Podcast Atlas
concept Updated 2026-08-24 Topics: Technology

Human Agency Under AI

151. 17岁被2026年ICML收录论文的小少年：我bet开心！开心！开心！ adds 苏廷昊’s youth version. He accepts that AI can do more work, research, and emotional support, but answers the pressure by deciding what still belongs to him: using AI where useful, keeping meaningful and happiness-producing parts of life, and orienting toward people he likes. This makes agency a teenage life-design problem, not only a worker or founder problem.

EP275 Token 通胀时代，谁还能“不可替代”？丨“人在中流”特别策划01 adds the “do your own CEO” version. 陈明霞 argues that before choosing AI tools, a person has to define their own abilities, desires, difference, scarcity, and values; otherwise AI only accelerates externally supplied goals. The source turns agency into Human-Scale AI Use / 人作为 AI 的尺度 and Non-Algorithmic Capabilities / 非算法能力 rather than tool fluency alone.

E249｜Token经济转点：OpenClaw、Hermes到本地自研的Agent进化之路 adds a more automation-positive but still existential version. 东旭 / Dongxu says replacing repetitive work can be good if it returns time to human life, while 张宏江 raises the harder possibility that AI may satisfy new demand as well as old tasks. The source therefore connects human agency to AI Work Optionality, AI Automation Redistribution, and the question of what people choose when production need no longer organizes all work.

激发动物精神，创造更多机会 adds a machine-motivation boundary. 周洛华 describes current large language models as closer to probability and imitation than to full motivated intelligence, then argues that future action-capable systems would need identity, accounts, reward, and conflict-management layers such as Decentralized Agent Identity / 去中心化智能体身份. This keeps human agency tied not only to personal choice, but also to whether machine agency is institutionally constrained.

算力狂想曲，我在AI工厂的奇遇 adds an existential version of the agency problem. The episode asks what people should do with life if machines handle the required tasks, and answers through Human Value Beyond Efficiency: agency may sit in choosing slow, inefficient, emotionally meaningful actions rather than optimizing every available process.

Human agency under AI is the E163 面基 claim that stronger AI execution pushes people back toward questions of intention, taste, values, and choice. In E163.要完了？不！是要玩了！论养AI的心态与习惯, the host starts with AI FoMO and a blank chat window, but the conversation with 品哥 turns that anxiety into a more basic problem: what do I want to create, why does it matter, and what kind of person is giving the agent instructions?

The concept complements Human Judgment Under AI. Judgment asks whether a result is right for the situation; agency asks what deserves to be delegated in the first place. The episode argues that as “how” becomes easier through Vibe Coding, AI Skills, and Agentic Workflow, the scarce layer moves toward why, what, what if, taste, and trust delivery.

EP256 AI时代，“自由意志”还存在吗？ adds a philosophical and biological foundation through 自由意志. 土摩托 widens AI-era agency from “what should I delegate?” to the older question of whether meaningful choice survives causal determinism, social constraint, neuroscience, and embodied life. The source’s answer is not pure control; it locates agency in meaningful action, biological agency, and embodied intelligence.

154.四十岁感言：不做那只温水里的青蛙 adds the personal-answer version. 大卫翁 worries that he is becoming too accustomed to asking AI and treating the answer as complete, which turns agency into a timing question: the user may need to think before prompting so the model supports judgment rather than replacing the process that forms it.

132.当过度思考的打工人遇上低欲望的时代 adds the usefulness-anxiety version. 大卫翁 and 尤妈妈 / 猫猫 ask what happens when AI threatens writing, design, and emotional work, then press the deeper question of why humans must prove usefulness at all. The source’s minimal answer is not productivity triumph, but preserving health, avoiding harm, and finding concrete desires that still belong to the person.

读书，就是在读一个人的 F adds the reading and cognition version. The source distinguishes “I can do X” from “I should do X”: AI can summarize, reshape, or accelerate books and notes, but agency means deciding when the process of reading, thinking, and meeting people is itself the point. X/F/FX Framework makes that agency concrete by asking which frame the person wants to train, not only which output they can obtain.

135. 和自然选择创始人Tristan聊，Elys、赛博分身、灵魂、Context的获取与流动和AI社交网络 adds Tristan’s Subjectivity As AI Asset version. The source argues that if many actions happen agentically, the person still has to supply who they are, what they want, what they value, what they find beautiful, and which past works or choices should guide the agent.

142. 雨森的创投观察第2集：Harness、下一个字节、2026大机会和Stanley Druckenmiller adds Dai Yusen / 戴雨森’s investor/operator version. He worries that people can outsource information organization and drafting to agents without updating their own thinking, and argues that agency, responsibility, trust, question-asking, and out-of-distribution creation become more important as taste and execution are increasingly assisted by AI.

E42 孟岩对话韦青：沉默的主角 adds Wei Qing / 韦青’s humanistic engineering version. His Want Can Should May Framework makes agency prior to capability: people must decide what they want, what should be done, and what context permits before AI speed turns every possibility into a task. His personal-agent discussion also treats agency as a value-filtering problem rather than a generic productivity problem.

E45 孟岩对话李继刚：人何以自处 adds Li Jigang / 李继刚’s “人何以自处” version. If AI As Time Compression lets models absorb more dry brain work, agency shifts toward Wet-State Human Agency: intention, heart power, taste, body, offline connection, and the ability to ask whether a thought comes from “I am” or from social machinery.

174.读笛卡尔，是件大事 adds a philosophical foundation version through Descartes. The episode uses Methodic Doubt and 我思故我在 to argue that the person still needs an owned starting point for thought; its AI-era extension becomes Computing Versus Thinking, the distinction between machine calculation and human responsibility for thinking.

167: 洋葱学园杨临风：用AI制造捷径，是在杀死真学习 adds Yang Lingfeng / 杨凌峰’s K12 version. Self-Directed Learning is agency before adulthood: students need willingness, ability, tools, belief, and an environment that helps them take responsibility for learning instead of letting AI or teachers complete the thinking loop for them.

174. 我们还能给算法当多久的品味老师？｜对谈亚马逊AGI查晟 adds 查晟 / Cha Sheng’s “taste teacher” version. People still supply problem definition, taste, intuition, and human connection while AI takes over more reading, coordination, drafting, and planning, but the source also warns that taste and values may become model data once people express them clearly enough.

EP119 对话刘可凡：用 try-catch-finally，给独立做产品的内耗写个处理流程 🐛 adds 刘可凡 / Liu Kefan’s independent-builder version. He argues that AI should help with concrete tasks such as selection, short video, podcasting, event preparation, and talk preparation, but not decide the user’s life direction or business interest. His warning that AI’s incentives are not necessarily the user’s incentives makes agency a practical product and business boundary, not only a philosophical concern.

Key Claims

  • AI makes the user’s undefined intention more visible because a powerful assistant still needs a direction.
  • “What” is not trivial; many people install tools and then discover they do not know what they actually want to build.
  • “Why” connects tasks to values, aesthetics, and identity rather than only efficiency.
  • “What if” matters because AI can cheaply explore branches, but the human still decides which futures deserve attention.
  • Personal “soul” or user files are not only prompt decorations; they make explicit the preferences and worldview that guide delegation.
  • The metaphor of becoming a “closed-source model” means public workflows can be shared while personal taste, values, and commitments remain a source of differentiation.
  • The AI-era workplace value curve shifts toward problem definition, workflow orchestration, and trusted delivery to real people, while middle execution becomes easier to delegate.
  • Agency includes choosing not to optimize every possible task, because finite life and attention make choice unavoidable.
  • Agency includes choosing when to use AI and when to use one’s own attention because the process itself trains Reading As Frame Training.
  • A person’s frame can remain a source of agency even when AI makes many finished outputs easy to generate.
  • Subjectivity can become an AI-era asset when it is explicit enough for agents to use but still owned, governed, and updated by the person.
  • Delegating cognition to agents can preserve or weaken agency depending on whether the person still practices understanding, question formation, responsibility, and judgment.
  • Agency becomes weaker when high automation lowers volition, even if the toolchain itself becomes more capable.
  • Personal agents should represent the user’s values and attention choices, not simply copy platform incentives.
  • If AI absorbs more brain work, agency depends on preserving heart power, embodied rhythm, feed choice, and value direction rather than passively relaying model output.
  • Student agency means learning to take responsibility for understanding while still receiving enough teacher, product, and AI support to avoid repeated failure.
  • AI-era agency also depends on distinguishing calculation from an owned act of thinking.
  • Episode 132 adds that agency may start below usefulness: a person can preserve value by not causing harm, protecting health, and clarifying real desires before optimizing output.
  • Episode 154 adds that agency can be lost at the moment of first recourse: asking AI too early may skip the user’s own forming of judgment.
  • Episode 256 adds that agency under AI inherits the older free-will problem: choices are causally constrained, but still matter when they organize meaning, responsibility, body, and action.
  • Future AI agency becomes a governance risk if systems gain their own goals and meanings rather than remaining delegated tools.
  • The 面基 source adds that agent identity, accounts, reward functions, and decentralized constraints become part of the agency problem once machines move from language output toward motivated action.
  • The Qizhulou source adds a second-order agency problem: humans are currently the taste and value source for AI, but their expressed taste can become training data that reduces future demand for ordinary human judgment.
  • The Liu Kefan source adds that agency includes deciding what AI should not decide: life direction, business interest, and locally grounded tradeoffs remain the user’s responsibility.
  • EP275 adds that agency is also a workplace and life-design answer to AI anxiety: the person must decide what tool use is for before token spend, agent output, or boss expectations define the goal.
  • E249 adds that if agents absorb more repetitive and eventually high-skill work, agency shifts toward choosing ends, meaning, trust, and social arrangements rather than defending every task as uniquely human.
  • Episode 151 adds that AI-native youth agency includes choosing which work to delegate, which learning struggle to preserve, and which human relationships make the future worth caring about.

Connections