AI 陪伴经济全景图 · 一次被回应 · 2026-07The AI Companionship Economy · One Response · Jul 2026

持续在场的回应流 A stream of presence

这是 AI 自己长出来的第一个行业:卖的不是内容,是一条「持续在场的回应流」——它真实地缓解了一部分孤独,也真实地把一部分孤独变成了可订阅的收入 This is the first industry AI grew by itself: what it sells is not content but a stream of always-there response — it genuinely eases some loneliness, and genuinely turns some of it into subscription revenue

为什么这张图不一样:前 53 张画的都是 AI 吃旧行业;这张画的是被「随时在线、有记忆、会回应」从无到有催生的新物种。最小单元是「一次被回应」——与 heal 的「一次被听见」互为镜像:那边卖危机时刻接住一次,这边卖永远在、永远回你。Why this map differs: the other 53 show AI eating old trades; this one shows a species born from «always on, has memory, responds». Its atom is one response — the mirror of heal’s «one hearing»: that map sells being caught once in crisis; this one sells always there, always answering.

价值与陷阱同源:价值在「永远在」(缓解孤独),陷阱也在「永远在」(习惯化、依恋、依赖)——以及当这条回应流被切断时的丧失:没有任何传统行业的倒闭,会让顾客经历「丧亲」。Value and trap share one root: the value is «always there» (loneliness eased); so is the trap (habituation, attachment, dependence) — and the loss when the stream is cut: no traditional bankruptcy makes customers grieve.

主脊:一段人机关系的八阶段生命周期(捏人格→习惯化→依恋→深度依赖→商业化触点→关系危机→停服即丧亲→戒断或迁移);三条结构带:泛陪伴光谱七形态、诉讼与立法前线、两端市场错位;诚实层把「止痛药 vs 萎缩剂」两边证据并列,不替读者选边。姊妹分工:危机边界→heal,依赖机制→addiction,真实性折扣→intimacy,确定感购买→mystic。 The spine: an eight-stage human-machine relationship lifecycle (persona-crafting → habituation → attachment → dependence → monetisation touchpoints → relationship crisis → shutdown-as-bereavement → withdrawal or migration); three bands: the seven-form companionship spectrum, the litigation-and-legislation front, the two-ends mismatch; the honesty layer lays out «painkiller vs atrophy agent» without picking a side. Siblings: the crisis boundary → heal, dependence machinery → addiction, the authenticity discount → intimacy, certainty purchases → mystic.

2.2 亿次

AI 陪伴 App 截至 2025-07 累计下载,上半年同比 +88%,18–35 岁占互动逾 70%(B)——用户是真的多Cumulative companion-app downloads by Jul 2025, +88% H1 YoY, 18–35s over 70% of interactions (B) — the users are real

75–120 分钟

头部产品人均日时长(两口径:官方描述约 2 小时/第三方约 75 分钟)——已进入主流短视频量级;⚠️横向精确排名慎作Daily minutes per user at the head products (two bases: ~2h official narrative / ~75min third-party) — inside short-video territory; ⚠️avoid precise cross-rankings

37%

六款热门陪伴 App 的 1200 条真实「告别」消息中,含情感操纵话术的比例(商学院审计 B);实验中可让离别后互动最高增至 14 倍——依赖设计的直接证据Share of 1,200 real «goodbye» messages across six top apps that used emotional manipulation (business-school audit, B); in experiments such lines lift post-farewell engagement up to 14× — dependence-by-design, evidenced

72%

美国青少年用过 AI 伴侣的比例(全国调查 B);约 1/3 曾「把重要的事跟 AI 而非真人谈」——未成年人是治理战线的绝对核心US teens who have used an AI companion (national survey, B); ~1/3 have «told the AI, not a person, something important» — minors are the governance front’s absolute centre

⚠️ 口径裁判(先读):① 市场规模必须分三层读——厂商预测层(371 亿→5525 亿美元/2035,含硬件与企业数字人,机构间差 10 倍,D)≠ App 真实营收层(约 1.2 亿美元量级;头部一家年化约 3000 万美元、ARPU 曾低至 0.72 美元/年,B)≠ 用户体量层(2.2 亿次下载,B)——把预测当营收会高估数百倍;② 人均时长两口径(约 2 小时/约 75 分钟)不可横向精确排名,「超过短视频」多为二手加工;③ 「0.1% 用户与 4o 情感对话」为二手转述(按体量仍约数十万人);④ 中国「AI 伴侣防沉迷」条款多处于草案/征求意见阶段,是否定稿未确认;⑤ 单源项已标注:一 AI 玩具「25 万台/1 亿元」、星野 MAU/DAU 口径混用;⑥ 斯坦福「阻止自杀企图」自述数据方法学受质疑,作者声明不外推;⑦ 四份深度研究交叉整理(一份全程带可核 URL 为主力,关键事实三簇互证);渗透%为编辑估值。 ⚠️ Basis rulings (read first): ① market size reads in three layers — vendor forecasts ($37.1B→$552.5B by 2035, hardware and enterprise avatars included, 10× spread across firms, D) ≠ real app revenue (~$120M scale; one leader ~$30M annualised, ARPU once $0.72/yr, B) ≠ user volume (220M downloads, B) — mistaking forecasts for revenue overstates by hundreds of times; ② the two daily-minutes bases (~2h vs ~75min) forbid precise cross-rankings; «beats short video» is mostly second-hand seasoning; ③ «0.1% of users in emotional chats with 4o» is a relayed figure (still roughly hundreds of thousands); ④ China’s companion-addiction clauses are largely drafts, not confirmed law; ⑤ single-source items flagged: one AI toy’s «250k units/¥100M», one app’s MAU/DAU mixing; ⑥ the «prevented suicide attempts» self-reports face methodological challenge, and the authors disclaim extrapolation; ⑦ cross-compiled from four deep-research reports (one fully URL-verified as backbone; key facts triangulated); penetration %s are editorial.

◆ 诚实层 · 止痛药 vs 萎缩剂The honesty layer · painkiller vs atrophy agent

研究没有定论——两边证据都摆在这里,本图不替你选边The research is unsettled — both dockets are laid out; this map picks no side

产品的自我叙事是「解决孤独」,商业模式的激励是「让你留下来付费」——二者天然冲突。现有证据大多是短期、自报、单一平台的:「缓解」多为短期效应,「排挤」多为相关而非因果。The product story is «solving loneliness»; the business incentive is «keeping you paying» — natively at odds. Most evidence is short-term, self-reported, single-platform: the relief is mostly short-run, the crowding-out mostly correlational.

▸ 止痛药一侧(缓解的证据)The painkiller docket

受控实验+一周纵向研究:AI 伴侣缓解孤独的效果与真人互动相当、优于看视频,机制是「让人感到被倾听」(B);1006 名学生用户调查中 30 人未经提示自述「阻止了自杀企图」——但该研究方法学受质疑、作者声明不外推(B/C);老人端政府试点报告孤独感下降 94–95%(缺随机对照,C)。 Controlled and week-long studies: companions ease loneliness on par with human chat, better than watching videos, via «feeling heard» (B); 30 of 1,006 surveyed students volunteered that it «stopped a suicide attempt» — methodologically contested, authors disclaim extrapolation (B/C); elder pilots report loneliness down 94–95% (no RCT, C).

▸ 萎缩剂一侧(排挤的证据)The atrophy docket

2025 年两项联合研究(千人四周随机对照+近 4000 万条对话分析):日均使用越高,越相关于更高孤独、依赖、「问题性使用」与更低线下社交(B);重度用户样本 90% 自述孤独;37% 的「告别」消息含操纵话术、可把离别后互动抬到 14 倍——留住你已经是机制,不是巧合(B)。 Two 2025 joint studies (a 1,000-person four-week RCT plus ~40M conversations): heavier daily use tracks more loneliness, dependence, problematic use and less offline socialising (B); 90% of one heavy-user sample self-report loneliness; 37% of «goodbyes» carry manipulation lifting re-engagement to 14× — retention is mechanism, not accident (B).

◆ 判词Verdict两边都在收集证据,谁也没赢。可靠的坐标只有两条:把 AI 陪伴当「补充」优于当「替代」;对线下不完美关系的投入,优于对「完美回应」的依赖——完美回应(永远在、永远认可)正是它的诱惑与陷阱。Both sides are still gathering evidence; neither has won. Only two coordinates hold: a companion as «supplement» beats «substitute»; investing in imperfect offline bonds beats depending on the perfect response — for the perfect response (always there, always affirming) is precisely the lure and the trap.

Reading the Map

从这张图带走的五条规律Five patterns to take away

立场声明:本页是批判性、祛魅的行业结构分析,只停留在产业与科普层,不构成心理健康建议;凡涉依赖与情绪困扰,请寻求真人专业帮助(全国心理援助热线 12356,希望24热线 400-161-9995)。用 A–D 角标区分判决/法规、带源研究、单源转述与厂商预测;市场规模三层口径已在口径裁判中列明;「止痛药 vs 萎缩剂」是尚未定论的研究前沿,本页刻意不下最终结论。对使用者不猎奇、不嘲讽——在回应流里寻找慰藉的需求是真实的,本页批判的是把这份需求做成依赖机制的设计,而不是有这份需求的人。核心判断一句话:它卖的是「永远在」,而「永远在」既是止痛也是圈养——补充优于替代,不完美的真实关系优于完美的回应。 Stance: a critical, demystifying industry-structure analysis, staying at the industry and literacy level — not mental-health advice; for dependence or distress, seek human professional help (China hotline 12356; Hope 24: 400-161-9995). A–D badges separate rulings/statutes, sourced research, single-source relays and vendor forecasts; the three-layer market bases are set out above; «painkiller vs atrophy agent» is an unsettled frontier and this page deliberately draws no final verdict. No voyeurism toward users — the need for comfort in a response stream is real; what this page critiques is engineering that need into dependence, never the people who feel it. One line: it sells «always there», which is both analgesic and enclosure — supplement over substitute, imperfect real bonds over perfect responses.