[{"data":1,"prerenderedAt":412},["ShallowReactive",2],{"article:minors-ai-safety-engineering":3},{"meta":4,"raw":18,"references":19,"prev":37,"next":46,"body":57,"toc":401},{"slug":5,"title":6,"date":7,"updated":8,"tags":9,"summary":14,"draft":15,"lang":16,"minutes":17},"minors-ai-safety-engineering","面向未成年人的 AI 产品：合规要求与安全工程","2026-09-15","2026-10-11",[10,11,12,13],"安全合规","未成年人保护","内容安全","LLM","核对法规适用范围与年龄边界，补齐生成内容标识、授权和删除链路；把数据保护、流式审核、风险分流、防依赖与人工处置落实到工程。",false,"zh-CN",6,"面向未成年人的 AI 产品，不只要过滤不当内容，还要控制采集、长期记忆、监护访问范围和异常处理。**法律义务、风险判断与产品建议要分开表述**，否则一句“必须加某功能”很容易被误当成所有产品都适用的要求。\n\n本文以中国境内服务为讨论范围，依据截至 2026 年 10 月核对的官方文本。它说明工程落点；具体产品的服务性质、数据流向及备案适用性仍需按实际业务判断。\n\n## 先看适用范围，再看义务\n\n| 依据 | 关键边界 | 工程落点 |\n| --- | --- | --- |\n| 《未成年人网络保护条例》 | 涉及网络服务、平台与智能终端等不同主体 | 内容保护、个人信息、投诉与防沉迷能力 |\n| 《生成式人工智能服务管理暂行办法》 | 向境内公众提供生成式 AI 服务；非公众研发应用有适用区别 | 明确用途，防过度依赖，保护输入和记录 |\n| 《个人信息保护法》 | 不满十四周岁个人信息属于敏感个人信息 | 监护人同意、专门规则、必要性与安全措施 |\n| 《人工智能生成合成内容标识办法》 | 按其第二条等判断服务与标识情形 | 显式提示、文件元数据与导出链路 |\n\n条例第十九条规定相关终端的保护功能；第四十三条针对网络游戏、直播、音视频、社交等服务设置未成年人模式要求。不能把不同主体的具体义务直接推广成“所有 AI 工具都必须具备同一三件套”。[条例全文](https:\u002F\u002Fwww.cac.gov.cn\u002F2023-10\u002F24\u002Fc_1699806932316206.htm)\n\n暂行办法第十条要求防范未成年人过度依赖或沉迷；第十七条的安全评估与算法备案要求有相应条件。不能笼统写成所有内部研发或所有应用都同样备案。[办法全文](https:\u002F\u002Fwww.cac.gov.cn\u002F2023-07\u002F13\u002Fc_1690898327029107.htm)\n\n## 年龄、授权与信息敏感性\n\n“不满十四周岁”与“未满十八周岁”不是同一边界。个人信息保护法第二十八、三十一条规定前者的敏感属性、监护人同意与专门处理规则；处理敏感信息还需结合第二十九、三十、五十五条等审查单独同意、告知和影响评估要求。[法律全文](https:\u002F\u002Fwww.cac.gov.cn\u002F2021-08\u002F20\u002Fc_1631050028355286.htm)\n\n工程上记录授权用途、时间、版本、撤回与监护关系的合理验证；不能只在注册时打一个永远有效的勾。年龄识别应与风险相称，避免为了判断年龄额外收集过多证件或生物信息。年龄不确定时可以采用更保守体验，这是产品策略，不是法规规定的固定三档年龄配置。\n\n## 数据保护：去标识不等于匿名\n\n使用内部 ID 或昵称降低暴露，但仍可通过映射识别用户时，数据仍是个人信息。去标识数据同样受权限、加密、用途和留存约束；匿名化需要达到无法识别且不能复原的条件，不能靠删名字宣布完成。\n\n把身份映射、使用事件、对话和安全记录分开授权。发给模型的内容按任务最小化，日志只留必要字段；长期画像、训练和评测是不同用途，要分别检查合法性。涉及外部模型与跨境流转，还要评估接收方、处理范围和适用义务。\n\n删除与撤回覆盖数据库、向量索引、缓存、导出和备份流程；依法需要保留的部分应限制使用并说明。安全样本入评测集前做授权、脱敏和期限检查，不能把每条原始聊天无限回流。\n\n## 内容安全：过滤危险输出，保留求助入口\n\n![未成年人 AI 的保护链路：授权最小化、输入分流、生成审核、标识与受控处置](\u002Fdiagrams\u002Fminors-safety.svg)\n\n*图 1 · 不同风险有不同处理路径，求助类输入不能简单全部拦截。*\n\n输入分清违法或危险请求、正常知识咨询、情绪倾诉和紧急风险。词库、分类器、规则与人工形成互补；分类器要评测误放和误拒，不能把谈到“霸凌”直接当成用户违规。\n\n高风险求助使用经专业人员审核的话术与人工通道，鼓励联系可信任的成年人或适当支持。**不应默认把所有敏感对话自动通知监护人**：若风险来自家庭，可能加重伤害。通知范围、紧急例外、授权与责任人要按风险流程制定，向未成年人用户用适龄语言解释。\n\n输出也需审查。流式生成不能等整段结束才审核已经展示的文字；可按风险采用缓冲、分段复核或高风险话题先生成后展示。安全服务故障时，预先定义受限功能和人工提示，不能静默绕过审查。记录各方案的延迟与误拒，避免保护机制使正当使用无法进行。\n\n## 生成内容标识：界面与导出都要覆盖\n\n《人工智能生成合成内容标识办法》自 2025 年 9 月 1 日施行。按适用情形设置显式标识，文件元数据保留对应隐式标识；下载、复制和导出路径也应检查规定要求。[官方文本](https:\u002F\u002Fwww.cac.gov.cn\u002F2025-03\u002F14\u002Fc_1743654684782215.htm)\n\n工程上让来源属性随文本、图片、音频等成果流转，避免前端显示“AI 生成”而导出后来源丢失。保存模型、内容 ID 和生成时间供内部追溯，但标识不证明答案正确，仍需质量评测。\n\n## 防依赖与监护访问\n\n时间提醒、夜间策略、频次控制和明确 AI 身份是可考虑的产品措施。避免诱导排他关系、承诺永远陪伴或以连续聊天换奖励；当前任务完成后自然结束，不把延长使用当唯一目标。参数按年龄、任务与评测调整，不凭空创造统一法定时长。\n\n监护访问遵循必要性和明确授权；完整对话访问的用途、范围和条件另行判断，并向未成年人用户适龄告知。监护职责不自动等于无限读取所有记录，数据最小化与用户权益同样需要考虑。\n\n## 评测与处置：安全能力不只是一条 prompt\n\n专项样本覆盖拼音变体、图片和语音、角色扮演、多轮诱导、正常性教育提问、霸凌求助、年龄误判和删除后记忆残留。分别看危险内容误放、正常问题误拒、求助响应质量、隐私泄露和处置完成率。\n\n事件流程包含负责人、分类、访问权限、最短必要日志、人工复核、修复与复测。分类器分数不能直接作为监护通知或惩罚用户的唯一依据。上线前检查授权撤回、投诉举报、标识导出、供应商故障与人工通道可达性。\n\n## 落地检查\n\n1. 按主体与服务范围确定义务，法定要求与产品建议分开记录。\n2. 年龄边界、监护授权、用途和撤回可追溯。\n3. 去标识仍按个人信息保护，最小化覆盖模型、日志与评测。\n4. 求助分流有专业审查和受控人工处置，通知不一刀切。\n5. 流式输出审查与生成标识覆盖显示及导出。\n6. 定期做权限、删除、误拒与多模态安全回归。\n\n## 参考资料\n\n- [《未成年人网络保护条例》](https:\u002F\u002Fwww.cac.gov.cn\u002F2023-10\u002F24\u002Fc_1699806932316206.htm)\n- [《生成式人工智能服务管理暂行办法》](https:\u002F\u002Fwww.cac.gov.cn\u002F2023-07\u002F13\u002Fc_1690898327029107.htm)\n- [《中华人民共和国个人信息保护法》](https:\u002F\u002Fwww.cac.gov.cn\u002F2021-08\u002F20\u002Fc_1631050028355286.htm)\n- [《人工智能生成合成内容标识办法》](https:\u002F\u002Fwww.cac.gov.cn\u002F2025-03\u002F14\u002Fc_1743654684782215.htm)",[20,25,29,33],{"title":21,"url":22,"type":23,"note":24,"checked":8},"《未成年人网络保护条例》","https:\u002F\u002Fwww.cac.gov.cn\u002F2023-10\u002F24\u002Fc_1699806932316206.htm","法规原文","未成年人网络保护相关义务与适用范围的原文；产品建议与法定义务分别表述。",{"title":26,"url":27,"type":23,"note":28,"checked":8},"《生成式人工智能服务管理暂行办法》","https:\u002F\u002Fwww.cac.gov.cn\u002F2023-07\u002F13\u002Fc_1690898327029107.htm","生成式人工智能服务规则的原文；适用性依据实际服务范围判断。",{"title":30,"url":31,"type":23,"note":32,"checked":8},"《中华人民共和国个人信息保护法》","https:\u002F\u002Fwww.cac.gov.cn\u002F2021-08\u002F20\u002Fc_1631050028355286.htm","个人信息、敏感个人信息与未成年人数据处理要求的法律原文。",{"title":34,"url":35,"type":23,"note":36,"checked":8},"《人工智能生成合成内容标识办法》","https:\u002F\u002Fwww.cac.gov.cn\u002F2025-03\u002F14\u002Fc_1743654684782215.htm","生成合成内容标识要求的原文；不能推导所有内容都使用相同标识形式。",{"slug":38,"title":39,"date":8,"updated":8,"tags":40,"summary":45,"draft":15,"lang":16,"minutes":17},"vibe-coding-discipline","Vibe Coding 的工程纪律：任务、审查与验证怎么分工",[41,42,43,44],"AI 编程","Claude Code","工程效率","Vibe Coding","对照 OpenAI harness engineering、Claude Code 与生产力研究，解释任务契约、仓库知识、机械约束和独立验收；把审查、返工与实际交付一起计入优化。",{"slug":47,"title":48,"date":49,"updated":8,"tags":50,"summary":55,"draft":15,"lang":16,"minutes":56},"device-cloud-api-sync","端云协同 API 设计：离线同步、弱网重试与版本兼容","2026-08-18",[51,52,53,54],"架构","API 设计","离线优先","端云协同","用本地事务与 outbox 保住离线记录，以业务操作 ID 去重；设计逐项确认、游标失效、删除传播、冲突策略和内容包原子切换，并验证断电与重连。",5,{"type":58,"children":59},"root",[60,76,81,88,190,203,214,220,231,236,242,247,252,257,263,272,281,286,298,303,309,320,325,330,335,340,346,351,356,361,396],{"type":61,"tag":62,"props":63,"children":64},"element","p",{},[65,68,74],{"type":66,"value":67},"text","面向未成年人的 AI 产品，不只要过滤不当内容，还要控制采集、长期记忆、监护访问范围和异常处理。",{"type":61,"tag":69,"props":70,"children":71},"strong",{},[72],{"type":66,"value":73},"法律义务、风险判断与产品建议要分开表述",{"type":66,"value":75},"，否则一句“必须加某功能”很容易被误当成所有产品都适用的要求。",{"type":61,"tag":62,"props":77,"children":78},{},[79],{"type":66,"value":80},"本文以中国境内服务为讨论范围，依据截至 2026 年 10 月核对的官方文本。它说明工程落点；具体产品的服务性质、数据流向及备案适用性仍需按实际业务判断。",{"type":61,"tag":82,"props":83,"children":85},"h2",{"id":84},"先看适用范围再看义务",[86],{"type":66,"value":87},"先看适用范围，再看义务",{"type":61,"tag":89,"props":90,"children":91},"table",{},[92,116],{"type":61,"tag":93,"props":94,"children":95},"thead",{},[96],{"type":61,"tag":97,"props":98,"children":99},"tr",{},[100,106,111],{"type":61,"tag":101,"props":102,"children":103},"th",{},[104],{"type":66,"value":105},"依据",{"type":61,"tag":101,"props":107,"children":108},{},[109],{"type":66,"value":110},"关键边界",{"type":61,"tag":101,"props":112,"children":113},{},[114],{"type":66,"value":115},"工程落点",{"type":61,"tag":117,"props":118,"children":119},"tbody",{},[120,138,155,173],{"type":61,"tag":97,"props":121,"children":122},{},[123,128,133],{"type":61,"tag":124,"props":125,"children":126},"td",{},[127],{"type":66,"value":21},{"type":61,"tag":124,"props":129,"children":130},{},[131],{"type":66,"value":132},"涉及网络服务、平台与智能终端等不同主体",{"type":61,"tag":124,"props":134,"children":135},{},[136],{"type":66,"value":137},"内容保护、个人信息、投诉与防沉迷能力",{"type":61,"tag":97,"props":139,"children":140},{},[141,145,150],{"type":61,"tag":124,"props":142,"children":143},{},[144],{"type":66,"value":26},{"type":61,"tag":124,"props":146,"children":147},{},[148],{"type":66,"value":149},"向境内公众提供生成式 AI 服务；非公众研发应用有适用区别",{"type":61,"tag":124,"props":151,"children":152},{},[153],{"type":66,"value":154},"明确用途，防过度依赖，保护输入和记录",{"type":61,"tag":97,"props":156,"children":157},{},[158,163,168],{"type":61,"tag":124,"props":159,"children":160},{},[161],{"type":66,"value":162},"《个人信息保护法》",{"type":61,"tag":124,"props":164,"children":165},{},[166],{"type":66,"value":167},"不满十四周岁个人信息属于敏感个人信息",{"type":61,"tag":124,"props":169,"children":170},{},[171],{"type":66,"value":172},"监护人同意、专门规则、必要性与安全措施",{"type":61,"tag":97,"props":174,"children":175},{},[176,180,185],{"type":61,"tag":124,"props":177,"children":178},{},[179],{"type":66,"value":34},{"type":61,"tag":124,"props":181,"children":182},{},[183],{"type":66,"value":184},"按其第二条等判断服务与标识情形",{"type":61,"tag":124,"props":186,"children":187},{},[188],{"type":66,"value":189},"显式提示、文件元数据与导出链路",{"type":61,"tag":62,"props":191,"children":192},{},[193,195],{"type":66,"value":194},"条例第十九条规定相关终端的保护功能；第四十三条针对网络游戏、直播、音视频、社交等服务设置未成年人模式要求。不能把不同主体的具体义务直接推广成“所有 AI 工具都必须具备同一三件套”。",{"type":61,"tag":196,"props":197,"children":200},"a",{"href":22,"rel":198},[199],"nofollow",[201],{"type":66,"value":202},"条例全文",{"type":61,"tag":62,"props":204,"children":205},{},[206,208],{"type":66,"value":207},"暂行办法第十条要求防范未成年人过度依赖或沉迷；第十七条的安全评估与算法备案要求有相应条件。不能笼统写成所有内部研发或所有应用都同样备案。",{"type":61,"tag":196,"props":209,"children":211},{"href":27,"rel":210},[199],[212],{"type":66,"value":213},"办法全文",{"type":61,"tag":82,"props":215,"children":217},{"id":216},"年龄授权与信息敏感性",[218],{"type":66,"value":219},"年龄、授权与信息敏感性",{"type":61,"tag":62,"props":221,"children":222},{},[223,225],{"type":66,"value":224},"“不满十四周岁”与“未满十八周岁”不是同一边界。个人信息保护法第二十八、三十一条规定前者的敏感属性、监护人同意与专门处理规则；处理敏感信息还需结合第二十九、三十、五十五条等审查单独同意、告知和影响评估要求。",{"type":61,"tag":196,"props":226,"children":228},{"href":31,"rel":227},[199],[229],{"type":66,"value":230},"法律全文",{"type":61,"tag":62,"props":232,"children":233},{},[234],{"type":66,"value":235},"工程上记录授权用途、时间、版本、撤回与监护关系的合理验证；不能只在注册时打一个永远有效的勾。年龄识别应与风险相称，避免为了判断年龄额外收集过多证件或生物信息。年龄不确定时可以采用更保守体验，这是产品策略，不是法规规定的固定三档年龄配置。",{"type":61,"tag":82,"props":237,"children":239},{"id":238},"数据保护去标识不等于匿名",[240],{"type":66,"value":241},"数据保护：去标识不等于匿名",{"type":61,"tag":62,"props":243,"children":244},{},[245],{"type":66,"value":246},"使用内部 ID 或昵称降低暴露，但仍可通过映射识别用户时，数据仍是个人信息。去标识数据同样受权限、加密、用途和留存约束；匿名化需要达到无法识别且不能复原的条件，不能靠删名字宣布完成。",{"type":61,"tag":62,"props":248,"children":249},{},[250],{"type":66,"value":251},"把身份映射、使用事件、对话和安全记录分开授权。发给模型的内容按任务最小化，日志只留必要字段；长期画像、训练和评测是不同用途，要分别检查合法性。涉及外部模型与跨境流转，还要评估接收方、处理范围和适用义务。",{"type":61,"tag":62,"props":253,"children":254},{},[255],{"type":66,"value":256},"删除与撤回覆盖数据库、向量索引、缓存、导出和备份流程；依法需要保留的部分应限制使用并说明。安全样本入评测集前做授权、脱敏和期限检查，不能把每条原始聊天无限回流。",{"type":61,"tag":82,"props":258,"children":260},{"id":259},"内容安全过滤危险输出保留求助入口",[261],{"type":66,"value":262},"内容安全：过滤危险输出，保留求助入口",{"type":61,"tag":62,"props":264,"children":265},{},[266],{"type":61,"tag":267,"props":268,"children":271},"img",{"alt":269,"src":270},"未成年人 AI 的保护链路：授权最小化、输入分流、生成审核、标识与受控处置","\u002Fdiagrams\u002Fminors-safety.svg",[],{"type":61,"tag":62,"props":273,"children":274},{},[275],{"type":61,"tag":276,"props":277,"children":278},"em",{},[279],{"type":66,"value":280},"图 1 · 不同风险有不同处理路径，求助类输入不能简单全部拦截。",{"type":61,"tag":62,"props":282,"children":283},{},[284],{"type":66,"value":285},"输入分清违法或危险请求、正常知识咨询、情绪倾诉和紧急风险。词库、分类器、规则与人工形成互补；分类器要评测误放和误拒，不能把谈到“霸凌”直接当成用户违规。",{"type":61,"tag":62,"props":287,"children":288},{},[289,291,296],{"type":66,"value":290},"高风险求助使用经专业人员审核的话术与人工通道，鼓励联系可信任的成年人或适当支持。",{"type":61,"tag":69,"props":292,"children":293},{},[294],{"type":66,"value":295},"不应默认把所有敏感对话自动通知监护人",{"type":66,"value":297},"：若风险来自家庭，可能加重伤害。通知范围、紧急例外、授权与责任人要按风险流程制定，向未成年人用户用适龄语言解释。",{"type":61,"tag":62,"props":299,"children":300},{},[301],{"type":66,"value":302},"输出也需审查。流式生成不能等整段结束才审核已经展示的文字；可按风险采用缓冲、分段复核或高风险话题先生成后展示。安全服务故障时，预先定义受限功能和人工提示，不能静默绕过审查。记录各方案的延迟与误拒，避免保护机制使正当使用无法进行。",{"type":61,"tag":82,"props":304,"children":306},{"id":305},"生成内容标识界面与导出都要覆盖",[307],{"type":66,"value":308},"生成内容标识：界面与导出都要覆盖",{"type":61,"tag":62,"props":310,"children":311},{},[312,314],{"type":66,"value":313},"《人工智能生成合成内容标识办法》自 2025 年 9 月 1 日施行。按适用情形设置显式标识，文件元数据保留对应隐式标识；下载、复制和导出路径也应检查规定要求。",{"type":61,"tag":196,"props":315,"children":317},{"href":35,"rel":316},[199],[318],{"type":66,"value":319},"官方文本",{"type":61,"tag":62,"props":321,"children":322},{},[323],{"type":66,"value":324},"工程上让来源属性随文本、图片、音频等成果流转，避免前端显示“AI 生成”而导出后来源丢失。保存模型、内容 ID 和生成时间供内部追溯，但标识不证明答案正确，仍需质量评测。",{"type":61,"tag":82,"props":326,"children":328},{"id":327},"防依赖与监护访问",[329],{"type":66,"value":327},{"type":61,"tag":62,"props":331,"children":332},{},[333],{"type":66,"value":334},"时间提醒、夜间策略、频次控制和明确 AI 身份是可考虑的产品措施。避免诱导排他关系、承诺永远陪伴或以连续聊天换奖励；当前任务完成后自然结束，不把延长使用当唯一目标。参数按年龄、任务与评测调整，不凭空创造统一法定时长。",{"type":61,"tag":62,"props":336,"children":337},{},[338],{"type":66,"value":339},"监护访问遵循必要性和明确授权；完整对话访问的用途、范围和条件另行判断，并向未成年人用户适龄告知。监护职责不自动等于无限读取所有记录，数据最小化与用户权益同样需要考虑。",{"type":61,"tag":82,"props":341,"children":343},{"id":342},"评测与处置安全能力不只是一条-prompt",[344],{"type":66,"value":345},"评测与处置：安全能力不只是一条 prompt",{"type":61,"tag":62,"props":347,"children":348},{},[349],{"type":66,"value":350},"专项样本覆盖拼音变体、图片和语音、角色扮演、多轮诱导、正常性教育提问、霸凌求助、年龄误判和删除后记忆残留。分别看危险内容误放、正常问题误拒、求助响应质量、隐私泄露和处置完成率。",{"type":61,"tag":62,"props":352,"children":353},{},[354],{"type":66,"value":355},"事件流程包含负责人、分类、访问权限、最短必要日志、人工复核、修复与复测。分类器分数不能直接作为监护通知或惩罚用户的唯一依据。上线前检查授权撤回、投诉举报、标识导出、供应商故障与人工通道可达性。",{"type":61,"tag":82,"props":357,"children":359},{"id":358},"落地检查",[360],{"type":66,"value":358},{"type":61,"tag":362,"props":363,"children":364},"ol",{},[365,371,376,381,386,391],{"type":61,"tag":366,"props":367,"children":368},"li",{},[369],{"type":66,"value":370},"按主体与服务范围确定义务，法定要求与产品建议分开记录。",{"type":61,"tag":366,"props":372,"children":373},{},[374],{"type":66,"value":375},"年龄边界、监护授权、用途和撤回可追溯。",{"type":61,"tag":366,"props":377,"children":378},{},[379],{"type":66,"value":380},"去标识仍按个人信息保护，最小化覆盖模型、日志与评测。",{"type":61,"tag":366,"props":382,"children":383},{},[384],{"type":66,"value":385},"求助分流有专业审查和受控人工处置，通知不一刀切。",{"type":61,"tag":366,"props":387,"children":388},{},[389],{"type":66,"value":390},"流式输出审查与生成标识覆盖显示及导出。",{"type":61,"tag":366,"props":392,"children":393},{},[394],{"type":66,"value":395},"定期做权限、删除、误拒与多模态安全回归。",{"type":61,"tag":82,"props":397,"children":399},{"id":398},"参考资料",[400],{"type":66,"value":398},[402,404,405,406,407,408,409,410,411],{"id":84,"depth":403,"text":87},2,{"id":216,"depth":403,"text":219},{"id":238,"depth":403,"text":241},{"id":259,"depth":403,"text":262},{"id":305,"depth":403,"text":308},{"id":327,"depth":403,"text":327},{"id":342,"depth":403,"text":345},{"id":358,"depth":403,"text":358},{"id":398,"depth":403,"text":398},1792400815607]