⚑ 危机求助IN CRISIS本图为产业结构分析,不构成任何心理健康诊断或治疗建议。如你或身边人正处危机(自伤/伤人念头/失控),请立即联系真人专业帮助:This map is industry-structure analysis and is not any mental-health diagnosis or treatment advice. If you or someone near you is in crisis (self-harm / harm to others / loss of control), reach a real professional now:中国 12356 · 希望24 400-161-9995 · US/CA 988 · 线下急诊ER
「AI 吞噬世界」 · AI IS EATING THE WORLD— №49 · HEAL / 心理疗愈
一次被听见Being Heard
AI 成了世界上最大的树洞——却既非治疗,也不受监管 · 心理疗愈产业 AI 全景图 · 2026AI became the world’s largest place to be heard — yet it is neither therapy nor regulated · the mental-wellness AI atlas · 2026
这个产业的原子,是「一次被听见」——一单位被接住的情绪。AI 把它变成了零边际成本的商品:免费、全天候、不评判,于是它成了世界上最大的心理支持提供者——头部通用对话产品周活 8 亿,每周有逾百万人在里面留下明确的自杀计划或意图迹象,而它既非临床治疗,也基本不受监管。科学最诚实的那一格同时打了所有人的脸:疗效的大头不是流派技术,是关系与期待(r≈0.28)——这解释了 AI 树洞为何真的有用,也解释了它为何有用到不了治愈;更解释了为什么证据最弱的身心灵市场反而增速最快:卖确定感与意义,永远比卖疗效好卖。墙内站着最后的真人:治疗联盟、危机兜底、身体在场、医疗责任。被听见可以免费——被担责,仍然昂贵。This industry’s atom is «being heard» — a unit of emotion caught. AI turned it into a zero-marginal-cost commodity: free, round-the-clock, non-judging, so it became the world’s largest provider of mental support — a leading chat product has 800M weekly users, and over a million a week leave clear signs of a suicide plan or intent inside it, though it is neither clinical treatment nor much regulated. The most honest cell of science slaps everyone: the bulk of therapeutic effect is not technique but relationship and expectation (r≈0.28) — which explains why the AI confidant truly helps, and why it helps but cannot cure; and why the least-evidenced wellness market grows fastest: selling certainty and meaning always outsells selling efficacy. Behind the wall stand the last humans: the therapeutic alliance, crisis backstop, bodily presence, medical liability. Being heard can be free — being answerable is still expensive.
凌晨三点谁还在线:● AI 树洞 24h 在线 · 真人咨询 熄灯 · 热线 有限——「随时可及」是最大的供给优势,也是依赖养成区Who’s online at 3 a.m.: ● the AI confidant, 24/7 · human therapy, lights out · hotlines, limited — «always available» is the biggest supply edge, and the dependence-forming zone
✓ 本图相信This map believesAI 已是事实上最大的「被听见」提供者(第一大用途 / 8 亿周活 / 0.15% 危机信号);低风险场景的 AI 心理支持有元分析级初步证据;疗效大头来自关系与期待而非流派(r≈0.28,三源学术锚);证据与规模的倒挂真实存在(不足 4% 正念 App 引用证据 vs 6.8 万亿 wellness);AI 最务实的落地是咨询师后台而非「AI 治疗师」;中国的认证真空与证书乱象真实存在。AI is de-facto the largest provider of «being heard» (a top use / 800M weekly users / 0.15% crisis signals); AI support in low-risk settings has meta-analytic preliminary evidence; the bulk of effect comes from relationship and expectation, not school (r≈0.28, three academic anchors); the evidence-to-scale inversion is real (under 4% of mindfulness apps cite science vs a $6.8T wellness economy); AI’s most practical landing is the therapist’s back office, not an «AI therapist»; China’s credentialing vacuum and certificate chaos are real.
✗ 本图不相信This map does not believe不信「AI 治疗师已经到来」——FDA 授权的 1200+ AI 医疗设备中精神健康适应症为零,两州立法禁止,首个正面 RCT 的研究者自己说「没有任何生成式 AI 可完全自主运行」;不信「聊天=治疗」——r≈0.28 只解释约 8% 方差,倾诉是入口不是全部;不信拿最佳数字当行业均值(51%/48% 是受控原型与个别研究);不信「体验有效=临床有效」;不信任何 AI 在危机场景的兜底能力;不信未经核实的传播数字(「31%」系误传、「估值 40 亿」未证实,均已剔除)。Not «the AI therapist has arrived» — of 1,200+ FDA-authorised AI medical devices, zero have a mental-health indication, two states have banned it, and the author of the first positive RCT says «no generative AI can run fully autonomously»; not «chat = therapy» — r≈0.28 explains only ~8% of variance, venting is an entrance not the whole; not treating best-case figures as the industry mean (51%/48% are a controlled prototype and single studies); not «feels effective = clinically effective»; not any AI backstop in a crisis; not unverified viral numbers («31%» is a misquote, a «¥4B valuation» unconfirmed — both removed).
B
逾100万/周
头部对话产品中每周出现明确自杀计划或意图迹象的用户量级(自报×周活折算)——「最大树洞」的重量users/week showing clear suicide-plan or intent signs in a leading chat product (self-report × weekly actives) — the weight of the «largest confidant»
B
72%
美国青少年用过 AI 陪伴——下一代的默认求助对象已换人of US teens have used an AI companion — the next generation’s default listener has changed
A
r≈0.28
治疗联盟与疗效的相关(约 295 项研究)——疗效大头是关系,不是流派correlation of therapeutic alliance with outcome (~295 studies) — the bulk of effect is the relationship, not the school
B
6.4万
中国精神科医生总数(每 10 万人约 4 名,约为欧美三分之一;92% 重性精神障碍未获治疗)——真人供给的地基缺口psychiatrists in China (~4 per 100,000, about a third of Western levels; 92% of severe mental illness goes untreated) — the bedrock gap in human supply
⚖ 口径裁判 · 先读我Calibration notes · read me first①「逾 100 万人/周」=厂商自报比例(0.15%)× 周活(8 亿)的折算值,非独立统计;②市场总盘至少五套口径切勿相加混用——全球 wellness 6.8 万亿(2024)/mental wellness 2683 亿(2024)与 1810 亿→3300 亿(2022→2027)系两代口径 / 数字心理健康 278 亿 / EAP 56–103 亿 / 中国「情绪消费」2.72 万亿为超宽口径单源;③占星「22 亿 / 228 亿 / 40 亿」为三套互斥口径;④「治疗与陪伴第一大用途」为定性榜单,网传「31%」系误传,不上页;⑤Therabot 51% 为受控研究原型(候补对照)、48%/43% 为个别研究最佳值,行业口径 = 元分析「小到中等」;⑥Woebot 已于 2025 年中关停消费级 App 且从未获 FDA 正式批准,仅作关停案例;Wysa「突破性认定」≠获批;⑦诉讼情节均为原告单方指控,和解 ≠ 责任认定;⑧渗透% 与压缩速度排序为编者综合评估值,非统计值;⑨本页综合四份独立深度研究交叉整理,为批判性产业结构分析,不构成任何心理健康诊断或治疗建议;危机场景请务必联系真人专业帮助。① «1M+/week» is a self-reported ratio (0.15%) × weekly actives (800M), not an independent statistic. ② At least five market-size definitions must never be added or mixed — global wellness $6.8T (2024) / mental wellness $268.3B (2024) vs $181B→$330B (2022→2027) are two vintages / digital mental health $27.8B / EAP $5.6–10.3B / China’s «emotional consumption» ¥2.72T is a broad single source. ③ Astrology «$2.2B / $22.8B / $4B» are three mutually exclusive scopes. ④ «Therapy & companionship, a top use» is a qualitative ranking; the viral «31%» is a misquote and is kept off the page. ⑤ Therabot’s 51% is a controlled prototype (waitlist control); 48%/43% are single-study bests; the industry read is a meta-analytic «small-to-moderate». ⑥ Woebot shut its consumer app mid-2025 and never had formal FDA approval, used only as a shutdown case; Wysa’s «breakthrough designation» ≠ approval. ⑦ Litigation details are plaintiffs’ allegations; a settlement ≠ an admission of liability. ⑧ Penetration % and the compression ranking are editorial estimates, not statistics. ⑨ Compiled from four independent deep-research reports as critical industry-structure analysis; not any mental-health diagnosis or treatment advice; in a crisis, please reach a real professional.
◎ 中心裂缝 · 全图的坐标原点The central rift · origin of this map
最大的心理支持提供者,恰恰是最不受监管的那一个The largest mental-support provider is precisely the least regulated one
把这个产业沿「一次被听见的一生」摊开,会看到 AI 的进食顺序被直觉搞反了:大家怕的是「AI 取代咨询师」,可 AI 先做的是把「被听见」这个原子变成零边际成本商品——树洞、解释、确认、陪伴四件事,24 小时、不评判、近乎免费。于是它顺手成了世界上最大的提供者:免费、可及、不评判的三重优势,是你身边的人往往无法同时满足的。但科学诚实的那一格立刻拉住了所有人——疗效的大头(r≈0.28)来自「关系与期待」,而这个相关只解释约 8% 的疗效方差。翻译过来:AI 树洞真的有用,但有用到不了治愈;身心灵之所以「感觉有效」,是因为高价与仪式感拉高了期待,而期待本身就是疗效成分。被听见很便宜,被治愈很贵,被担责最贵。Unroll this industry along «the life of being heard» and AI’s feeding order is the reverse of intuition: people fear «AI replaces therapists», yet AI first turned the atom of «being heard» into a zero-marginal-cost commodity — a confidant, an explanation, validation, company, 24/7, non-judging, nearly free. So it became the world’s largest provider: free, accessible, non-judging — three advantages the people around you often cannot supply at once. But the honest cell of science pulls everyone back — the bulk of effect (r≈0.28) comes from «relationship and expectation», and that correlation explains only ~8% of outcome variance. Translated: the AI confidant genuinely helps, yet helps but cannot cure; the wellness track «feels effective» because high prices and ritual raise expectation, and expectation is itself an active ingredient. Being heard is cheap, being cured is dear, being answerable is dearest.
被听见层 · AI 近独占Heard · nearly AI-only
HEARD — 零边际成本区zero-marginal-cost zone
自助与树洞Self-help & confidant最高渗透 ~70–85%,凌晨可及peak penetration ~70–85%, available at 3 a.m.
第一入口First point of contact「我先问 AI」取代先搜索先问朋友(~65–75%)«I ask AI first» replaces search or a friend (~65–75%)
分诊与匹配Triage & matching平台胜负手从流量转向分层能力the platform’s edge shifts from traffic to tiering
被担责层 · 真人堡垒Answerable · the human fortress
HUMAN — 执照与在场区licence & presence zone
治疗联盟The therapeutic alliancer≈0.28,能受苦、能修复裂痕的主体在场r≈0.28, a subject present who can suffer and repair ruptures
危机兜底Crisis backstop执照 / 报警 / 到场——AI 没有可被惩罚性licence / calling for help / being present — AI has no punishability
精神科诊疗Psychiatric care诊断 / 处方 / 医疗责任(FDA 精神科适应症至今为零)diagnosis / prescription / medical liability (zero FDA psych indications to date)
8%
科学的诚实之锚:治疗联盟与疗效的相关 r≈0.28,平方后只解释约 8% 的疗效方差——其余 92% 未被这一项解释。这一个数字同时做了两件事:拆掉了循证轨的「流派优越感」(渡渡鸟裁决:正规疗法大致等效),也解释了 AI 树洞与身心灵为何「真的有用、却治不了病」。本图用它当标尺,评估这张图上的一切产品。Science’s honesty anchor: the alliance-outcome correlation r≈0.28, squared, explains only ~8% of outcome variance — the other 92% is not explained by it. This one number does two things at once: it dismantles the evidence track’s «school superiority» (the Dodo-bird verdict: bona-fide therapies are roughly equivalent) and explains why the AI confidant and the wellness track «truly help yet cannot cure». This map uses it as the ruler for every product on the page.
一句话读法:AI 不是在链条某一点替换人,而是在两端(自助 / 树洞、维持)几乎独占,在中段(真正的治疗工作)只能辅助。三层信号灯贯穿全页:冷蓝=被听见(AI) · 琥珀=被接住(人机) · 暖光=被担责(真人)——越往真人越暖,暖光只给「有真人在场」的格子。被听见可以免费,被担责仍然昂贵。One-line reading: AI does not replace people at one point in the chain — it nearly monopolises the two ends (self-help / confidant, maintenance) and can only assist in the middle (the real therapeutic work). Three signal lights run through the page: cold blue = heard (AI) · amber = held (human+AI) · warm light = answerable (human) — the closer to a human, the warmer; warm light is reserved for cells where a real person is present. Being heard can be free; being answerable is still expensive.
ATHE JOURNEY · 主脊 · 一次被听见的一生The spine · the life of being heard
一段疗愈旅程八节点:两端 AI 独占,中段真人堡垒Eight nodes of a healing journey: AI owns the two ends, humans hold the middle
这个产业的原子不是一次咨询、一次课程、一次测评,而是「一次被听见」——一单位被接住的情绪。每个说「我不太好」的人都要走一遍旅程:痛苦觉察 → 自助 → 寻求帮助 → 评估匹配 → 正式工作 → 会间巩固 → 改变或脱落 → 结束维持。AI 在两端(树洞 / 维持)几乎独占,在中段(真正的治疗)只能辅助。每节点标承接强度:被听见 / 被接住 / 被担责;渗透 % 为编者评估值。The atom here is not a session, a course or a test but «being heard» — a unit of emotion caught. Everyone who says «I’m not okay» walks the journey: awareness → self-help → seeking help → assessment & matching → the real work → between-session consolidation → change or dropout → ending & maintenance. AI nearly monopolises the two ends (confidant / maintenance) and can only assist in the middle (the real therapy). Each node is tagged by strength of holding: heard / held / answerable; penetration % are editorial estimates.
BTWO INVERSIONS · 两个倒挂与一场收网Two inversions and a reckoning
证据与规模倒挂、后台文书革命,以及被诉讼倒逼的监管补课The evidence-to-scale inversion, the back-office revolution, and the litigation-driven regulatory catch-up
中篇是全图最沉重的三张牌:钱流向证据最弱的一端、AI 最务实的红利在后台而非前台、而 2025 是心理 AI 的「监管补课元年」——第三张必须与顶部危机提示条同屏读。The middle holds the map’s three heaviest cards: money flows to the least-evidenced end, AI’s most practical dividend is in the back office not the front, and 2025 became the «regulatory catch-up year» for mental-health AI — read the third alongside the crisis bar at the top.
CHARD BONES · 硬骨头Hard bones
AI 攻不进的七块:被听见可以免费,被担责仍然昂贵Seven things AI cannot breach: being heard is free, being answerable is not
利润与地位正向这七堵墙内迁移:能承担危机责任、能建立算数的关系、能到场的位置,最后失守。每一块都不是「模型还不够强」,而是「问题的结构不给模型留位置」。Profit and status migrate behind these seven walls: the positions that can bear crisis liability, build a relationship that counts, and be physically present fall last. Each is hard not because «models aren’t strong enough» but because the problem’s structure leaves a model no seat.
DSTANCE MAP · 立场地图Stance map
你站在哪:按角色 × 承接强度三层过滤Where you stand: filter by role × the three tiers of holding
选「我是谁」高亮相关卡;切三层信号灯——被听见(AI 独占) / 被接住(人机协作) / 被担责(真人堡垒)。开关本身就是「被听见可以免费,被担责仍然昂贵」论点的交互证明。Pick «who I am» to highlight cards; toggle the three lights — heard (AI-only) / held (human+AI) / answerable (the human fortress). The toggles are the interactive proof of «being heard is free, being answerable is expensive».
四份底稿对下面四题立场不一——各方并置,再给编者合议,不裁掉任何一方。The four sources disagree below — each side shown, then an editorial synthesis; no side is cut.
TIMELINE · 时间线Timeline
2015–2026:审批线 vs 事故线 的双轨2015–2026: the twin tracks of approval and incident
底稿对 2015–2016 极薄(冥想 App 与在线平台前史,标 ※),2023 后骤厚(平台冰火 / DTx 破产与报销破冰 / 诉讼与监管收网)——琥珀=审批与报销,红=事故与禁令,两轨交缠说明「审批 ≠ 存活、爆发 ≠ 安全」。薄区如实留稀。Sources are thin on 2015–2016 (the prehistory of meditation apps and online platforms, ※) and thicken after 2023 (platform boom-and-bust / DTx bankruptcies and reimbursement thaw / litigation and regulatory catch-up) — amber = approval & reimbursement, red = incident & ban; the two tracks interweave to show «approval ≠ survival, boom ≠ safe». Sparse zones left honestly sparse.
FIELD GUIDE · 产品指南Field guide
四种人的「现在能用 / 别碰 / 戒律」——每类第一条都是同一条红线«Use now / avoid / discipline» for four roles — the first line is the same red line for all
厂商效率数字一律 D 级;所有清单是行业结构信息,不构成心理健康建议。每类角色的最高优先都相同:任何危机场景 = 真人专业帮助(12356 / 988 / 希望24 / 急诊),这条红线高于本表一切内容。Vendor efficiency figures are grade D; these lists are industry-structure information, not mental-health advice. The top priority is identical for every role: any crisis = a real professional (12356 / 988 / hope-line / ER), above everything in this table.
FIVE LAWS · 五条规律Five laws
把整张图压缩成五句判断The whole map, compressed into five judgments
L1
最大的提供者,是最不受监管的那一个The largest provider is the least-regulated one
AI 先把「被听见」变成零边际成本商品,顺手成了世界上最大的心理支持提供者(周活 8 亿、逾百万人/周留下危机信号),而它既非临床治疗、也基本不受监管。危险不在于 AI 会不会存在——它已经存在;在于谁定义边界、何时必须转人、谁承担事故责任。AI first turned «being heard» into a zero-marginal-cost commodity and thereby became the world’s largest mental-support provider (800M weekly users, 1M+/week leaving crisis signals), while being neither clinical treatment nor much regulated. The danger is not whether AI will exist — it does; it is who defines the boundary, when a handoff to a human is mandatory, and who bears the liability for an incident.
L2
疗效大头是关系与期待,不是流派技术(r≈0.28 / 8%)The bulk of effect is relationship and expectation, not technique
治疗联盟与疗效相关 r≈0.28——它同时拆了循证轨的流派优越感,也解释了 AI 树洞与身心灵为何「感觉有效」;但 r≈0.28 只解释约 8% 方差。AI 能把「共同因素」做到极致(不疲惫、不拒绝、永远在),却有用到不了治愈。这是评估这张图上一切产品的标尺。Alliance-outcome correlation is r≈0.28 — it dismantles the evidence track’s school superiority and explains why the AI confidant and wellness «feel effective»; but r≈0.28 explains only ~8% of variance. AI can max out the «common factors» (never tired, never refusing, always there) yet helps but cannot cure. This is the ruler for every product on this map.
L3
倒挂:证据最弱的一端,增速最快The inversion: the least-evidenced end grows fastest
全球 wellness 6.8 万亿美元,而不足 4% 的正念 App 引用科学证据;数千元速成的「疗愈师」证与博士级临床训练同台竞争「接住情绪」这单生意——市场给了前者更快的增速。卖确定感与意义,比卖疗效好卖。越往身心灵与 AI 树洞,越是「感觉有效」压过「临床有效」。A $6.8T global wellness economy, yet under 4% of mindfulness apps cite science; a few-thousand-yuan «healer» certificate competes with doctoral clinical training for the same job of «catching emotion» — and the market gives the former faster growth. Selling certainty and meaning outsells selling efficacy. The further into wellness and the AI confidant, the more «feels effective» beats «clinically effective».
L4
AI 最务实的落地在后台,不在前台AI’s most practical landing is the back office, not the front
录音转纪要、个案概念化、督导保真度、随访依从——后台文书自动化(某工具已处理 520 万份纪要,⚑宣称省 50–60% 时间)才是 AI 在这个行业最稳的红利;「独立消费级 AI 治疗师」则撞上州法禁令、FTC 调查与 APA 警示。把 AI 省下的时间还给关系,而不是加密排班。Recording-to-notes, case formulation, supervision fidelity, follow-up adherence — back-office automation (one tool has processed 5.2M notes, claiming ⚑ 50–60% time saved) is AI’s steadiest dividend here; the «standalone consumer AI therapist» hits state bans, an FTC probe and APA warnings. Give the time AI saves back to the relationship, not to a denser schedule.
L5
审批 ≠ 存活:在医疗世界,报销码比下载量值钱Approval ≠ survival: in medicine, a billing code beats a download
数字疗法审批谱系成型(首个 PDT → 首个抑郁症 DTx),但先驱靠 3.1 万张处方救不活 1000 美元定价而破产,十亿估值跌去 96%。活下来的路径是嵌入支付方与临床路径(靠保险反转 / 靠 NHS / 靠文书自动化),而非独立消费级「AI 治疗师」。证据打不赢渠道。The DTx approval lineage formed (first PDT → first depression DTx), yet a pioneer went bankrupt because 31,000 prescriptions couldn’t sustain a $1,000 price, and a billion-dollar valuation fell 96%. The survivors embed into payors and clinical pathways (turning on insurance / on the NHS / on note automation), not a standalone consumer «AI therapist». Evidence does not beat distribution.
STANCE · 本图立场Where this map stands本图站在批判性祛魅一侧:既不替「AI 治疗师已经到来」的叙事抬轿,也不替「AI 只是玩具、真正的治疗永远属于人」的傲慢背书。我们把四份高度同构的深研底稿拆开重验:框架共识多为研究命题的回声,只有八条事实独立收敛——AI 已是最大被听见入口、疗效大头在关系与期待、证据与规模倒挂、初步有效与边界事故并存、危机必须真人执照、中国认证真空、审批 ≠ 存活、后台文书是最务实落地。这个产业最擅长把「感觉有效」讲成「临床有效」、把「被听见」讲成「被治愈」——本图的全部工作,是用科学最诚实的那个数字(r≈0.28、8% 方差)当标尺,把这些重新拆开。被听见可以免费,被治愈很贵,被担责最贵。This map stands on critical demystification: it carries the sedan chair neither for «the AI therapist has arrived» nor for the arrogance of «AI is a toy, real therapy will always belong to humans». We took four highly isomorphic reports apart and re-verified: most framework consensus echoes the commissioning prompt; only eight facts converge independently — AI is the largest «being heard» entrance, the bulk of effect is relationship and expectation, evidence and scale are inverted, preliminary efficacy coexists with boundary incidents, crisis requires a licensed human, China’s credentialing is a vacuum, approval ≠ survival, and the back office is the most practical landing. This industry excels at narrating «feels effective» as «clinically effective» and «being heard» as «being cured» — the whole job of this map is to take those apart using science’s most honest number (r≈0.28, 8% of variance) as the ruler. Being heard can be free, being cured is dear, being answerable is dearest.
⚑ 危机与边界:本图为批判性产业结构分析,不构成任何心理健康诊断或治疗建议。若你或身边人正处危机(自伤 / 伤人念头 / 幻觉妄想 / 失控),请立即联系真人专业帮助——全国统一心理援助热线 12356(中国大陆)· 24 小时希望24热线 400-161-9995 · 988(美国/加拿大)· 线下急诊优先。AI 没有执照、无法报警、不能到场;任何 AI 陪聊、疗愈课程或冥想 App 都不应把危机承接到底。⚑ Crisis and boundaries: this map is critical industry-structure analysis and is not any mental-health diagnosis or treatment advice. If you or someone near you is in crisis (self-harm / harm to others / hallucinations or delusions / loss of control), reach a real professional now — China’s unified support line 12356 · 24h hope line 400-161-9995 · 988 (US/Canada) · the ER first. AI has no licence, cannot call for help, cannot be present; no AI chat, healing course or meditation app should carry a crisis to the end.