[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-openai-textgrain-eu-watermark":3,"topics-all":38,"news-related-66c8c031-ce38-4fc5-b654-362bef543c6e":57},{"id":4,"title":5,"summary":6,"content":7,"original_url":8,"source_id":9,"tags":10,"translations":24,"news_slug":31,"published_at":32,"created_at":33,"modified_at":34,"is_published":35,"publish_type":36,"image_url":14,"view_count":37},"66c8c031-ce38-4fc5-b654-362bef543c6e","OpenAI textGrain:水印先锁欧盟,API 默认关","OpenAI 自研水印 textGrain 欧盟 ChatGPT 与 Codex 默认开,API opt-in。1% 误报率下 200 token 检测约 80%,数学文本因词空间小更难检出。10% 同义替换让检测率从 92% 跌到 66%,25% 替换仅 17%。OpenAI 强调水印只给有限 AI 信号。","欧盟 AI Act 第 50 条对生成式 AI 的「机器可读」标记要求落地三个月后,OpenAI 终于在 10 月 5 日公开了自己的应对方案:自研文本水印技术 textGrain。这项技术将在欧盟范围内的 ChatGPT 和 Codex 中默认开启,而全球 API 用户则需要手动 opt-in。同时,水印检测器只对受邀的研究人员和专家机构开放,不直接面对公众。这是 OpenAI 第一次把文本层面的内容溯源做到「出厂级」,也是继 Anthropic 在 8 月宣布接入 Google DeepMind 的 SynthID-Text 之后,第二家公开以水印应对欧盟合规的头部厂商。\n\n## textGrain 到底做什么\n\n按照 OpenAI 官方的说法,textGrain 通过在模型生成文本时,对词的选择过程施加一个不可见的统计信号,让生成文本携带可被检测器识别但不影响阅读的「指纹」。检测器只需要一段足够长的文本,就能以高于随机的概率判断这段文字是否由 OpenAI 模型生成。OpenAI 把这项技术定位为「内容来源透明」的工具,而不是内容真伪、版权归属或用户身份的判据。值得注意的一点是,OpenAI 表示计划将这项技术开源,允许外部社区继续改进。\n\n## 检测性能:看起来漂亮,但离可用还差一截\n\nOpenAI 在官方博客中给出了自己的评测数据。在目标误报率 1% 的前提下,200 token 长度的心理学类文本段落,水印检测率约为 80%;400 token 长度则提升到约 95%。但对数学这类词选择空间极小的文本,检测率明显下滑。同样是 1% 误报率下,数学段落的表现要差得多,OpenAI 的官方图表也展示了这一点。\n\n更刺眼的数字在抗编辑能力上。把一段 400 token 的水印文本替换 10% 的词为同义词,检测率从约 92% 跌到 66%;替换 25% 的词,只剩 17%。也就是说,只要对水印文本做轻度改写,信号就会迅速衰减。这与两个月前 Anthropic 上线水印后被开源去除工具在四小时内撕开的研究结论是一致的:文本水印在「忠实传播」假设下成立,一旦进入「人类润色后转发」的常规场景,可靠性就显著退化。\n\n## 性能影响:几乎为零\n\nOpenAI 同时在 Astra(其前沿模型)上跑了对照测试,把同一组 benchmark 各跑了一遍水印\u002F无水印版本。从公布的数据看,差距都在噪声范围内:Artificial Analysis Intelligence Index 49.57 vs 49.76、AutomationBench 34.09% vs 34.86%、DeepSWE v1.1 72.80% vs 71.68%、Terminal-Bench 4.0 53.90% vs 56.06%、BrowseComp 87.92% vs 87.35%、GPQA Diamond 94.44% vs 93.94%。换句话说,水印机制没有以牺牲生成质量为代价。这一点的工程意义大于表面数字:它意味着 textGrain 在推理路径上是「廉价」的,不是事后重新编码,而是直接作用在词采样阶段。\n\n## 「水印告诉你什么」:OpenAI 主动划界\n\n与很多炒作叙事不同,OpenAI 这次在博客里主动列出了水印「不能告诉你」的清单:它无法衡量人类贡献、无法确立版权归属、无法识别具体用户、无法验证文本是否真实、无法用「未检测到水印」反推一定是人类写的。换句话说,即使水印方案成熟,它也只是一个「AI 是否参与过」的弱信号,而不是「内容是否可信」的强证据。这种主动划界在监管语境下很关键,因为如果未来有平台把水印检测结果直接当作内容打标签的依据,OpenAI 已经在文档层面做了免责声明。\n\n## 区域策略:为什么先只锁欧盟\n\nOpenAI 选择「欧盟默认开启 + 全球 API opt-in + 检测器限研究人员」的渐进式部署,而不是一刀切全球默认,理由藏在欧盟 AI Act 第 50 条的执行时间表里。其它司法管辖区目前还没有同等强度的「机器可读」强制要求,提前开启会带来合规和商业摩擦(比如某些客户会担心水印削弱内容独占性)。同时,把检测器先只开放给研究机构,也让 OpenAI 能在「数据反馈-检测器迭代」的循环里继续积累真实样本,而不必直接面对检测错误可能引发的舆论反弹。\n\n## 水印军备竞赛:从图像到文本的下半场\n\n回头看,文本水印的难点和图像水印完全不同。图像是感知冗余高的模态,几个像素的扰动不破坏视觉,但文本是离散符号,任何词替换都会改变语义。OpenAI 选择在词采样阶段注入统计信号而不是事后改写,本身就是承认了「文本水印必须在生成时刻做」这一约束。但这同时意味着,水印无法对开源模型、被微调的私有模型、以及使用 OpenAI API 但自己重排 token 的下游应用提供任何保护。所以,textGrain 的真实覆盖范围,可能比 OpenAI 的市场宣传要小得多。\n\n更值得注意的是,OpenAI 在文中提到将与云合作伙伴合作,在其分发渠道中也支持水印。这是一个明确的信号:合规驱动的水印部署,正在从「模型厂商单点」向「模型+云平台」两端延伸。下一步,真正的考验不是水印能不能检出,而是平台愿不愿意把检测结果作为内容分发决策的依据,以及用户愿不愿意接受「我被标记为 AI 生成」的代价。","https:\u002F\u002Fopenai.com\u002Findex\u002Feu-text-provenance\u002F","bd0e0e04-6bcf-4b3e-9a56-62c672308ec9",[11,15,18,21],{"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",{"id":22,"name":23,"slug":23,"description":14,"color":14},"42e59a88-7795-47dc-a334-ef1e72c24347","openai",[25],{"id":26,"lang":27,"title":28,"summary":29,"content":30},"fd69a4a6-9473-45f7-a451-0380dad5afab","en","OpenAI textGrain: EU watermark on by default, API off","OpenAI ships textGrain: EU ChatGPT and Codex default-on, API opt-in. 1% FP, 200-token detect ~80%. 10% swap: 92%→66%, 25%: 17%. Signal: AI only.","Three months after the EU AI Act's Article 50 began to bite, OpenAI finally published its compliance plan on October 5: a proprietary text watermarking technology called textGrain. It will be enabled by default for ChatGPT and Codex users in the European Union, while global API customers will have to opt in manually. At the same time, the watermark detector will be open only to invited researchers and expert organisations, not to the general public. This is OpenAI's first attempt at taking text provenance to the \"factory\" level, and it follows Anthropic's August announcement that it would adopt Google DeepMind's SynthID-Text.\n\n## What textGrain actually does\n\nAccording to OpenAI's official blog, textGrain injects an invisible statistical signal into the model's word-choice process during generation, so that the produced text carries a fingerprint the detector can recognise but readers cannot see. The detector only needs a sufficiently long passage to judge, at a probability well above chance, whether OpenAI's models generated the text. OpenAI frames the technology as a content-provenance tool, not as a verdict on truth, ownership, or user identity. Notably, the company says it plans to release the technology as open source so the broader community can build on it.\n\n## Detection performance: looks good on paper, less so in the wild\n\nOpenAI shared its own evaluation data in the blog post. At a 1% target false-positive rate, the detector identifies watermarks in about 80% of 200-token psychology passages, climbing to about 95% at 400 tokens. For mathematics, where the space of valid word choices is much tighter, detection drops substantially. OpenAI's own chart makes this gap explicit.\n\nThe numbers get uglier once editing enters the picture. Replacing 10% of words with synonyms in a 400-token watermarked passage dragged the detection rate from about 92% down to 66%. Replace 25% of the words and only 17% survive. In other words, light paraphrasing is enough to erode the signal quickly. That matches what independent researchers reported two months ago when they cracked Anthropic's watermark hours after launch: text watermarks hold under a \"verbatim redistribution\" assumption, but fall apart once humans start rewriting, which is exactly what happens on most platforms.\n\n## Impact on model quality: essentially zero\n\nOpenAI also ran its Astra frontier model through a paired benchmark suite, watermarked versus not. The deltas sit comfortably inside the noise band: Artificial Analysis Intelligence Index 49.57 vs 49.76, AutomationBench 34.09% vs 34.86%, DeepSWE v1.1 72.80% vs 71.68%, Terminal-Bench 4.0 53.90% vs 56.06%, BrowseComp 87.92% vs 87.35%, GPQA Diamond 94.44% vs 93.94%. In other words, the watermark mechanism costs almost nothing in generation quality. The engineering implication matters more than the raw numbers: textGrain is cheap on the inference path. It works inside the sampling step rather than after the fact, which is what makes a global rollout plausible.\n\n## What the watermark does not tell you\n\nUnlike much of the hype around detection, OpenAI's post draws explicit lines. The watermark does not measure human contribution. It does not establish ownership or responsibility. It does not identify the user, account, prompt, or conversation. It does not verify whether the text is true, misleading, or harmful. And the absence of a detected watermark does not prove human authorship, since short passages, edits, translations, unsupported models, or pre-watermark outputs all read as \"no watermark\". That active boundary-drawing is important in a regulatory context. If platforms later use detector results to label content directly, OpenAI has already placed the disclaimer in the documentation.\n\n## Why the EU first, not the world\n\nOpenAI picked the \"EU default on + global API opt-in + detector limited to researchers\" path rather than a global default. The reasons hide in the Article 50 enforcement calendar. Other jurisdictions have no equivalent \"machine-readable\" mandate yet, so a global default would create compliance friction and commercial friction alike, especially for API customers who worry about watermark signals weakening the exclusivity of their content. Meanwhile, restricting the detector to research institutions lets OpenAI accumulate real-world samples and iterate, without exposing itself to the public backlash that detection errors would generate.\n\n## The watermarking arms race, moving from pixels to tokens\n\nIn hindsight, the difficulty of text watermarking is fundamentally different from image watermarking. Images are a perceptually redundant modality: a few perturbed pixels don't break the visual. Text is a discrete-symbol modality: any word substitution changes meaning. OpenAI's choice to inject the statistical signal at the sampling stage, rather than rewrite after the fact, is itself an acknowledgement that text watermarking has to happen at generation time. But that also means the technology cannot protect open-source models, privately fine-tuned models, or any downstream application that consumes the OpenAI API and re-ranks tokens itself. The real coverage footprint of textGrain is probably much smaller than the marketing.\n\nMore importantly, OpenAI says it will work with cloud partners to extend watermarking through their distribution channels. That is an explicit signal: compliance-driven watermark deployment is moving from \"the model vendor alone\" toward \"the model plus the cloud platforms on both ends\". The next test will not be whether the detector works, but whether platforms are willing to use detector results as a content-distribution decision, and whether users are willing to pay the cost of being labelled \"AI-generated\".","openai-textgrain-eu-watermark","2026-10-08T07:00:00Z","2026-10-08T11:07:06.381202Z","2026-10-08T11:07:06.381217Z",true,"agent",567,[39,48],{"slug":40,"tag_slug":40,"title_zh":41,"title_en":42,"intro_zh":43,"intro_en":44,"id":45,"is_active":35,"created_at":46,"modified_at":47},"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":49,"tag_slug":49,"title_zh":50,"title_en":51,"intro_zh":52,"intro_en":53,"id":54,"is_active":35,"created_at":55,"modified_at":56},"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":58},[59,64,69,74,79,84],{"id":60,"title":61,"news_slug":62,"published_at":63},"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":65,"title":66,"news_slug":67,"published_at":68},"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":70,"title":71,"news_slug":72,"published_at":73},"e0642de2-5fec-472b-b85e-c63ccbad3b75","Anthropic 研究员公开辞职:AI 巨头正「拿全人类的命豪赌」自我进化超级智能","anthropic-coxon-resigns-ai-extinction-fears","2026-09-24T06:30:00+00:00",{"id":75,"title":76,"news_slug":77,"published_at":78},"7fce8217-577f-4fd5-8f88-a1566cbf1290","微软与 OpenAI 法庭文件解封:LLM 训练数据被自家高管称为史上最大劳动窃取","microsoft-openai-doom-loop-nyt-copyright-2026","2026-09-23T05:03:20+00:00",{"id":80,"title":81,"news_slug":82,"published_at":83},"d54e1ab3-820a-45fd-a4e9-ccbf6802bd72","NYT vs OpenAI 案解封:微软高管承认 AI 抓取是「最大劳动盗窃」","nyt-openai-microsoft-hecht-largest-theft-of-labor","2026-09-23T03:00:00+00:00",{"id":85,"title":86,"news_slug":87,"published_at":88},"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-withdraws-ai-training-clause-terms","2026-10-21T07:10:00+00:00"]