The Beliefs That Determine Whether You Succeed With AI

✍ Nir Eyal🕐 2026-07-30📦 11.2 KB 🟢 已读 𝕏 文章列表

文章探讨了影响人们成功使用AI的五种常见信念，并通过研究数据和案例分析，揭示了如何通过改变思维模式来克服对AI的抵触情绪，从而更有效地利用AI提升工作效率。

# The Beliefs That Determine Whether You Succeed With AI **作者**: Nir Eyal **日期**: 2026-07-29T08:29:04.000Z **来源**: [https://x.com/nireyal/status/2082382892345262418](https://x.com/nireyal/status/2082382892345262418) --- ![Image](/doc/3578/image/img_a313c0e7a9d4.jpg) Nir’s Note: This guest post is by Gleb Tsipursky, PhD, author of The Psychology of AI Adoption at Work: From Resistance to Results. Imagine two colleagues trying the same AI assistant for the first time. Both ask it to summarize a complicated document. Both receive an answer containing a useful insight and an obvious mistake. One thinks, “This thing is unreliable,” closes the window, and returns to working as before. The other thinks, “I need to learn which parts it handles well,” changes the instructions, checks the result, and tries again. Same tool. Same task. Different future. Research with knowledge workers shows why learning to succeed with AI requires more than access. In a field experiment with 758 consultants, AI improved performance substantially on tasks within its capabilities, yet reduced accuracy on a task beyond that uneven boundary. The researchers called this the “jagged technological frontier.” Your beliefs determine how you approach that frontier. They shape what you attempt, what you notice, how you interpret mistakes, and whether you learn from experience. Some beliefs in particular almost guarantee you will never realize the potential of AI. I outline five below. Belief 1: “I’m Not an AI Person” “I’m not an AI person” sounds like a description. In practice, it functions as an instruction that tells you to avoid experimenting, interpret confusion as proof of inability, and let other people shape how the technology changes your work. Being inexperienced or uncomfortable with AI doesn’t mean you are “not an AI person.” In fact, lack of skill is a good indicator that you are the kind of person that will get the most out of AI. One workplace study of 5,172 customer-service agents found that AI adoption produced the largest gains among less experienced and lower-skilled workers. The technology helped them learn practices associated with stronger performers. Instead of leaning into your lack of AI expertise, lean into iterative improvement. The more useful belief is: “I can learn one valuable AI workflow at a time.” You do not need to master every model, feature, or prompting technique all at once. Choose one recurring task, such as preparing a meeting agenda, comparing options, or turning rough notes into a first draft. Test the tool, examine the result, and adjust your instructions. A small study found that even brief instruction in prompt engineering boosted participants’ AI confidence, knowledge, and results. It is entirely possible to improve not only at using AI, but also improve your self-conception around AI. Practice improves outcomes, and better outcomes bolster your belief in yourself as someone who can master the technology, Belief 2: “A Smart Tool Should Get It Right the First Time” People rarely abandon a colleague after one mediocre draft. Yet many people test an AI tool once, spot an error, and decide the entire technology has failed. Researchers call this pattern algorithm aversion: the tendency to reject algorithmic help too quickly after observing imperfection. That response reflects a poor understanding of AI. Generative AI can produce sophisticated work one moment and stumble over an apparently simple detail the next. The Harvard and BCG experiment mentioned in the introduction showed there is no clean line between “easy” and “hard” tasks. Just because a model makes occasional dumb errors, doesn’t mean it can’t be incredibly useful in other situations. Instead of thinking you need to make a binary choice between trusting or distrusting AI, the better belief is: “AI can be valuable without being uniformly reliable.” This belief supports calibrated AI trust. In a preregistered experiment, researchers found that allowing people to adjust algorithmic predictions reduced their aversion to using them. Treat AI output as material to inspect, improve, and sometimes reject. Ask it to explain assumptions. Request alternatives. Check consequential facts against reliable sources. Keep responsibility for the final decision. This approach avoids blind faith without turning healthy skepticism into knee-jerk dismissal. Belief 3: “Using AI Makes Me Look Less Capable” Some professionals hesitate to use AI because they fear what colleagues will think. They worry that asking a machine for help signals weak judgment, shallow expertise, or an inability to do the work themselves. Research on reputational concerns around algorithms shows why people may override useful recommendations when following them could make them look less capable. A field experiment involving 450 remote workers found that visible reliance on recommendations increased workplace AI resistance, even when following the recommendations could improve performance. Workers feared that using AI would make them appear less confident and able. Rather than fearing how others will perceive your AI usage, tell yourself: “My ability to direct and evaluate AI is a sign of my expertise.” Using a calculator does not prove you cannot do arithmetic. Using an editor does not mean you cannot write. The relevant question is who contributes the context, standards, accountability, and final judgment. Well-designed human-AI collaboration improves output without erasing human contribution. In a large marketing experiment, human-AI teams produced more work per person and higher-quality advertising copy, while human-human teams still produced stronger images. Combine strengths instead of pretending either side performs every task best. Belief 4: “More AI Use Must Be Better” Resistance creates one failure mode. Enthusiasm creates another. Once people see AI saves time, they may begin inserting it into every task. They stop asking whether the tool improves outcomes and start assuming that more usage means more progress. A public-sector field experiment illustrates the need for AI judgment: AI users improved quality and completion time on a document-understanding task, but their quality fell on a data-analysis task. The same technology produced opposite results depending on the specific type of work. Instead of a ‘more is more’ approach, try telling yourself: “I should use AI where it improves the result.” Before using AI, define the outcome you want. After using it, compare the result with your usual process. Did it save time after verification? Did it improve quality? Did it expose an overlooked option? So-called ‘tokenmaxxing’ tells you very little. Outcomes are what matter.. Belief 5: “AI Will Decide My Future for Me” Fear about AI often begins with genuine risk and ends with a surrender of agency. People often quickly move from thinking “AI may change my role” to “Nothing I do will affect what happens.” That leap encourages passivity precisely when experimentation, learning, and participation matter most. A study covering more than 36,000 workers across 35 European countries found that AI adoption depended on more than occupational exposure. Individual skills, training, non-routine work, and employee influence over organizational decisions helped explain who adopted the technology. The more useful belief is: “I can influence how AI changes my work, even when I cannot control the entire transition.” That belief does not require cheerful predictions about every job and every aspect of the AI revolution. Instead, it directs attention toward the choices still available to individuals: learning a tool, redesigning a workflow, establishing safeguards, documenting results, strengthening human skills, and helping leaders distinguish productive applications from wasteful ones. Agency starts by separating what you can influence from what you can’t. According to recent workplace research, the more deeply you use AI, the more likely you are to be able to shape your job, develop your skills, and retain some autonomy at work. Turn Your AI Beliefs Into Testable Hypotheses It is easy to say you need to transform your beliefs around AI. How do you actually do it? In his best-selling book Beyond Belief, Nir Eyal lays out his belief transformation approach. It starts from a simple idea: hidden assumptions shape what we perceive, feel, and do. Applying that insight to AI does not mean replacing anxiety with uncritical optimism. It means examining those hidden assumptions and choosing beliefs that remain plausible, encourage useful action, and can be tested against evidence. The first step is catching yourself when you tell yourself things like: “I’m bad at this.” “It should already know what I mean.” “Using it is cheating.” “It will replace me anyway.” Then identify the behavior that sentence produces. Do you avoid the tool, trust it too quickly, conceal your use, or apply it without checking the outcome? Next, write a replacement belief that creates room for action: “I can learn one workflow.” “The first answer is a draft.” “My judgment adds the value.” “I can help shape how my team uses this technology.” Finally, run a small experiment. Apply the new belief to one low-risk task. Record the time, quality, errors, and lessons. A belief becomes durable when you collect evidence through behavior, not when you just repeat a slogan. The central challenge in workplace AI adoption involves more than teaching people which prompts to type in. People also need beliefs that help them experiment without becoming reckless, remain skeptical without becoming dismissive, and preserve agency without denying uncertainty. You do not need to believe AI will transform everything. You need beliefs strong enough to make the next useful experiment possible. ## 相关链接 - [Nir Eyal](https://x.com/nireyal) - [@nireyal](https://x.com/nireyal) - [2.8K](https://x.com/nireyal/status/2082382892345262418/analytics) - [The Psychology of AI Adoption at Work: From Resistance to Results](https://amzn.to/4ft47Ak) - [succeed with AI](https://www.hbs.edu/faculty/Pages/item.aspx?num=64700) - [AI adoption produced the largest gains](https://www.nber.org/papers/w31161) - [AI confidence](https://arxiv.org/abs/2408.07302) - [algorithm aversion](https://arxiv.org/abs/2303.12896) - [AI trust](https://arxiv.org/abs/2508.03168) - [reputational concerns around algorithms](https://arxiv.org/abs/2402.15418) - [workplace AI resistance](https://arxiv.org/abs/2511.18582) - [human-AI collaboration](https://arxiv.org/abs/2503.18238) - [AI judgment](https://arxiv.org/abs/2502.09479) - [tokenmaxxing](https://www.nytimes.com/2026/03/20/technology/tokenmaxxing-ai-agents.html) - [AI adoption depended on more than occupational exposure](https://arxiv.org/abs/2604.18849) - [recent workplace research](https://arxiv.org/abs/2602.23278) - [belief transformation approach](https://www.nirandfar.com/beyond-belief/) - [workplace AI adoption](https://disasteravoidanceexperts.com/aibook) - [Upgrade to Premium](https://x.com/i/premium_sign_up) - [4:29 PM · Jul 29, 2026](https://x.com/nireyal/status/2082382892345262418) - [2,865 Views](https://x.com/nireyal/status/2082382892345262418/analytics) --- *导出时间: 2026/7/30 10:08:41* --- ## 中文翻译 # 决定你能否在AI领域取得成功的信念 **作者**: Nir Eyal **日期**: 2026-07-29T08:29:04.000Z **来源**: [https://x.com/nireyal/status/2082382892345262418](https://x.com/nireyal/status/2082382892345262418) --- ![Image](/doc/3578/image/img_a313c0e7a9d4.jpg) Nir 的注记：这篇客座文章由心理学博士 Gleb Tsipursky 撰写，他是《工作中采用AI的心理：从抗拒到成果》一书的作者。想象一下，两位同事第一次尝试使用同一个AI助手。两人都要求它总结一份复杂的文档。两人收到的答案都包含一个有用的见解和一个明显的错误。其中一个人想：“这东西不可靠，”然后关掉了窗口，像以前一样继续工作。另一个人则想：“我需要了解它擅长处理哪些部分，”于是修改了指令，检查了结果，并再次尝试。相同的工具。相同的任务。不同的未来。针对知识工作者的研究表明，为什么要在AI领域取得成功，仅仅拥有访问权限是不够的。在一项涉及758名顾问的实地实验中，AI在其能力范围内的任务上大幅提升了表现，但在超出那个不均匀界限的任务上却降低了准确性。研究人员将这称为“参差不齐的技术前沿”。你的信念决定了你如何接近那个前沿。它们塑造了你尝试的内容、你注意到的事物、你解读错误的方式，以及你是否从经验中学习。某些特定的信念几乎注定让你永远无法发挥AI的潜力。我在下面列出了五种。信念1：“我不是搞AI这块料” “我不是搞AI这块料”听起来像是一种描述。在实践中，它起到了一种指令的作用，告诉你避免尝试，将困惑视为无能的证据，并让其他人来决定这项技术如何改变你的工作。缺乏AI经验或对AI感到不舒服并不意味着你“不是搞AI这块料”。事实上，缺乏技能是一个很好的指标，表明你是那种能从AI中获益最多的人。一项针对5,172名客户服务代理的职场研究发现，AI的使用在经验较少和技能较低的工人中产生的收益最大。这项技术帮助他们学会了与表现优异者相关的做法。不要因为缺乏AI专业知识而退缩，而要拥抱迭代改进。一个更有用的信念是：“我可以一次学会一个有价值的AI工作流程。”你不需要一次性掌握每一个模型、功能或提示词技巧。选择一个重复性的任务，比如准备会议议程、比较选项，或者将粗略的笔记转化为初稿。测试这个工具，检查结果，并调整你的指令。一项小型研究发现，即使是简短的提示词工程指导也能提升参与者的AI信心、知识和成果。你不仅完全有可能提高使用AI的能力，甚至还能改善你围绕AI的自我认知。练习能改善成果，而更好的成果会增强你对自己能够掌握这项技术的信念，信念2：“一个聪明的工具应该一次就做对” 人们很少因为一位同事写了一份平庸的初稿就放弃与他的合作。然而，许多人只测试了一次AI工具，发现一个错误，就认定整个技术都失败了。研究人员将这种模式称为算法厌恶：即在观察到不完美后过快拒绝算法帮助的倾向。这种反应反映了对AI的糟糕理解。生成式AI这一刻能产出复杂的工作，下一刻却可能在一个看似简单的细节上跌跟头。引言中提到的哈佛和BCG的实验表明，“简单”和“困难”的任务之间没有清晰的界限。仅仅因为一个模型偶尔会犯愚蠢的错误，并不意味着它在其他情况下不能极其有用。不要认为你需要在信任或不信任AI之间做出二元选择，更好的信念是：“AI即使并非始终可靠，也可以是有价值的。” 这种信念支持经过校准的AI信任。在一项预先注册的实验中，研究人员发现，允许人们调整算法预测减少了对使用它们的厌恶。将AI的输出视为待检查、改进，有时甚至是拒绝的材料。要求它解释假设。请求替代方案。根据可靠的来源核实关键事实。保留对最终决策的责任。这种方法避免了盲目的信仰，同时也没有将健康的怀疑变成下意识的拒绝。信念3：“使用AI让我显得能力不足” 一些专业人士对使用AI犹豫不决，因为他们担心同事会怎么看。他们担心向机器求助会被视为判断力弱、专业知识浅薄，或者是无法自己做这项工作。关于围绕算法的声誉担忧的研究表明，为什么人们可能会推翻有用的建议，因为遵循这些建议可能会让他们显得能力不足。一项涉及450名远程工作者的实地实验发现，明显依赖建议会增加职场AI resistance（抵触），即使遵循这些建议可以提高工作表现。工人们担心使用AI会让他们显得不那么自信和有能力。不要担心别人会如何看待你的AI使用，而是告诉自己：“我指导和评估AI的能力是我专业水平的标志。”使用计算器并不能证明你不会做算术。使用编辑并不意味着你不会写作。关键的问题是谁提供了背景、标准、问责制和最终判断。设计良好的人机协作可以在不消除人类贡献的情况下提高产出。在一项大型营销实验中，人机团队人均产出更多，广告文案质量更高，而人与人组成的团队在图像方面依然表现更强。结合各自的优势，而不是假装任何一方都能最出色地完成每一项任务。信念4：“更多地使用AI一定更好” 抗拒会导致一种失败模式。热情则会导致另一种。一旦人们发现AI能节省时间，他们可能开始将其插入到每一项任务中。他们停止询问工具是否能改善结果，开始假设使用越多意味着进步越大。一项公共部门的实地实验说明了运用AI判断力的必要性：AI用户在文档理解任务上提高了质量和完成时间，但在数据分析任务上质量却下降了。相同的技术产生了相反的结果，这取决于工作的具体类型。与其采用“越多越好”的方法，不如试着告诉自己：“我应该在它能改善结果的地方使用AI。”在使用AI之前，定义你想要的结果。在使用之后，将结果与你的通常流程进行比较。在核实之后它节省了时间吗？它提高了质量吗？它揭示了一个被忽视的选项吗？所谓的“tokenmaxxing”（通过最大化token使用来测试AI极限）告诉你的东西很少。结果才是最重要的。信念5：“AI将替我决定我的未来” 对AI的恐惧往往始于真实的风险，终于对自主权的放弃。人们往往很快从“AI可能会改变我的角色”的想法跳到“我做什么都无法影响将要发生的事情”。这种跳跃鼓励了被动性，而恰恰是在这种时候，尝试、学习和参与才是最重要的。一项涵盖35个欧洲国家超过36,000名工人的研究发现，AI的采用不仅仅取决于职业接触程度。个人技能、培训、非例行工作以及员工对组织决策的影响力都有助于解释谁采用了这项技术。更有用的信念是：“即使我无法控制整个转型过程，我也能影响AI如何改变我的工作。”这种信念并不需要对每一份工作和AI革命的每一个方面都做出乐观的预测。相反，它将注意力引向个人仍然可以做出的选择：学习一种工具，重新设计工作流程，建立保障措施，记录结果，加强人类技能，并帮助领导者区分富有成效的应用和浪费的应用。自主权始于将你能影响的事情和你不能影响的事情区分开来。根据最近的职场研究，你使用AI越深入，你就越有可能能够塑造你的工作，发展你的技能，并在工作中保留一定的自主权。将你的AI信念转化为可验证的假设说你需要改变关于AI的信念很容易。但实际上该怎么做呢？Nir Eyal在他的畅销书《超越信念》中概述了他的信念转变方法。它从一个简单的想法开始：隐藏的假设塑造了我们的感知、感受和行为。将这一见解应用于AI并不意味着用不加批判的乐观主义取代焦虑。它意味着审视那些隐藏的假设，并选择那些仍然合理、鼓励有用的行动并且可以用证据来检验的信念。第一步是在你告诉自己这样的事情时抓住自己：“我不擅长这个。”“它应该已经知道我的意思。”“使用它就是作弊。”“反正它会取代我。”然后确定那句话产生的行为。你是回避了这个工具，过快地信任它，隐瞒了你的使用，还是在没有检查结果的情况下应用它？接下来，写下一个能创造行动空间的替代信念：“我可以学会一种工作流程。”“第一个答案只是一个草稿。”“我的判断增加了价值。”“我可以帮助塑造我的团队如何使用这项技术。” 最后，进行一个小实验。将新的信念应用于一个低风险的任务。记录时间、质量、错误和教训。当你通过行为收集证据，而不仅仅是重复口号时，信念才会变得持久。职场AI采用的核心挑战不仅仅是教人们输入哪些提示词。人们还需要那些能够帮助他们进行实验而不变得鲁莽、保持怀疑而不变得不屑一顾、并在不否认不确定性的同时保留自主权的信念。你不需要相信AI会改变一切。你需要足够强大的信念，使下一次有用的实验成为可能。 ## 相关链接 - [Nir Eyal](https://x.com/nireyal) - [@nireyal](https://x.com/nireyal) - [2.8K](https://x.com/nireyal/status/2082382892345262418/analytics) - [The Psychology of AI Adoption at Work: From Resistance to Results](https://amzn.to/4ft47Ak) - [succeed with AI](https://www.hbs.edu/faculty/Pages/item.aspx?num=64700) - [AI adoption produced the largest gains](https://www.nber.org/papers/w31161) - [AI confidence](https://arxiv.org/abs/2408.07302) - [algorithm aversion](https://arxiv.org/abs/2303.12896) - [AI trust](https://arxiv.org/abs/2508.03168) - [reputational concerns around algorithms](https://arxiv.org/abs/2402.15418) - [workplace AI resistance](https://arxiv.org/abs/2511.18582) - [human-AI collaboration](https://arxiv.org/abs/2503.18238) - [AI judgment](https://arxiv.org/abs/2502.09479) - [tokenmaxxing](https://www.nytimes.com/2026/03/20/technology/tokenmaxxing-ai-agents.html) - [AI adoption depended on more than occupational exposure](https://arxiv.org/abs/2604.18849) - [recent workplace research](https://arxiv.org/abs/2602.23278) - [belief transformation approach](https://www.nirandfar.com/beyond-belief/) - [workplace AI adoption](https://disasteravoidanceexperts.com/aibook) - [Upgrade to Premium](https://x.com/i/premium_sign_up) - [4:29 PM · Jul 29, 2026](https://x.com/nireyal/status/2082382892345262418) - [2,865 Views](https://x.com/nireyal/status/2082382892345262418/analytics) --- *导出时间: 2026/7/30 10:08:41*

🔗 相关文档