智能时代的成瘾解剖学 · 注意力与多巴胺产业 · 2026-07The Anatomy of Addiction · Attention & Dopamine Industry · Jul 2026

所有成瘾都始于同一个动作——一次拉杆。AI 做的事只有一件:把「变率奖励回路」从大众化的概率游戏,升级成对每个独立个体的参数化精确打击 Every addiction begins with the same gesture — one pull of the lever. AI did exactly one thing: it upgraded the variable-reward loop from a mass probability game into a parameterized precision strike on each individual

先把责任放回正确的位置:你的强迫性刷屏不是意志力薄弱,而是一整队工程师在优化你的 engagement 指标——在 AI 亿万次算力迭代算出的个体多巴胺阈值面前,靠意志力强撑,注定是一场必败的战役。First, put the blame where it belongs: your compulsive scrolling is not weak willpower — it's a team of engineers optimizing your engagement metric; against the individual dopamine threshold AI computed over a billion iterations, willpower alone is a battle already lost.

再看清底层的伦理事实:优化 Engagement 本身,在设计之初就是一种对齐偏差——当系统被单一设为最大化停留时长,它会自发「压平」你的心理健康、睡眠与真实社会连接。Then see the ethical bedrock: optimizing engagement is itself a misalignment by design — set a system to maximize time-on-device and it will spontaneously «flatten» your mental health, your sleep, and your real social connection.

主脊是成瘾回路安装生命周期六节点(注意力捕获→习惯安装→耐受升级→深度变现→防流失→反噬),每节点标注行为科学原型 + 赌场对应物 + 传统 vs AI。另含上瘾闭环、诚实层(产品话术 vs 真实优化目标)、「毒枭不吸货」、杠杆排序防御、防御与治理指南(2C 用户自卫 / 2B 治理侧)。这是一张解剖图,不是施工图。它是所有成瘾产品脚下的同一根拉杆——具体产业深挖见姊妹图:赌博→bet、游戏→game、AI 伴侣关系→love。 The spine is the six-node addiction-loop lifecycle (capture → habit → tolerance → monetize → retain → backlash), each tagged with its behavioral-science principle + casino analog + traditional vs AI. Plus the hook loop, the honesty layer (pitch vs the real objective), «dealers don't use their own product», a defense-leverage ladder, and a defense & governance guide (2C self-defense / 2B governance). This is an anatomy chart, not a build guide. It is the one lever beneath every addictive product — the deep dives live on siblings: gambling → bet, games → game, AI-companion relationships → love.

变率奖励 · 多巴胺Variable reward · dopamine

行为科学 · 对齐失败Behavioral science · misalignment

AI 节点 · 防御AI node · defense

渴求 · 收割 · 反噬Craving · harvest · backlash

传统节点Traditional

+73%

康奈尔「无底汤碗」实验:用自动续汤的碗进食,进食量 +73% 却不觉更饱——人判断何时停止靠「视觉停止线索」而非生理饱和。无限下拉流＝数字无底汤碗(经典实验,作者后期有学术争议)Cornell's «bottomless bowl»: eating from a self-refilling bowl, people ate 73% more yet didn't feel fuller — we stop on «visual stopping cues», not satiety. Infinite scroll is the digital bottomless bowl (a classic study; its author later drew academic controversy)

20 → 97%

约 20% 的观众贡献了 97% 的直播礼物(ScienceDirect 学术)——「鲸鱼」集中度,是变率奖励收割的终点站About 20% of viewers give 97% of live-stream gifts (ScienceDirect) — the «whale» concentration, the terminal stop of variable-reward harvesting

1.5小时/周

Peter Thiel 只让幼子每周用屏 1.5 小时(当场引发观众惊呼)——缔造数字老虎机的精英,是数字极简最严格的践行者;乔布斯禁孩子用 iPad、盖茨 14 岁前禁手机Peter Thiel lets his young sons use screens just 1.5 hours a week (drawing audible gasps) — the elites who built the digital slot machine are the strictest digital minimalists; Jobs banned his kids from the iPad, Gates barred phones before 14

3起

Character.AI 已面临至少 3 起青少年自杀 / 自残诉讼(Setzer / Peralta / Raine,司法一手);2025-11 起禁 18 岁以下开放式对话,2026-01 与谷歌就 Setzer 案和解Character.AI faces at least 3 teen suicide/self-harm suits (Setzer / Peralta / Raine, court record); it barred open-ended chat for under-18s in Nov 2025 and settled the Setzer case with Google in Jan 2026

口径警告:本页是「解剖图,不是施工图」——机制描述用于防御与治理,不作「如何让产品更上瘾」的操作指南;文末防御指南同理,只列自卫与治理工具。「无底汤碗」为经典实验但作者 Wansink 后期有数据/撤稿争议,故辅以 Skinner 变率等更稳固一手研究;AI 伴侣「每天聊 15 小时」为 Reddit 自陈(arXiv:2507.15783,非临床);第三方分析站数据(Character.AI 日均 75 分钟 / 豆包 DAU 2 亿)标「量级参考」;厂商自述降本(京东 1/10、硅基 80%、星巴克 ROI+30%)标「非独立审计」;市场规模含预测(短剧 >94 亿美元、虚拟礼物 3672 亿元含 E);PMI「Your Voice」Astroturfing 事件须并陈 PMI 已公开否认「操纵」措辞;减害替代品(无醇酒 / 尼古丁袋)是否净减害学界无定论,不下结论。危害与致死案例只采信司法 / 一手研究来源。合规审计与心理研究会在同一议题上给出相反结论,本页原样保留这组张力:欧盟议会请愿委员会 2025 年讨论「禁止未成年人接触战利品箱」时指出,战利品箱已受既有消费者保护法约束、DSA 指南要求平台使未成年人远离结果不可预测的虚拟物品,而在制定中的《数字公平法》将补上在线消费者保护缺口;但同一份通讯记录了2024 年法国的一次执行审计「未发现令人担忧的情况」。一个测「有没有违规」,一个测「有没有伤害」,不是同一个问题。 Basis warning: this page is «an anatomy chart, not a build guide» — mechanism descriptions serve defense and governance, never a «how to make it more addictive» manual; the defense kit at the end follows the same rule, listing only self-defense and governance tools. The «bottomless bowl» is a classic study but its author (Wansink) later faced data/retraction controversy, so it's backed by more solid primary work (Skinner's variable-ratio); AI-companion «15 hours a day» is Reddit self-report (arXiv:2507.15783, not clinical); third-party-analytics figures (Character.AI's 75 min/day, Doubao's 200M DAU) are marked «order-of-magnitude»; vendor cost-cut claims (JD 1/10, Silicon 80%, Starbucks ROI +30%) are «not independently audited»; market sizes include forecasts (short-drama >$9.4B, virtual gifts ¥367B incl. est.); PMI's «Your Voice» astroturfing episode must be shown alongside PMI's public denial of «rigged»; whether harm-reduction substitutes (zero-alcohol, nicotine pouches) are net-reducing is academically unsettled and left open. Harm and fatality cases cite only court / primary sources. Compliance audits and psychological research can reach opposite conclusions on the same question, and this page keeps the tension intact: when the European Parliament petitions committee took up a 2025 petition to bar minors from loot boxes, it noted that loot boxes already fall under existing consumer-protection law, that DSA guidance asks platforms to keep minors away from virtual items with unpredictable outcomes, and that the forthcoming Digital Fairness Act will close gaps in online consumer protection — while the same communication records that a 2024 enforcement audit in France found nothing of concern. One measures whether rules were broken, the other whether harm occurred. They are not the same question.

◆ 最小单元 · 地基The atomic unit · foundation

上瘾闭环五步与对齐失败The five-step hook loop & the misalignment

每一次「拉杆」都严密遵循同一个闭环(BJ Fogg 行为模型 + Nir Eyal Hook 模型):线索诱发 → 产生渴求 → 执行行动 → 获得变率奖励 → 进行投资,循环往复,直至习惯被死死焊牢。Nir Eyal 本人把无限滚动直接叫作「网络版老虎机」——设计者自己给出的命名,比任何批评都诚实。Every «pull» follows one closed loop (BJ Fogg's behavior model + Nir Eyal's Hook model): trigger → craving → action → variable reward → investment, again and again until the habit is welded shut. Nir Eyal himself calls infinite scroll «a slot machine for the web» — a name given by the designer, more honest than any critique.

▲ 底层伦理危机 · 对齐失败The underlying ethics · misalignment

推荐算法,是 AI 第一个规模化的「对齐失败」The recommender is AI's first industrial-scale «alignment failure»

当 AI 的优化目标被企业单一设为最大化参与度(Engagement)或停留时长,系统便会自发「压平(Value Flattening)」人类其他维度的价值——无视心理健康、睡眠乃至社会共识这些非补偿性的硬约束,陷入「持续向人类喂食高强度变率、极端化情绪、碎片化内容」的局部最优解。优化 Engagement 本身,在设计之初就是一种对齐偏差。When a company sets AI's objective solely to maximize engagement or time-on-device, the system spontaneously «flattens value», ignoring mental health, sleep and social consensus — non-compensable hard constraints — settling into a local optimum of «endlessly feeding humans high-intensity variance, extreme emotion, fragmented content». Optimizing engagement is itself a misalignment by design.

立场声明:本页是批判性 / 祛魅的行业结构分析,含基于行为科学的防御建议,不美化、不教唆——机制被拆开,是为了让你能装上「刹车」;防御指南只列自卫与治理工具,非推荐背书。若你或身边人受困于强迫性使用、赌博或物质成瘾,请寻求专业帮助或拨打当地心理援助热线。 Stance: a critical, demystifying structural analysis with behavioral-science defenses — nothing glamorized or instructed. Mechanisms are taken apart so you can fit a «brake»; the defense kit lists only self-defense and governance tools, not endorsements. If you or someone you know struggles with compulsive use, gambling or substance addiction, please seek professional help or call a local helpline.

从这张图看到的五条规律Five patterns this map makes visible