[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"article-openai-content-filtering-labeling-factory-en":3,"tags-openai-content-filtering-labeling-factory-en":30,"related-lang-openai-content-filtering-labeling-factory-en":38,"related-posts-openai-content-filtering-labeling-factory-en":42,"series-industry-ea07c233-f907-44b1-8fad-bb682295f775":79},{"id":4,"title":5,"content":6,"summary":7,"source":8,"source_url":9,"author":10,"image_url":11,"keywords":12,"language":18,"translated_content":10,"views":19,"is_premium":20,"created_at":21,"updated_at":21,"cover_image":11,"published_at":22,"rewrite_status":23,"rewrite_error":10,"rewritten_from_id":24,"slug":25,"category":26,"related_article_id":27,"status":28,"google_indexed_at":29,"x_posted_at":10,"tweet_text":10,"title_rewritten_at":10,"title_original":10,"key_takeaways":10,"topic_cluster_id":10,"embedding":10,"is_canonical_seed":20},"ea07c233-f907-44b1-8fad-bb682295f775","OpenAI内容过滤器背后的标注工厂","\u003Cp>2021年11月起，\u003Ca href=\"https:\u002F\u002Fopenai.com\" target=\"_blank\" rel=\"noopener\">OpenAI\u003C\u002Fa>把数万条文本片段发给肯尼亚外包公司进行标注，这些材料里有暴力、仇恨言论和性虐待内容。目标很直接：训练一个检测器，让它在用户看到之前先拦住类似内容。\u003C\u002Fp>\u003Cp>这件事很容易被阴谋论包裹，但真正值得看的不是“AI里是不是藏了谁的意识”，而是内容审核这门生意到底怎么运转。它依赖大量人工判断、脏数据清洗、模型分类器和产品层过滤，整个链条都很朴素，也很残酷。\u003C\u002Fp>\u003Ch2>这套系统到底在做什么\u003C\u002Fh2>\u003Cp>OpenAI这次做的，不是训练一个会聊天的模型，而是训练一个用于识别有害文本的检测器。简单说，就是先给一堆样本贴标签，再让模型学会分辨相似文本，最后把结果接进\u003Ca href=\"https:\u002F\u002Fchat.openai.com\" target=\"_blank\" rel=\"noopener\">ChatGPT\u003C\u002Fa>的内容过滤流程里。\u003C\u002Fp>\n\u003Cfigure class=\"my-6\">\u003Cimg src=\"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1775142603758-yydg.png\" alt=\"OpenAI内容过滤器背后的标注工厂\" class=\"rounded-xl w-full\" loading=\"lazy\" \u002F>\u003C\u002Ffigure>\n\u003Cp>这种做法在AI行业里很常见。大模型本身不会“理解”什么是有害内容，它只是从人工标注里学到统计模式。只要样本够多，模型就能对某些侮辱、骚扰、暴力、色情剥削类文本做出高召回率判断。\u003C\u002Fp>\u003Cp>这类系统通常会被放在两处：一处在生成前做输入侧检查，另一处在生成后做输出侧审核。前者拦截用户提示词，后者过滤模型回复。两层都上，误放行的概率才会下降。\u003C\u002Fp>\u003Cul>\u003Cli>训练目标：识别暴力、仇恨、性虐待等文本\u003C\u002Fli>\u003Cli>数据来源：数万条文本片段\u003C\u002Fli>\u003Cli>处理方式：人工标注后再训练分类器\u003C\u002Fli>\u003Cli>部署位置：ChatGPT内容过滤链路\u003C\u002Fli>\u003C\u002Ful>\u003Ch2>为什么偏偏要找外包人工标注\u003C\u002Fh2>\u003Cp>原因并不神秘：这类工作需要人眼判断，而且要有人能接受长时间接触恶心内容。机器可以做筛选，但第一批标签往往还是得靠人来定。\u003C\u002Fp>\u003Cp>肯尼亚外包公司参与这类工作，说明AI产业链早就全球化了。训练数据、标注劳动力、审核流程，分别分布在不同国家。用户在美国、欧洲或亚洲看到的一个“安全”功能，背后可能是一群远程标注员在逐条看极端文本。\u003C\u002Fp>\u003Cp>这也解释了为什么很多AI公司会强调“安全”与“对齐”。这些词听上去抽象，落到执行层面，就是把大量脏活拆成标准化任务，再交给标注团队和审核系统处理。\u003C\u002Fp>\u003Cblockquote>“The internet is the first thing that humanity has built that humanity doesn’t understand, the largest experiment in anarchy that we have ever had.” — Eric Schmidt\u003C\u002Fblockquote>\u003Cp>这句话虽然不是专门谈内容审核，却很适合这里。互联网内容太多、太杂、太快，任何想做过滤的公司都得面对同一个现实：先把混乱变成可分类的数据，再谈规则。\u003C\u002Fp>\u003Ch2>和其他内容审核方案比，差别在哪\u003C\u002Fh2>\u003Cp>OpenAI这类做法的重点，是把人工经验转成可复用的分类器。和纯人工审核比，它的速度更快；和纯规则过滤比，它更能识别变体写法、拼写变形和语义绕过。\u003C\u002Fp>\n\u003Cfigure class=\"my-6\">\u003Cimg src=\"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1775142601170-i2t5.png\" alt=\"OpenAI内容过滤器背后的标注工厂\" class=\"rounded-xl w-full\" loading=\"lazy\" \u002F>\u003C\u002Ffigure>\n\u003Cp>但代价也明显。分类器会误杀正常内容，也会漏掉新型规避表达。尤其是涉及政治隐喻、黑话、俚语时，模型常常比人更笨。为了减少误伤，产品团队通常得不断回收样本、重新标注、再训练。\u003C\u002Fp>\u003Cp>如果把它和常见的审核路径放在一起看，差异会更清楚：\u003C\u002Fp>\u003Cul>\u003Cli>纯人工审核：准确率高，但慢，成本也高\u003C\u002Fli>\u003Cli>关键词规则：便宜，速度快，绕过也最容易\u003C\u002Fli>\u003Cli>机器分类器：覆盖面广，能处理变体，但需要持续迭代\u003C\u002Fli>\u003Cli>多层混合方案：最常见，成本和效果最平衡\u003C\u002Fli>\u003C\u002Ful>\u003Cp>从工程角度看，OpenAI这类系统并不神秘。真正难的是把它做得足够稳定，同时别把正常用户体验弄坏。审核太松，平台会被垃圾内容淹没；审核太严，用户会觉得模型像个动不动就罢工的保守派。\u003C\u002Fp>\u003Ch2>为什么阴谋论总会缠上AI\u003C\u002Fh2>\u003Cp>AI很容易被神秘化，因为大多数人看不到训练过程，只能看到最终输出。输入、标注、清洗、微调这些环节都藏在后台，外界只看见一个会说话的接口，于是很自然地开始脑补“它到底吃了什么”。\u003C\u002Fp>\u003Cp>但从这条新闻本身看，最重要的信息其实很普通：OpenAI在做内容过滤训练，而且用了人工标注。这个流程说明的是工业化审核，不是超自然秘密。\u003C\u002Fp>\u003Cp>真正值得警惕的，是人们对AI黑箱的误解会被反复利用。有人拿它编故事，有人拿它制造恐慌，还有人借机把正常的工程问题说成阴谋。结果是，大家讨论的重点被带偏，真正该问的问题反而没人问：这些标注员的工作条件怎么样，数据处理合规吗，过滤器误伤率有多高。\u003C\u002Fp>\u003Cp>如果你关心的是产品安全，那么更应该盯住两个指标：误报率和漏报率。前者决定用户会不会被过度拦截，后者决定平台会不会放出真正危险的内容。AI审核不是玄学，就是一场持续调参的工程活。\u003C\u002Fp>\u003Ch2>结论：别被神秘叙事带跑\u003C\u002Fh2>\u003Cp>把“失踪人口意识”这类说法放到这条新闻里，基本属于把普通的数据标注工作往神秘主义方向硬拽。更合理的解释很无聊，也更接近现实：OpenAI在用人工标注训练内容过滤器，目的就是让ChatGPT更少输出危险文本。\u003C\u002Fp>\u003Cp>接下来更值得关注的，不是这些文本“像不像某种秘密材料”，而是这类审核系统会不会继续扩大到更多产品、更多语言和更多地区。如果未来你发现模型越来越谨慎，背后多半不是“意识被抽出来了”，而是标注、过滤和审核这三件事又被加码了一轮。\u003C\u002Fp>","OpenAI把数万条有害文本送去人工标注，用来训练ChatGPT过滤器。它为什么要这样做？","www.zhihu.com","https:\u002F\u002Fwww.zhihu.com\u002Fquestion\u002F2022623696783774161\u002Fanswer\u002F2022632267613312315",null,"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1775142603758-yydg.png",[13,14,15,16,17],"OpenAI","ChatGPT","内容审核","数据标注","文本过滤","en",1,false,"2026-04-02T15:09:36.871742+00:00","2026-04-02T15:09:36.526+00:00","done","e7a88f53-bfc7-4756-afb1-ad94b8c53878","openai-content-filtering-labeling-factory-en","industry","8b08524b-22a3-4f8e-8376-feacb8fdf2a5","published","2026-04-08T09:00:51.527+00:00",[31,33,35,36,37],{"name":13,"slug":32},"openai",{"name":14,"slug":34},"chatgpt",{"name":15,"slug":15},{"name":17,"slug":17},{"name":16,"slug":16},{"id":27,"slug":39,"title":40,"language":41},"openai-content-filtering-labeling-factory-zh","OpenAI內容過濾器的標註工廠","zh",[43,49,55,61,67,73],{"id":44,"slug":45,"title":46,"cover_image":47,"image_url":47,"created_at":48,"category":26},"6ff3920d-c8ea-4cf3-8543-9cf9efc3fe36","circles-agent-stack-targets-machine-speed-payments-en","Circle’s Agent Stack targets machine-speed payments","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1778871659638-hur1.png","2026-05-15T19:00:44.756112+00:00",{"id":50,"slug":51,"title":52,"cover_image":53,"image_url":53,"created_at":54,"category":26},"1270e2f4-6f3b-4772-9075-87c54b07a8d1","iren-signs-nvidia-ai-infrastructure-pact-en","IREN signs Nvidia AI infrastructure pact","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1778871059665-3vhi.png","2026-05-15T18:50:38.162691+00:00",{"id":56,"slug":57,"title":58,"cover_image":59,"image_url":59,"created_at":60,"category":26},"b308c85e-ee9c-4de6-b702-dfad6d8da36f","circle-agent-stack-ai-payments-en","Circle launches Agent Stack for AI payments","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1778870450891-zv1j.png","2026-05-15T18:40:31.462625+00:00",{"id":62,"slug":63,"title":64,"cover_image":65,"image_url":65,"created_at":66,"category":26},"f7028083-46ba-493b-a3db-dd6616a8c21f","why-nebius-ai-pivot-is-more-real-than-hype-en","Why Nebius’s AI Pivot Is More Real Than Hype","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1778823055711-tbfv.png","2026-05-15T05:30:26.829489+00:00",{"id":68,"slug":69,"title":70,"cover_image":71,"image_url":71,"created_at":72,"category":26},"b63692ed-db6a-4dbd-b771-e1babdc94af7","nvidia-backs-corning-factories-with-billions-en","Nvidia backs Corning factories with billions","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1778822444685-tvx6.png","2026-05-15T05:20:28.914908+00:00",{"id":74,"slug":75,"title":76,"cover_image":77,"image_url":77,"created_at":78,"category":26},"26ab4480-2476-4ec7-b43a-5d46def6487e","why-anthropic-gates-foundation-ai-public-goods-en","Why Anthropic and the Gates Foundation should fund AI public goods","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1778796645685-wbw0.png","2026-05-14T22:10:22.60302+00:00",[80,85,90,95,100,105,110,115,120,125],{"id":81,"slug":82,"title":83,"created_at":84},"d35a1bd9-e709-412e-a2df-392df1dc572a","ai-impact-2026-developments-market-en","AI's Impact in 2026: Key Developments and Market Shifts","2026-03-25T16:20:33.205823+00:00",{"id":86,"slug":87,"title":88,"created_at":89},"5ed27921-5fd6-492e-8c59-78393bf37710","trumps-ai-legislative-framework-en","Trump's AI Legislative Framework: What's Inside?","2026-03-25T16:22:20.005325+00:00",{"id":91,"slug":92,"title":93,"created_at":94},"e454a642-f03c-4794-b185-5f651aebbaca","nvidia-gtc-2026-key-highlights-innovations-en","NVIDIA GTC 2026: Key Highlights and Innovations","2026-03-25T16:22:47.882615+00:00",{"id":96,"slug":97,"title":98,"created_at":99},"0ebb5b16-774a-4922-945d-5f2ce1df5a6d","claude-usage-diversifies-learning-curves-en","Claude Usage Diversifies, Learning Curves Emerge","2026-03-25T16:25:50.770376+00:00",{"id":101,"slug":102,"title":103,"created_at":104},"69934e86-2fc5-4280-8223-7b917a48ace8","openclaw-ai-commoditization-concerns-en","OpenClaw's Rise Raises Concerns of AI Model Commoditization","2026-03-25T16:26:30.582047+00:00",{"id":106,"slug":107,"title":108,"created_at":109},"b4b2575b-2ac8-46b2-b90e-ab1d7c060797","google-gemini-ai-rollout-2026-en","Google's Gemini AI Rollout Extended to 2026","2026-03-25T16:28:14.808842+00:00",{"id":111,"slug":112,"title":113,"created_at":114},"6e18bc65-42ae-4ad0-b564-67d7f66b979e","meta-llama4-fabricated-results-scandal-en","Meta's Llama 4 Scandal: Fabricated AI Test Results Unveiled","2026-03-25T16:29:15.482836+00:00",{"id":116,"slug":117,"title":118,"created_at":119},"bf888e9d-08be-4f47-996c-7b24b5ab3500","accenture-mistral-ai-deployment-en","Accenture and Mistral AI Team Up for AI Deployment","2026-03-25T16:31:01.894655+00:00",{"id":121,"slug":122,"title":123,"created_at":124},"5382b536-fad2-49c6-ac85-9eb2bae49f35","mistral-ai-high-stakes-2026-en","Mistral AI: Facing High Stakes in 2026","2026-03-25T16:31:39.941974+00:00",{"id":126,"slug":127,"title":128,"created_at":129},"9da3d2d6-b669-4971-ba1d-17fdb3548ed5","cursors-meteoric-rise-pressures-en","Cursor's Meteoric Rise Faces Industry Pressures","2026-03-25T16:32:21.899217+00:00"]