UniPert 到 GPT-Rosalind 的硬核进化","AI for Science 2026: from UniPert to GPT-Rosalind","生命科学、化学材料、物理世界模型——AI 正在从\"语言工具\"变成\"实验伙伴\"。本专题收录 AI 在三大科学方向的关键节点：UniPert 统一基因与化学扰动空间、GPT-Rosalind 端到端生命科学推理、达摩院 AI 智能体 28 小时找到 4 种超导新材料、Anthropic Claude Science 把工作台做成标准品。","From language tool to lab partner — AI is reshaping life sciences, chemistry\u002Fmaterials, and physical world models. This topic covers the key milestones: UniPert unifying genetic-chemical perturbation spaces, GPT-Rosalind's end-to-end life-sciences reasoning, DAMO's AI agent discovering 4 superconducting materials in 28 hours, and Anthropic's Claude Science workbench going mainstream.","988a4300-5fab-41c4-b5d8-63711a2dc757","2026-09-10T01:34:15.296649Z","2026-09-10T01:34:15.296663Z",{"slug":46,"tag_slug":46,"title_zh":47,"title_en":48,"intro_zh":49,"intro_en":50,"id":51,"is_active":32,"created_at":52,"modified_at":53},"h3-series","MiniMax H3 系列：从开源权重到 35 倍吞吐","MiniMax H3 Series: from open weights to 35x throughput","MiniMax H3 自 2026 年 8 月开源以来节奏密集：官方把生成、参考与编辑收回一个模型；ComfyUI 当天压进 RTX 3060；摩尔线程 3 小时完成国产 GPU 适配；fal 后训练版把吞吐拉到 35 倍；FastH3 蒸馏再砍推理成本。本专题持续追踪 H3 的发布—开源—蒸馏—部署全链路。","Since MiniMax open-sourced H3 in August 2026 the pace has been relentless: one unified omni-modal model, same-day ComfyUI support down to an RTX 3060, a 3-hour Day-0 port to Moore Threads GPUs, fal's post-trained H3 Max at 35x throughput, and FastH3 distillation cutting inference cost further. This topic tracks the full H3 chain — release, open weights, distillation, deployment.","83ef0daa-3c31-4cb3-86ed-e5ee58654d5f","2026-09-08T07:33:19.942193Z","2026-09-08T07:33:19.942209Z",{"items":55},[56,61,66,71,76,81],{"id":57,"title":58,"news_slug":59,"published_at":60},"c2c96e3a-cc8d-4970-8d9d-e9584bbc9237","WhatsApp 帮助页藏 102 块 LLM 暗指令","whatsapp-hidden-llm-instructions-help-pages","2026-10-19T16:00:00+00:00",{"id":62,"title":63,"news_slug":64,"published_at":65},"66c8c031-ce38-4fc5-b654-362bef543c6e","OpenAI textGrain:水印先锁欧盟,API 默认关","openai-textgrain-eu-watermark","2026-10-08T07:00:00+00:00",{"id":67,"title":68,"news_slug":69,"published_at":70},"e67e70ff-6e41-48fb-a22d-9249eca411ae","AI 炒作叙事不需要全部买账","ai-hype-mythos-counter","2026-10-05T05:00:00+00:00",{"id":72,"title":73,"news_slug":74,"published_at":75},"ea6de12d-e8bf-4013-a299-14211805ba30","ChatGPT __obi cookie 把你带到站外:1 年同站标识","openai-obi-cookie-cross-site-tracking","2026-09-30T00:00:00+00:00",{"id":77,"title":78,"news_slug":79,"published_at":80},"9cae377e-d048-4d0e-af78-8cba3e1cf5b4","法庭文件揭微软高管私下定性 AI 抓取为最大盗窃","ai-scraping-largest-theft-hecht-memo","2026-09-29T04:00:00+00:00",{"id":82,"title":83,"news_slug":84,"published_at":85},"7f31f734-c682-4d49-8fda-43606ecd9546","「失控模型」话术拆解:AI 炒作正在替公司免责","rogue-model-hype-accountability-escape","2026-09-25T13:07:00+00:00"]