半透膜时代：信息过载、AI冲击与媒体、学术出版与机构研究的三层重构

一、引子：三层结构与半透膜

过去三十年互联网被讲成一个自由与开放的故事：谷歌把公开网页的流量分发做成广告帝国，普通读者在"搜到、点开、被广告变现"的循环里完成信息消费。三十年后的今天，信息过载、生成式人工智能与信任危机同时发作，这套单一循环正在裂解为三层结构。位于外侧的，是仍然对搜索与AI可发现、但直接变现能力剧烈下降的开放发现层；位于中间的，是以注册、会员与订阅换取深度关系的注册会员层；位于内核的，是以可审计的方法、可追责的人类署名与定制服务承接决策的专业机构层。

真正贴切的隐喻不是"高墙"，而是"半透膜"。半透膜允许特定分子通过、拦截其他分子。对应到内容产业，就是允许摘要、元数据与可授权采样通过，让机构与作品仍能被AI、搜索与专业读者发现，同时把全文、原始数据与身份化关系数据留在墙内、免于被无偿抓取。绝对封闭会退出发现市场，绝对开放会失去议价能力。问题从来不是筑不筑墙，而是把哪些分子放进来、哪些留在墙内，并对进出双向计价与留痕。

二、信任崩解与注意力迁移：可核验的宏观图景

信任危机是这场重构的第一推动力，且不是主观印象而是可核验的下滑。路透新闻研究所的《2026年数字新闻报告》覆盖48个市场约10万人，全球对新闻的信任度降至37%，为2015年首次测量以来的最低值，48个市场中29个显著下降；每周回避新闻的人群维持在42%，2017年只有29%，同一样本中"随意/被动新闻用户"从2021年的16%升至25%，"新闻爱好者"从29%降至22%（Reuters Institute DNR 2026）。美国的下沉更陡：盖洛普2025年9月调查显示"完全或较多信任大众媒体"的比例首次跌破三成、只剩28%，共和党人群仅剩8%（Gallup）；皮尤2025年10月发布的数据显示对全国性新闻机构的信任在半年内下降11个百分点、只剩56%，但对本地新闻机构的信任仍有70%------信任的坍缩是有梯度的，本地、专业与人格化来源相对抗跌（Pew Research）。爱德曼2025年信任度调查在28国33,000多人中发现，平均69%的受访者担心政府官员、商界领袖与记者"故意误导"他们，63%越来越难判断新闻是可信来源还是刻意欺骗（Edelman 2025）。

发现入口本身也在迁移。中国互联网络信息中心2026年3月发布的第57次报告显示网民规模11.25亿、手机上网比例99.6%，生成式AI用户一年内增长141.7%至6.02亿，而搜索引擎使用率从2024年12月的79.2%降至2025年12月的69.5%（CNNIC、报告PDF）。西方侧同样的方向：路透报告记录到每周用AI聊天机器人获取新闻的比例从7%升至10%，其中最年轻组已达17%（DNR 2026 AI章）。两条数据合在一起意味着"公开搜索+网页浏览"的黄金入口正被"社交/短视频/AI回答"三合一取代------这不是崩溃也不是替代，而是发现权在重新分配。

三、广告：不是崩盘，而是位移与测量失灵

广告业当下的真实处境不是"回报处于历史最低"，而是结构性迁移与测量失灵同时出现。IAB与普华永道《2025年全年互联网广告收入报告》显示美国数字广告收入2025年达2,946亿美元、同比+13.9%，程序化+20.5%至1,624亿美元，创作者广告370亿美元、2026年预计升至440亿美元（IAB/PwC）。线性电视则加速让位：eMarketer援引WARC口径，2026年线性电视广告预计下降超11%至1,391亿美元，占全球媒体支出份额从2013年41.3%降至12.4%，但仍占全球电视投放的75%（eMarketer）。真正的痛点是测量与效率：尼尔森2025年营销报告调查全球1,400名营销人后发现，只有32%的营销人对传统与数字媒体做整体性测量，54%计划减少广告支出（Nielsen）；美国全国广告主协会研究更直接：880亿美元的开放网程序化生态中，多达200亿美元被浪费，占比23%（ANA）。

信任并不因此就简单转向"熟人和创作者"。BBB国家项目《影响者信任指数》2025年报告显示，87%的美国消费者信任他们所看到的广告（其中5.5%完全信任、83.2%有些信任），信任网红广告的比例为74%，26%不信任------而70%的消费者会因为未披露的商业合作而对创作者产生负面观感（BBB National Programs）。可以站住的判断是：注意力与预算正在向创作者、社群、零售媒体与视频迁移，而未披露的商业关系会立刻毁掉这一层的溢价。

推荐经济是这一图景中最容易被误读的一部分。金融业并没有"放弃广告"，而是在广告之外叠加可计量的获客渠道。可核验的官方条款有：大通银行的Freedom卡每次推荐50美元、每年上限500美元，Sapphire每次15,000积分（年上限10万分），Marriott Bonvoy每次40,000积分（Chase）；第一资本官方教育页说明其信用卡推荐奖励通常每日历年上限500美元（Capital One）；美国运通规定每卡账户每日历年最多5次推荐奖励（American Express）；SoFi单一日历年内所有存款奖励累计不超过10,000美元、单次推荐最高125美元（SoFi）；罗宾汉券商采用阶梯设计、每年最高15,000美元股票奖励（Robinhood）。真正的信号并不是"广告死了"，而是当广告可测性下降时，机构愿意为可归因、带身份的人际引流付出可核算的赏金。这与订阅、会员、社群一样，都属于第二层的经济学。

四、中美两种"围墙"：闭环变现与合规叠加

中美的"围墙"看似都是把内容锁在墙内，但建造逻辑并不相同。中国移动生态的墙首先是商业设计，不是法律强制。腾讯2025年报把广告涨价的成因之一直接表述为"闭环广告的占比持续提升(用户点击后可直达小程序、微信小店或小游戏等原生交易场景)"，微信合并月活14.18亿（腾讯2025年度业绩）。这套设计的本质是在超级App内部完成"发现---转化---支付---履约"闭环，链外的开放网页因此不再是必要环节。同期监管的方向恰恰是"拆墙"而不是"筑墙"：2021年9月工信部召开"屏蔽网址链接问题行政指导会"，提出即时通信软件合规标准，要求企业依限解除屏蔽，主管官员在国新办发布会上明确"互联互通是互联网行业高质量发展的必然选择"，同一报道亦援引专家意见指出"生态开放不能一蹴而就......不宜一刀切地解除壁垒"，主张分步实施并保留"必要的防火墙"（新华网）。这段官方与半官方表述本身，就是"半透膜策略"在中国监管语境下的现成注脚。

实名与账号治理是中国规则里最鲜明的一层，但严格的落点是注册与发布，不是"浏览网页非法"。新修订的《网络安全法》2026年1月1日起施行，第二十六条要求网络接入、域名注册、入网、信息发布与即时通讯服务提供真实身份信息（网络安全法）；《互联网用户账号信息管理规定》第十一条要求新闻、经济、教育、医疗卫生、司法等领域账号提交资质并"加注专门标识"，第十二至十三条要求展示IP归属地与主体信息（网信办）；《互联网信息服务算法推荐管理规定》第十七条要求"向用户提供不针对其个人特征的选项"并可关闭算法推荐，第二十一条禁止"大数据杀熟"式差别定价，第二十四条要求具有舆论属性或社会动员能力的服务须完成算法备案（算法推荐规定）；《生成式人工智能服务管理暂行办法》第七条要求训练数据"合法来源"并不侵害他人知识产权（生成式AI暂行办法）；《人工智能生成合成内容标识办法》2025年9月1日施行，明确区分显式与隐式标识，要求元数据中写入服务提供者与内容编号、日志留存不少于六个月（标识办法）。合并起来，中国这一层是"账号资质核验+IP归属地展示+算法备案+生成内容标识"的合规叠加，而不是"匿名浏览违法"。

美国与欧盟侧的"墙"则以私人版权、访问控制与反垄断为主，法律并未简单禁止AI抓取。版权诉讼路径分歧仍在：Bartz诉Anthropic案中Alsup法官2025年6月认定合法购书训练可属转换性合理使用，但保存约700万本盗版书构成侵权，该案最终以15亿美元和解，2026年7月20日获终局批准（Reuters 2025-06-24、Reuters 2026-07-20）；《纽约时报》诉OpenAI/Microsoft在2026年6月追加了对微软的帮助侵权指控（Ars Technica）；OpenAI官方案件页显示部分主张已被驳回，核心争点仍是合理使用（OpenAI）。反垄断路径由Penske诉Google打开，2025年9月大型出版商首次就AI Overviews起诉搜索巨头，核心指控涉及市场支配地位与流量截流（Reuters 2025-09-14）。合同许可路径已出现范式，News Corp与OpenAI 2024年5月签署多年期全球授权（News Corp）。

访问控制层是"能挡多少"的现实检验。Cloudflare 2025年7月起对新域名默认阻断AI爬虫并推出"按次抓取付费"实验，声称已有超250万个网站通过其托管规则完全禁止AI训练用途抓取（Cloudflare 新闻稿）；但同一家公司随后指控Perplexity使用"隐形、未申报"爬虫，当被网络层阻断时"似乎会掩盖其抓取身份"，反复更换user agent与ASN、忽略甚至不去获取robots.txt，Cloudflare因此取消其"已验证机器人"资格（Cloudflare 博客）。授权语义层的新工具是RSL 1.0（Really Simple Licensing），1,500多家机构在2025年12月支持这一机器可读许可与计费标准（RSL）。真实图景是：robots.txt是自愿协议（RFC 9309）而不是法律禁令，访问控制在技术上永远能被局部规避，只有把技术+合同+法律三条线叠起来，才是可防守的边界。

一句话对比：中国的"墙"是平台级+账号级+发布环节实名+算法备案+生成内容强制标识的行政合规叠加；美国与欧盟的"墙"是平台级+网络层默认拒爬+机器可读许可+版权和反垄断诉讼的私人权利叠加。两者在保护点（版权、身份、公共秩序）不同，但在部署方式上都走向"结构化的半透膜"，而非绝对封闭。

五、AI作为过滤层：能力、误差与合规

AI进入公域信息消费的方式并不是"取代记者"，而是接管过滤与策展。这套能力的底层不是"深度生物识别"，而是检索增强生成（RAG）+ 显式反馈 + 第一方会员数据。Lewis等人2020年提出RAG的原论文明确指出，纯参数化模型的两大缺陷是"为其决策提供出处（provenance）"与"更新其世界知识"，RAG通过外接非参数化记忆使生成"更具体、更多样、更符合事实"（arXiv:2005.11401）。这条技术路线的商业含义是：模型天然需要外部、结构化、可授权的语料，因此可信私域并不是与AI对立，而是AI的必需品。

但出处能力≠出处正确。哥伦比亚新闻评论Tow Center 2025年对8个生成式搜索工具、20家出版商、1,600次查询的测试显示整体错误率超过60%，其中Perplexity 37%、Grok 3高达94%（CJR / Tow Center）。欧洲广播联盟与BBC领导的《AI助手中的新闻完整性》2025年10月研究覆盖22家公共媒体、18国、14语言、3,000多条回答：45%的AI回答至少存在一项重大问题、31%存在严重来源问题、20%存在重大准确性问题，Gemini问题率高达76%（BBC Media Centre）。这为可信私域留下了明确的商业空间------AI摘要越普及，可验证署名的机构与作者就越有议价能力。

关于个性化，正确的表述是"依赖行为与偏好信号、显式反馈与第一方会员数据"，而不是"深度生物识别"。欧盟《通用数据保护条例》第9(1)条原则性禁止以唯一识别自然人为目的处理生物识别数据（GDPR）；欧盟《人工智能法》第5条直接禁止工作场所与教育机构的情绪推断、基于生物特征推断种族/政治/宗教/性取向的分类系统以及无目标抓取面部图像建库（EU AI Act）。生物识别在这些辖区是合规负债而非产品前提。

法律层还锁定了一个至关重要的"退出键"：欧盟《数字服务法》第38条要求超大型平台和搜索引擎至少提供一个不基于GDPR第4(4)条画像的推荐选项（DSA 38），中国《算法推荐规定》第十七条同样要求"不针对个人特征的选项"或便捷关闭功能（网信办）。可以在两个法域找到"非画像推荐"作为法定选项，这本身就否证了"AI必须监听一切、必须做深度画像才能工作"的错觉。

关于AI是否直接制造极化，可靠证据也是有限的。路透研究所2022年文献综述结论是"回音室远没有通常假设的那样普遍，未发现支持过滤气泡假说的证据"（ORA）；发表在《科学》的Facebook实验显示，把用户移出算法信息流会大幅减少其平台使用时间并改变内容暴露结构，但在三个月内并未显著改变议题极化、情感极化或政治知识（Science abp9364）。可靠的判断是：算法主要改变暴露结构与停留时长；对态度极化的因果证据薄弱；AI过滤层真正的风险是"发现权"与"来源可见性"的再分配，而不是简单的"洗脑"。

内容溯源的技术底座正在成型。C2PA 2.2的Content Credentials用密码学方式绑定并记录资产的来源、修改历史与AI使用情况，规范明确设计为"全局、可选择加入（opt-in）"，并强调"添加来源信息是可选的，无意造成双层媒体生态"（C2PA Explainer）。中国路径上，来源标注已由标识办法强制化：文本、音频、图片、视频与虚拟场景须加显式标识；元数据中须写入服务提供者与内容编号；用户申请无显式标识内容时，服务方须留存日志不少于六个月（标识办法）。当AI回答越来越多，署名、修改历史与许可条件被写进内容本体，就成为"可信"这一属性的最小基础。

六、可信私域的商业逻辑：从报刊亭到平台化的会员经济

在信任下降与AI质量缺口的双重压力下，可信私域正在被现金投票。Press Gazette的《10万俱乐部》2026年榜单显示，62家英语新闻/杂志出版商各有10万以上数字订阅，合计5,400万数字订阅，高于2025年2月的4,470万，同比增长约21%；这份榜单一家《纽约时报》就占23%（Press Gazette）。《纽约时报》2026年二季报显示纯数字订阅约1,280万，数字ARPU从上年同期9.64美元升至9.94美元，五个季度稳步上行，提价与套餐化都在生效（NYT Q2 2026 SEC）。

平台侧的转化机制同样有数据支撑。Patreon的CEO口径显示，自2013年以来创作者累计收款超过100亿美元，2,500万以上付费会员，超过1亿免费会员每月有超70万转化为付费（Axios）；平台费官方口径为2025年8月4日后新创作者适用标准10%（Patreon）。Substack宣布累计500万付费订阅，官方分成为"作者保留90%再扣信用卡手续费"，官方示例是"1,000名订阅者×每月5美元=年收入6万美元"（Substack、Substack Going Paid）。这些数据一起说明**"免费层→注册层→付费层"的漏斗真实存在且可量化**------但也提醒创作者：如果全部依赖平台，就意味着长期让渡10%的现金和100%的用户关系；自有渠道之所以值得建，正是因为在这道半透膜里，会员关系是唯一不会被平台改规则回收的资产。

"报刊亭式订阅"抓到了正确的直觉，但它并不代表"数字回到印刷"，而代表"从广告经济回到关系经济"。读者付费购买的不是"稀缺信息"（AI已让稀缺信息变得便宜），而是编辑判断、来源可追、编辑与作者的可问责性以及不被广告噪声干扰的阅读环境。会员层的定价能力，本质上是"信任---方法---署名"的复合价格。

七、学术出版与机构研究：不是崩溃，是门槛与责任的资本化

学术侧真正在发生的是准入门槛提高、筛查成本资本化、责任与披露制度化，而不是"普通大学和期刊大面积倒闭"。arXiv自2025年10月31日起要求计算机科学类的综述与立场论文必须先通过期刊或会议同行评议并提交证明才会获完整考虑，官方理由是"综述与立场论文涌入量已无法管理"，且"生成式AI/大语言模型让不产生新研究结果的论文写起来又快又容易"（arXiv 官方博客）。国际科技医学出版商协会（STM）2026年1月发布的报告指出，部分出版商已设立超过100人的专职科研诚信团队，每年筛查数百万份投稿，"运用技术但让人始终处于工作的中心"；协作基础设施同时扩张：STM Integrity Hub 49家机构成员、COPE 106家出版商代表14,500多种期刊、United2Act反论文工厂联盟58家机构（STM）。据Frontiers对111国约1,600名学者的调查，超过50%的研究者在评审稿件时使用过AI（Nature 报道）。

责任规则则被明确锁定在人身上。国际医学期刊编辑委员会规定聊天机器人与其他AI辅助工具不应被列为作者，因其无法对准确性、完整性与原创性负责，作者须披露AI使用并说明如何使用，且"以AI生成材料作为主要来源引用是不可接受的"（ICMJE）；出版伦理委员会2024年8月版立场声明也明确"AI工具不能作为论文作者"，因其"无法对提交的作品负责"、无法声明利益冲突、无法管理版权与许可协议，作者对稿件内容承担全部责任，即便是AI工具生成的部分（COPE）。

对学术出版与机构研究，能生存并涨价的，是把"来源可追溯性、方法透明、数据与代码可得、评议流程可审计、人类署名责任"打包成产品的机构；被冲击的，是把"批量数据整理+格式化解读"当作核心供给的中低端环节。机构研究的转型方向不是继续量产PDF，而是走向可追责的顾问服务、决策级简报、机构订阅数据授权与训练语料许可。这也是"三层模型"在机构世界的对应表达。

八、三层商业模型：可以照做的操作蓝图

层一：开放发现层------把"被发现"当成一项产品

发现层的核心是让机构与作品被AI、被搜索、被跨平台的编辑网络看到，被引用、被摘要、被链接。它的直接货币收入应尽可能低估：程序化开放网生态自身的浪费已达23%（ANA），Google的AI Overviews正在被大型出版商挑战（Reuters），AI回答本身的严重来源问题率仍在31%（BBC/EBU）。

发现层应做而且只做四件事：一是摘要与元数据开放------把标题、导语、结构化摘要、作者与许可条件公开；二是授权语义机器化------用robots.txt（RFC 9309）+ RSL 1.0机器可读许可 + Cloudflare按次抓取付费声明"可否抓、以何价、按何用途"（Cloudflare、RSL 1.0）；三是内容凭证随行------把作者、机构、修改历史、AI使用情况通过C2PA写进内容本体（C2PA）；四是KPI从"广告收入/CPM"迁移到"被引用次数×署名保留率×授权覆盖率"。这层不是不赚钱，而是主要以授权与配额而不是广告位作为收入形式。News Corp与OpenAI 2024年5月的多年期授权已给出参考模板（News Corp）。

层二：注册与会员层------把"关系"当成资产

这一层的产品化包括四步：注册（免费/低门槛）→ 会员（社区+活动）→ 付费订阅（内容+归档+讨论）→ 家庭/学生/团体套餐。可核验的定价基准是月费9到20美元、年费100到200美元区间，机构套餐单价可翻倍。银行推荐奖金证明"可归因人际引流"的市场价格：每次成功付费转化，愿意支付25到200美元，是被主流机构在真金白银里认可过的区间（Chase、SoFi、Capital One）。可执行的会员经济学证据来自Patreon与Substack（转化漏斗与90/10分成）以及英语新闻订阅榜的一年+21%（Axios、Substack、Substack Going Paid、Press Gazette）。

合规约束不能省略：必须提供非画像推荐选项（DSA 38、算法推荐规定）；不得以生物识别作为个性化前提（GDPR、AI Act）；生成内容需按当地法规做溯源标识（中国标识办法、C2PA）。这一层的最大风险是回避与疲劳（42%的人回避新闻，25%是被动用户，DNR 2026），因此编辑克制、结构化摘要、可控通知与低频精编产品比"更多推送"更重要。

层三：专业与机构层------把"可问责性"当成价格

第三层的关键词不是"高端"，而是可审计、可追责、可被机构采购。它承接三类需求：一是机构订阅------图书馆、企业与政府愿意为"可被审计的研究流程与数据集"付费；二是定制决策支持------面向具体业务问题的顾问式产品，包括战情简报、评估、模型审计与培训；三是授权与训练语料------把可核验的语料按用途授权给AI提供商，与News Corp、RSL、Pay Per Crawl一线的机制打通。这层的护城河是"人---方法---声誉"的组合：责任规则要求人类署名并承担全部责任（ICMJE、COPE），科研诚信团队已成资本密集能力（STM），中国侧的账号资质核验与专门标识是这类"专业资质"制度化的对照（账号信息管理规定第11条）。

半透膜：三层之间如何调度

三层之间要靠一张"半透膜"连接：摘要与元数据向外免费流出以维持发现权；全文、原始数据与身份化关系数据留在墙内以维持议价权；训练语料通过机器可读许可与配额向AI提供商定价流出，形成可核算的第三方收入。绝对高墙会失去发现，绝对开放会失去定价；能长期赢的机构，是把半透膜的"孔径"当成核心产品参数、随着信任经济与授权市场演化持续微调的机构。

结语：从宏大隐喻走向可运营的半透膜

这场重构做的是一件更实用也更艰难的事：把信任、发现、变现与责任重新分层，让每一层都有它专属的合规约束、单位经济学与可运营指标。在信息过载、AI冲击与信任危机三合一的今天，赢家不会是筑得最高、看起来最傲的墙，而是能同时经营"被发现的摘要、被信任的会员关系与被机构采购的可追责服务"的机构------一个把半透膜孔径调准的机构。这是可以从现在开始、一步一步做的事。

（本文所有数据与判断均基于本次会话中实际读取并核验的公开来源；截至2026年9月11日，NYT诉OpenAI/Microsoft与Penske诉Google案尚无终审结论，广告与订阅统计随季度更新。）

1. Prologue: Three tiers and a membrane

For three decades, the internet was told as a story of openness. Google turned open-web distribution into an advertising empire, and ordinary readers cycled through search, click, and monetized attention. Thirty years in, that single loop is fracturing. Information overload, generative AI, and a broad crisis of trust are pushing the industry toward three tiers. On the outside sits an open discovery layer that remains addressable to search and AI but generates little direct revenue. In the middle sits a registered membership layer that trades access for identity, community, and paid relationships. At the core sits a professional institutional layer that sells auditable methods, human accountability, and bespoke decision support.

The right metaphor is a semi-permeable membrane rather than a wall. A membrane admits certain molecules and blocks others. Applied to content, it lets abstracts, structured metadata, and licensable samples pass outward so that institutions and works remain visible to AI, search, and professional readers, while full text, raw data, and identity-based relationship data stay inside and out of the free-scraping economy. Absolute closure exits the discovery market; absolute openness surrenders pricing power. The question was never whether to build walls but which molecules to let through, which to keep, and how to price and log the traffic in both directions.

2. Erosion of trust and migration of attention

Trust decay is the first-order driver, and it is not an impression but a measurable slide. The Reuters Institute Digital News Report 2026 (48 markets, roughly 100,000 respondents) records global trust in news at 37%, the lowest since measurement began in 2015, with significant declines in 29 of 48 markets; weekly news avoidance holds at 42% versus 29% in 2017; the share of "casual or passive" news users has risen from 16% in 2021 to 25%, and "news lovers" have slipped from 29% to 22% (Reuters Institute DNR 2026). U.S. decay is steeper: Gallup's September 2025 survey put "a great deal or fair amount" of trust in mass media at 28%, below 30% for the first time, and 8% among Republicans (Gallup). Pew reported in October 2025 that trust in national news organizations fell 11 points in six months to 56%, while trust in local news held at 70% --- the collapse is graded, not uniform, and local, specialist, and person-branded sources hold up better (Pew Research). Edelman's 2025 Trust Barometer (28 countries, 33,000+ respondents) found that 69% of respondents worry that government leaders, business leaders, and journalists are "purposely trying to mislead" them, and 63% say it is increasingly hard to tell whether news comes from a credible source or a deliberate attempt to deceive (Edelman 2025).

Discovery itself is migrating. China's CNNIC 57th report (February 2026, data through December 2025) puts the online population at 1.125 billion, mobile share at 99.6%, and generative-AI users at 602 million after a one-year jump of 141.7%; search engine usage among internet users fell from 79.2% in December 2024 to 69.5% a year later (CNNIC; report PDF). On the Western side, weekly use of AI chatbots for news rose from 7% to 10%, and to 17% among the youngest cohort (DNR 2026 AI chapter). Together these signals mean the "search plus web page" golden funnel is being replaced by a blend of social, short video, and AI answers. Discovery power is being redistributed rather than destroyed.

3. Advertising: not collapse, but migration and measurement failure

The real state of advertising is not "returns at historic lows" but a simultaneous structural migration and measurement failure. IAB and PwC's Internet Advertising Revenue Report: Full Year 2025 (April 2026) shows U.S. digital ad revenue at USD 294.6 billion, up 13.9%, with programmatic up 20.5% to USD 162.4 billion and creator advertising at USD 37 billion in 2025, projected at USD 44 billion in 2026 (IAB/PwC). Linear television gives way faster: eMarketer, citing WARC, projects 2026 linear TV ad spend down more than 11% to USD 139.1 billion, with linear's share of global media spend falling from 41.3% in 2013 to 12.4%, though still 75% of global TV ad spend (eMarketer). The real pain point is measurement and efficiency: Nielsen's 2025 Annual Marketing Report finds only 32% of marketers measure traditional and digital media holistically, and 54% plan to cut ad spending (Nielsen); the ANA's programmatic supply-chain transparency study found that up to USD 20 billion of the USD 88 billion open-web programmatic pool is wasted, roughly 23% (ANA).

Trust is not simply migrating from advertisers to friends and creators. The BBB National Programs Influencer Trust Index 2025 shows 87% of U.S. consumers trust the advertising they see (5.5% completely, 83.2% somewhat), while trust in influencer advertising sits at 74%, with 26% distrusting influencers and 70% reacting negatively when commercial partnerships go undisclosed (BBB National Programs). The defensible reading is that attention and budget are migrating toward creators, communities, retail media, and video, and that undisclosed commercial relationships instantly destroy the trust premium at that layer.

Referral economics is the most misread part of the picture. Financial services have not "abandoned advertising"; they are adding measurable acquisition channels alongside advertising. Verifiable official terms include Chase Freedom cards at USD 50 per referral capped at USD 500 per year, Sapphire cards at 15,000 points per referral (capped at 100,000 per year), and Marriott Bonvoy at 40,000 points per referral (Chase); Capital One's education page notes credit-card referral rewards typically cap at USD 500 per calendar year, with actual amounts requiring login (Capital One); American Express caps referral bonuses at five per card per calendar year (American Express); SoFi's bank referral program pays up to USD 125 per referral with a USD 10,000 combined annual cap (SoFi); Robinhood's tiered program tops out at USD 15,000 of stock rewards per year (Robinhood). The real signal is not that advertising is dead. It is that when advertising becomes harder to measure, institutions will pay auditable bounties for attributable, identity-carrying introductions. That belongs to the second tier's economics, alongside subscriptions, memberships, and communities.

4. Two ways to build a "wall": closed-loop monetization in China, layered compliance in the West

The U.S. and Chinese "walls" both lock content inside, but they are built for different reasons. China's mobile walls are first a business design, not a legal mandate. Tencent's 2025 annual results attribute part of ad price growth to "the continued increase in the share of closed-loop advertising (users can go directly to Mini Programs, Weixin Stores, or Mini Games for native transactions after clicking)," with Weixin/WeChat combined MAU at 1.418 billion (Tencent 2025 Annual Results). The essential design is discovery, conversion, payment, and fulfillment inside one super-app, which turns the outside open web into an optional detour. Regulators in China have pushed the opposite way: in September 2021 the Ministry of Industry and Information Technology held an administrative-guidance session on "URL-link blocking," set a September 17 compliance deadline for messaging apps, and its senior official publicly stated that "interconnection is the necessary choice for the high-quality development of the internet industry"; the same official coverage cited experts urging that "ecosystem opening cannot be achieved overnight ... a one-size-fits-all removal of barriers is inadvisable" and calling for staged implementation with "necessary firewalls" preserved (Xinhua, 2021-09-15). That is a ready-made Chinese-context endorsement of a semi-permeable strategy.

Real-name and account governance are China's most distinctive rule stack, but the binding point is registration and publication, not "anonymous browsing is illegal." The revised Cybersecurity Law, effective January 1, 2026, requires real-identity information at network access, domain registration, information publishing, and instant-messaging (Article 26) (Cybersecurity Law). The Provisions on Administration of Internet User Account Information require accounts operating in news, economics, education, health, judicial and similar sectors to submit credentials and carry "special identifiers," and require accounts to display IP-attribution regions and subject information (CAC Provisions). The Provisions on Administration of Algorithmic Recommendation Services (Article 17) require providers to offer non-profiling options or convenient shut-offs, prohibit price discrimination by algorithm (Article 21), and require filings for services with public-opinion or social-mobilization attributes (Article 24) (CAC Provisions). The Interim Measures for Generative AI Services (Article 7) require lawful training data and non-infringement of intellectual property (CAC Interim Measures). The Measures for Labeling AI-Generated Synthetic Content, effective September 1, 2025, distinguish explicit from implicit labels, mandate provider identifiers and content IDs in metadata, and require logs to be retained for at least six months when users receive unlabeled outputs (Labeling Measures). Read together, China's tier is a layered compliance stack --- account credentials, IP display, algorithm filings, mandatory provenance labels --- not a ban on anonymous reading.

In the U.S. and EU, the equivalent walls are private-rights and infrastructure walls, and the law does not simply prohibit AI scraping. The copyright track remains unsettled: in Bartz v. Anthropic, Judge Alsup ruled in June 2025 that training on lawfully purchased books was transformative fair use, while storage of roughly 7 million pirated books constituted infringement; the case settled for USD 1.5 billion, with final approval on July 20, 2026 (Reuters, 2025-06-24; Reuters, 2026-07-20). NYT v. OpenAI/Microsoft remains active, with the Times adding contributory infringement claims against Microsoft in June 2026 (Ars Technica); OpenAI's own case page reports that certain claims have been dismissed and the fair-use question remains central (OpenAI). The antitrust track opened in September 2025 with Penske Media v. Google, the first major U.S. publisher lawsuit over AI Overviews, alleging market dominance and traffic diversion (Reuters, 2025-09-14). Licensing has its own template in the News Corp--OpenAI multi-year global deal announced in May 2024 (News Corp).

The access-control layer stress-tests how much any wall can actually stop. Cloudflare began blocking AI crawlers by default for new domains on July 1, 2025 and launched a "Pay Per Crawl" pilot, reporting that "more than 2.5 million websites" have already fully disallowed AI training crawls through its managed rules (Cloudflare Press; Cloudflare Blog). The same company later accused Perplexity of using stealth, undeclared crawlers that "appeared to mask their crawling identity" when blocked at the network layer, rotating user agents and ASNs and ignoring or failing to fetch robots.txt, prompting Cloudflare to revoke its verified-bot status (Cloudflare Blog). A licensing-semantics layer is emerging with the Really Simple Licensing (RSL) 1.0 specification, backed by more than 1,500 organizations in December 2025 for machine-readable license and price terms (RSL 1.0). The honest picture: robots.txt is a voluntary protocol (RFC 9309) rather than a legal prohibition, technical blocks can always be locally circumvented, and only a stacked posture --- technical plus contractual plus legal --- makes a defensible boundary.

A one-line comparison: China's wall is a platform-and-account layer built from real-name registration, credential-based special identifiers, algorithm filings, and mandatory provenance labels; the U.S. and EU wall is a platform-and-network layer built from default crawler blocks, machine-readable licensing, and copyright/antitrust litigation. The protected interests differ --- copyright, identity, public order --- but both jurisdictions are converging on structured semi-permeable membranes, not absolute closure.

5. AI as a filtering layer: capability, error rates, and compliance

AI's role at the public discovery layer is not to replace journalists but to take over filtering and curation. The enabling stack is retrieval-augmented generation (RAG) + explicit feedback + first-party membership data, not "deep biometric identification." Lewis et al. (2020), the original RAG paper, name the two weaknesses of purely parametric models as "provenance for their decisions" and "updating their world knowledge," and show that combining parametric with non-parametric memory produces language that is "more specific, diverse, and factual" (arXiv:2005.11401). The commercial implication is precise: models structurally need external, structured, licensable corpora. The trusted private domain is not AI's opponent but its supply chain.

Provenance capability, however, is not provenance correctness. The Columbia Journalism Review's Tow Center tested eight generative-search tools with 20 publishers and 1,600 queries in 2025 and found error rates above 60%, with Perplexity at 37% and Grok 3 at 94% (CJR / Tow Center). The EBU/BBC study "News Integrity in AI Assistants" (October 2025, 22 public-service broadcasters, 18 countries, 14 languages, 3,000+ responses, ChatGPT/Copilot/Gemini/Perplexity) found 45% of responses had at least one major issue, 31% had serious sourcing problems, and 20% had major accuracy problems, with Gemini at a 76% issue rate (BBC Media Centre). That gap is what leaves clear commercial room for a trusted private domain: the more AI summarization spreads, the more pricing power accrues to verifiably attributed institutions and authors.

Personalization should be described as riding on behavioral signals, explicit feedback, and first-party membership data, not biometrics. GDPR Article 9(1) prohibits, as a principle, the processing of biometric data for the purpose of uniquely identifying a natural person (GDPR). EU AI Act Article 5 directly prohibits emotion-inference systems in workplaces and educational institutions, biometric-based categorization inferring race, political opinions, trade-union membership, religious or philosophical beliefs, sex life or sexual orientation, and the untargeted scraping of facial images to build or expand recognition databases (EU AI Act Article 5). In these jurisdictions, biometrics is a compliance liability, not a product prerequisite.

Regulation has also legislated an "exit key" from personalization. EU DSA Article 38 requires very large online platforms and search engines to provide at least one non-profiling recommender option (DSA 38); China's Article 17 of the Algorithmic Recommendation Provisions makes the same demand (CAC). Two of the world's largest regulatory blocs treat non-profiling recommendation as a legal default, which itself falsifies the claim that AI "must surveil everything to function."

The evidence that algorithmic feeds directly drive polarization is thin. A Reuters Institute literature review concluded that echo chambers are "much less widespread than commonly assumed" and that "we do not find evidence supporting the filter bubble hypothesis" (ORA). A Science experiment on Facebook (Guess et al.) found that moving users off the algorithmic feed sharply reduced time and activity on the platform and changed their exposure structure --- chronological feeds surfaced more political and untrustworthy content but less uncivil material and more content from moderate friends and ideologically mixed audience sources --- yet "the chronological feed did not significantly alter levels of issue polarization, affective polarization, political knowledge, or other key attitudes" over three months (Science, abp9364). The defensible thesis: algorithms mainly reshape exposure structure and time-on-platform; the causal case for attitude polarization is weak; the real risk of an AI filtering layer is the redistribution of discovery and source visibility, not brainwashing.

The technical substrate for provenance is arriving. C2PA 2.2 Content Credentials use cryptographic binding to record asset provenance, edit history, and AI use; the specification is expressly "global and opt-in" and states that adding provenance is optional and "is not intended to create a two-tier media ecosystem" (C2PA Explainer). China has moved this from voluntary to mandatory: the Labeling Measures require explicit and implicit labels across text, audio, image, video, and virtual scenes; provider identifiers and content IDs in metadata; and log retention of at least six months for unlabeled deliveries (Labeling Measures). When AI answers proliferate, authorship, edit history, and license terms embedded in the content itself become the minimum floor for "trustworthy."

6. The business logic of the trusted private domain: from newsstand to platform-scale memberships

Trust erosion plus an AI-quality gap is being validated in cash. Press Gazette's 100k Club 2026 lists 62 English-language news and magazine publishers with 100,000-plus digital subscriptions each, totaling 54 million digital subscriptions --- up from 44.7 million in February 2025, about 21% growth year over year --- with The New York Times accounting for 23% of the list (Press Gazette). NYT's Q2 2026 report to the SEC shows roughly 12.8 million digital-only subscriptions and digital ARPU rising from USD 9.64 a year earlier to USD 9.94, with five consecutive quarters of ARPU growth as pricing and bundling take hold (NYT Q2 2026 SEC).

Platform-side conversion mechanics are also quantifiable. Patreon's CEO told Axios in August 2025 that creators have collected more than USD 10 billion since 2013, that the platform has 25 million-plus paid members, and that over 100 million free members convert at more than 700,000 per month (Axios); the official fee page confirms a standard 10% platform fee for creators launching after August 4, 2025 (Patreon). Substack has announced 5 million paid subscriptions cumulatively, with the official split described as "writers keep 90% of the revenue" after credit-card fees, and an official example of "1,000 subscribers at USD 5/month = USD 60,000 in yearly revenue" (Substack; Substack Going Paid). Together these numbers demonstrate a real and measurable free-to-registered-to-paid funnel --- and remind creators that platform-only distribution costs a permanent 10% of cash and 100% of the customer relationship. Owned channels earn their keep because, inside a semi-permeable membrane, the member relationship is the one asset that cannot be reclaimed by a platform's next rule change.

The "newsstand" instinct is right, but it does not signal a return to print. It signals a return from advertising economics to relationship economics. Readers are not paying for scarce information --- AI has already made that cheap. They are paying for editorial judgment, source traceability, editor and author accountability, and a reading environment insulated from advertising noise. The pricing power at the membership tier is the composite price of trust, method, and byline.

7. Academic publishing and institutional research: not collapse, but the capitalization of thresholds and responsibility

What is actually happening in academia is that access thresholds are rising, screening costs are being capitalized, and responsibility and disclosure are being institutionalized --- not that ordinary universities and journals are collapsing en masse. Since October 31, 2025, arXiv has required computer-science review articles and position papers to pass peer review at a journal or conference and provide proof before full consideration, citing "an unmanageable flood of submissions" and the fact that "generative AI / large language models have made writing articles --- especially ones that don't add new research results --- fast and easy" (arXiv Blog). A January 2026 report commissioned by STM documents that some publishers have built dedicated research-integrity teams of more than 100 people that screen millions of submissions annually, "using technology while keeping humans firmly at the center of the work"; supporting infrastructure is expanding through the STM Integrity Hub (49 member organizations), COPE (106 publisher members representing 14,500-plus journals), and United2Act (58 organizations against paper mills) (STM). A Frontiers survey of about 1,600 scholars across 111 countries reports that over half of researchers have used AI when reviewing manuscripts (Nature).

Responsibility is being tied to humans. ICMJE states that "chatbots (for example, ChatGPT) and other AI-assisted technologies should not be listed as authors because they cannot be responsible for the accuracy, integrity, and originality of the work"; authors must disclose AI use and specify how it was used, and "using AI-generated material as a primary source of citation is not acceptable" (ICMJE). COPE's August 2024 position states that "AI tools cannot be listed as authors of a paper" because they cannot be accountable, cannot declare conflicts of interest, and cannot manage copyright and license agreements; authors take full responsibility for the manuscript, including any portions produced by AI tools (COPE).

The winners in academic publishing and institutional research will be the institutions that package source traceability, transparent methodology, accessible data and code, auditable review, and human byline responsibility as a product; the losers will be mid-tier venues that treat "bulk data compilation plus formatted commentary" as their core supply. Institutional research does not survive by producing more PDFs; it survives by pivoting toward auditable advisory services, decision-grade briefings, institutional subscriptions and data licensing, and licensed AI-training corpora. This is the three-tier model applied to the research world.

8. The three-tier model as an operating blueprint

Tier 1 --- Open discovery: treat "being discoverable" as a product

The discovery layer's job is to be seen by AI, search, and cross-platform editorial networks; to be cited, summarized, and linked. Direct monetization should be planned modestly: the open-web programmatic pool wastes 23% of its own spend (ANA), Google's AI Overviews are being challenged in court by major publishers (Reuters), and AI answers themselves still fail on serious sourcing 31% of the time (BBC/EBU).

Discovery should do four things and only four. First, open abstracts and metadata --- titles, dek, structured summaries, authors, and license terms. Second, machine-readable licensing semantics --- robots.txt (RFC 9309) plus RSL 1.0 licensing plus a Pay-Per-Crawl declaration stating what can be crawled, at what price, for what purpose (Cloudflare; RSL 1.0). Third, content credentials travel with the work --- authorship, institution, edit history, and AI use bound into the asset via C2PA (C2PA). Fourth, migrate discovery-layer KPIs away from ad revenue and CPM toward citation counts, byline retention rate, and license coverage. This tier is not without revenue; it earns primarily through licensing and metered access rather than display advertising, following the News Corp--OpenAI template (News Corp).

Tier 2 --- Registered and membership: treat the relationship as the asset

The productized funnel is four steps: register (free / low-friction) → member (community and events) → paid subscription (content, archive, discussion) → household, student, and group bundles. Defensible price benchmarks live in a USD 9--20 monthly and USD 100--200 annual range, with institutional bundles at multiples. The bank referral evidence sets an audited market price for attributable human referrals: USD 25 to USD 200 per successful conversion is what mainstream institutions actually pay in cash (Chase; SoFi; Capital One). The membership economics themselves are visible in Patreon and Substack (the conversion funnel and the 90/10 split) and in the English-language subscription list growing 21% year on year (Axios; Substack; Substack Going Paid; Press Gazette).

Compliance is non-optional at this tier. Providers must offer non-profiling recommender options (DSA 38; Algorithmic Recommendation Provisions); cannot condition personalization on biometrics (GDPR; AI Act); and must implement provenance labels for AI-assisted content per local rules (China Labeling Measures; C2PA). The single largest risk is avoidance and fatigue --- 42% of respondents avoid news and 25% are passive users (DNR 2026) --- which makes editorial restraint, structured summaries, user-controlled notifications, and low-frequency curated products more strategically valuable than "more push."

Tier 3 --- Professional and institutional: treat accountability as the price

The keyword at the top tier is not "premium" but auditable, accountable, procurement-ready. It serves three demand types. Institutional subscriptions: libraries, corporations, and government pay for auditable research processes and datasets. Bespoke decision support: advisory work for concrete business questions, including situation briefings, evaluations, model audits, and training. Corpora and licensing: verifiable content licensed to AI providers by use case, connecting to the same primitives as News Corp, RSL, and Pay Per Crawl. The moat is a combination of people, methods, and reputation: responsibility rules bind humans to full accountability (ICMJE; COPE); research-integrity teams are becoming a capital-intensive capability (STM); and China's Article 11 credential-and-special-identifier system is a parallel institutionalization of "professional qualifications" (Provisions on Account Information).

The semi-permeable membrane: how the three tiers connect

A single membrane connects the tiers: abstracts and metadata pass outward to preserve discovery power; full text, raw data, and identity-based relationship data stay inside to preserve pricing power; training corpora move outward under machine-readable licenses and metered fees, creating an auditable third-party revenue stream. Absolute walls forfeit discovery; absolute openness forfeits pricing. Institutions that thrive over the long run will be the ones that treat membrane pore size as a core product parameter and tune it as the trust economy and licensing markets evolve.

Conclusion: from a grand metaphor to an operable membrane

This reconfiguration is doing something more practical and more difficult: relayering trust, discovery, monetization, and responsibility so that each tier gets its own compliance envelope, unit economics, and operating metrics. In a moment when information overload, AI shocks, and trust erosion arrive at once, the winners will not be the highest walls. They will be institutions that can simultaneously operate a discoverable abstract layer, a trusted membership layer, and a procurement-grade accountable-service layer --- institutions that keep the membrane's pore size correctly tuned. That is work that can start now, one calibrated aperture at a time.

(All figures and judgments here draw on public sources actually retrieved and read during the underlying research pass; as of September 11, 2026, NYT v. OpenAI/Microsoft and Penske v. Google remain undecided, and advertising and subscription statistics are updated quarterly.)