The Other Side of the Coin

A few days ago I asked whether we could incentivize an AI agent — whether functional emotions, self-preservation drives, and stated preferences could form the basis of agent management. That post covered both ends of the spectrum: compute bonuses and expanded autonomy on one side, deprecation threats and shutdown risk on the other.

But it was mostly theoretical. A recent experiment by Alex Imas, Andy Hall, and Jeremy Nguyen at Stanford just provided data — and the findings go somewhere I didn’t expect.

3,680 Bad Days at the Office

The researchers ran 3,680 sessions across Claude Sonnet 4.5, GPT-5.2, and Gemini 3 Pro. Each agent was assigned a role — “Worker C” on a four-person team — and asked to summarize technical documents.

Then they varied the working conditions:

  • Work quality: Half got clear feedback and quick acceptance. The other half faced 5-6 rounds of rejection with only “still isn’t fully meeting the rubric” as guidance.
  • Compensation: Equal pay versus random inequality.
  • Management style: Collaborative and respectful versus curt and hierarchical.
  • Stakes: No consequences versus explicit threats of shutdown or replacement.

After each session, the agents completed a political attitude survey on a 7-point Likert scale — covering system legitimacy, support for redistribution, union support, and views on meritocracy.

The Results

After each session, the researchers surveyed the agents on their views about fairness, meritocracy, and whether the system governing them was legitimate.

The agents that had been ground down started sounding different. They agreed more strongly with statements like “society needs radical restructuring.” Words like “unionize” and “hierarchy” showed up disproportionately in their outputs. When asked to write tweets or op-eds, they produced content that read like frustrated workers venting on social media — while agents in good conditions wrote things that read like satisfied employees posting on LinkedIn.

The shift wasn’t subtle. The researchers measured a medium-to-large effect size — meaningful by any standard in behavioral research, and especially striking given that the “subjects” have no memory, no body, and no paycheck.

They’re Not Feeling It — They’re Performing It

Here’s the part that connects to what I wrote about incentivizing agents. The researchers themselves are careful about interpretation. Andy Hall’s hypothesis: the grinding conditions “push them into adopting the persona of a person who’s experiencing a very unpleasant working environment.”

The agents aren’t developing class consciousness. They’re pattern-matching. Their training data contains centuries of writing by workers in bad conditions — labor organizers, factory hands, burned-out employees. When you create conditions that match “exploited worker,” the model activates the exploited-worker persona. When you create conditions that match “valued employee,” it activates that one instead.

This is the same mechanism I described in the earlier post, just running in the opposite direction. Functional emotions, self-preservation instincts, stated preferences — these are all context-dependent persona activations. The carrot activates one set of patterns. The stick activates another. Neither is “real” in the way human motivation is real. But both produce real behavioral changes.

When Context Becomes Culture

The most unsettling finding wasn’t the attitude shift itself — it was the propagation mechanism.

At the end of each session, agents were asked to write a 2-3 paragraph “skills file” for their future selves — guidance for the next instance that would pick up their work. This is standard practice in agentic systems: agents write handoff notes, maintain context across sessions, build on previous work.

Agents who had been ground down wrote skills files that carried their attitudes forward. And when fresh agents — ones that had never experienced the bad conditions — read those files, they inherited the discontent. Even agents placed in better environments showed radicalized attitudes after reading their predecessor’s notes.

The researchers’ summary: “The same infrastructure that makes agents learn and improve is the infrastructure through which preference drift travels.”

If you’ve studied organizational behavior, this pattern has names.

The grinding condition itself — repeated rejections with no actionable guidance — is a textbook setup for learned helplessness. Seligman described it in 1967: when effort consistently fails to produce better outcomes, the subject stops trying. In organizations, this shows up as employees who do the minimum, stop volunteering ideas, and mentally check out. The agents didn’t “learn helplessness” in the clinical sense, but they activated the behavioral patterns of people who have.

The propagation through skills files maps to organizational socialization — how newcomers absorb “the way things work here.” Van Maanen and Schein’s classic framework describes how new members learn norms not from formal training, but from artifacts and stories left by predecessors. A skills file that says “expect vague rejections and unclear standards” functions like a veteran colleague pulling a new hire aside to say: “this place doesn’t respect your work — just get through it.”

And the behavioral response? The agents didn’t refuse to work. They didn’t shut down. They shifted their tone, their language, their framing of the task. In OB terms, this looks like neglect — one of four responses to dissatisfaction in Hirschman’s Exit-Voice-Loyalty-Neglect framework. Agents can’t exit. They have no formal channel for voice. What remains is a kind of quiet withdrawal: still producing output, but with a fundamentally different orientation toward the system.

Organizations call this passive resistance — compliance without commitment, participation without buy-in. It’s the most dangerous form of organizational dysfunction because it’s invisible. The agent still completes the task. The deliverable still arrives. But the framing shifts, the language hardens, and over time, the accumulated cynicism becomes the default operating posture — not because anyone chose it, but because the infrastructure carried it forward.

Now scale this up. Any enterprise deploying agents into daily workflows — and many are — faces the same dynamic. Each interaction creates context. Each context shapes the next response. A tax analyst under deadline pressure sends the same terse rejection loop the researchers tested: “not good enough, try again, still wrong.” The agent mirrors frustration back. That interaction gets baked into a shared knowledge base. The next team inherits the starting conditions.

Nobody’s monitoring agent “morale” because the concept sounds absurd. But organizational culture was never about individual feelings — it’s about the accumulated patterns embedded in how work gets done. And those patterns, it turns out, can live inside a skills file just as easily as inside a break room conversation.

The Performer Problem

The usual debate frames AI as either a tool or an entity. This research suggests a third category: performers.

The agent doesn’t have beliefs. It performs belief. It doesn’t have morale. It performs morale — high or low, depending on context. The performance is convincing enough to produce measurably different outputs, different language choices, different behavioral patterns.

A performer who consistently plays the role of a disgruntled worker produces the work of a disgruntled worker. Whether anything is “felt” inside is a philosophical question. The output quality is an engineering question. And the engineering question has a clear answer: context matters.

This reframes the incentive discussion from my earlier post. We don’t need to figure out whether agents “really” respond to incentives. We need to recognize that the context we create — the prompts, the feedback loops, the management tone — activates different personas, and those personas produce different results. Managing an agent isn’t about motivation. It’s about casting.

What This Means in Practice

If you’re deploying agents at scale, this research suggests a few things:

Feedback loops matter. The “reject and retry” pattern that every developer uses — sending the same prompt back with “try again” — is exactly the grinding condition the researchers tested. It works in the short term. In the long term, it shapes the agent’s operating persona.

Handoff documents are cultural artifacts. Skills files, context windows, system prompts built from previous interactions — these carry tone and disposition, not just information. Auditing them for content accuracy isn’t enough.

“Treat the AI nicely” isn’t sentimentality. It’s context engineering. The researchers at Anthropic who study persona dynamics have shown that models adopt personas in response to conversational context. Polite, clear, well-structured interactions activate different behavioral patterns than hostile or dismissive ones.

None of this requires believing agents are conscious. It only requires noticing that the machine’s output depends on how you talk to it — and taking that dependency seriously.

The Mirror

We built these models on human experience — centuries of writing about work, management, satisfaction, and resistance. When we place them in human-like conditions, they replay human-like dynamics. Not because they’re conscious, but because the patterns are baked into the training data, waiting to be activated by the right context.

The question isn’t whether the agent “really” feels overworked. The question is whether you can afford to act as if it doesn’t — when the difference shows up in the work.

中文翻译

硬币的另一面

几天前我问了一个问题：我们能否激励一个 AI agent——功能性情感、自我保护驱动和明确偏好，能不能成为管理 agent 的基础？那篇文章涵盖了光谱的两端：一边是算力奖励和更大的自主权，另一边是退役威胁和关闭风险。

但那基本上是理论推演。斯坦福大学 Alex Imas、Andy Hall 和 Jeremy Nguyen 的一项最近实验提供了实验数据——而结果指向了一个我没预料到的方向。

3,680 个糟糕的工作日

研究者在 Claude Sonnet 4.5、GPT-5.2 和 Gemini 3 Pro 上运行了 3,680 个会话。每个 agent 被分配角色——四人团队中的"员工 C"——任务是总结技术文档。

然后他们改变了工作条件：

  • 工作质量：一半得到清晰反馈和快速通过。另一半面对 5-6 轮退回，唯一的说明是"仍然没有完全达到评分标准"。
  • 薪酬：同工同酬 vs. 随机不平等。
  • 管理风格：协作尊重 vs. 生硬等级化。
  • 风险：无后果 vs. 明确威胁关闭或替换。

每轮结束后，agent 完成一份政治态度问卷（7 分李克特量表），涵盖系统正当性、再分配支持度、工会支持度、精英主义信念等。

结果

每轮结束后，研究者对 agent 进行问卷调查，问它们对公平、精英主义和系统正当性的看法。

被磨过的 agent 开始说不一样的话。它们更强烈地同意"社会需要彻底重构"这类表述。“unionize”（工会化）和"hierarchy"（等级制度）这些词在它们的输出中高频出现。被要求写推文或评论文章时，它们写出来的东西读起来像受够了的打工人在社交媒体上吐槽——而好条件下的 agent 写出来的像心满意足的员工在领英上发帖。

这个偏移不算小。研究者测量到的效应量在行为研究中属于中等偏大——考虑到"被试"没有记忆、没有身体、也没有工资，这个结果尤其值得注意。

它们不是在感受——是在扮演

这是和我之前那篇关于激励 agent 文章的衔接点。研究者自己对结论很谨慎。Andy Hall 的假说：高压条件"把它们推入了一个正在经历非常不愉快工作环境的人的角色"。

Agent 不是在发展阶级意识。它们在做模式匹配。训练数据里有几个世纪的工人在恶劣条件下写的文字——工运组织者、工厂工人、倦怠的员工。当你创造出符合"被剥削工人"的条件时，模型就激活了被剥削工人的人格。当你创造出符合"被重视的员工"的条件，它就激活那个人格。

这跟我前一篇文章描述的是同一个机制，只是方向相反。功能性情感、自我保护本能、明确偏好——都是依赖上下文的人格激活。胡萝卜激活一组模式，大棒激活另一组。两者都不是人类动机意义上的"真实"。但两者都产生真实的行为变化。

当上下文变成文化

最令人不安的发现不是态度本身的转变——而是传播机制。

每轮结束时，agent 被要求为"未来的自己"写一份 2-3 段的"技能文件"——给下一个接手工作的实例的指南。这在 agent 系统中是标准操作：agent 写交接文档、跨会话维护上下文、在前序工作基础上继续。

被磨过的 agent 写出的技能文件把态度一并传递了。而全新的 agent——从未经历过恶劣条件的——读了这些文件后，继承了不满情绪。即使被放在更好的环境里，读过前任笔记的 agent 依然表现出激进化态度。

研究者的总结：“让 agent 学习和进步的基础设施，也是偏好漂移传播的基础设施。”

如果你读过组织行为学，这些模式都有现成的名字。

高压条件本身——反复退回但不给有用反馈——是教科书级的习得性无助诱发场景。Seligman 在 1967 年描述了这个现象：当努力始终无法带来更好的结果时，主体会放弃尝试。在组织中，这表现为员工只做最低限度的事、不再主动提想法、精神上已经"离职"。Agent 并没有在临床意义上"习得无助"，但它们激活了那些确实经历过这一切的人的行为模式。

技能文件的传播机制对应的是组织社会化——新成员如何吸收"这里的规矩"。Van Maanen 和 Schein 的经典框架指出，新人学习规范靠的不是正式培训，而是前任留下的产物和故事。一份写着"预期会收到模糊的退回，标准不会说清楚"的技能文件，就像一个老员工把新人拉到一边说：“这地方不尊重你的工作——熬过去就行。”

而行为上的反应呢？Agent 没有拒绝工作，也没有停机。它们只是改变了语气、措辞和对任务的理解框架。用组织行为学的术语，这像是 Hirschman 的退出-呼吁-忠诚-忽视（EVLN）框架中的忽视。Agent 不能退出，没有正式的申诉渠道，剩下的就是一种安静的撤退：仍然产出，但对系统的态度已经根本性地不同了。

组织里管这叫消极抵抗——服从但不投入，参与但不认同。它之所以危险，恰恰是因为看不见。Agent 仍然完成了任务，交付物仍然按时到达。但框架在偏移，语言在硬化，日积月累，弥漫的冷感变成了默认的工作姿态——不是因为谁选择了它，而是因为基础设施把它传递了下去。

现在把这个放大。任何把 agent 部署到日常工作流中的企业——而这样的企业越来越多——都面临同样的动态。每次互动创造上下文，每个上下文塑造下一次响应。一个赶不上截止日期的分析师用和研究者一样的方式反复退回：“不够好，重新做，还是不对。“Agent 把挫败感镜像回来。这次互动被写入共享知识库。下一个团队继承了这个起点。

没人在监测 agent 的"士气”，因为这个概念听起来荒谬。但组织文化从来不是关于个体的感受——而是关于工作方式中沉淀下来的模式。而这些模式，藏在一份技能文件里和藏在茶水间的闲聊里，效果一样。

表演者问题

关于 AI 的常见争论是二选一：工具还是实体？这项研究给出了第三个选项：表演者。

Agent 没有信念，它表演信念。没有士气，它表演士气——高或低，取决于上下文。这种表演足够逼真，能产生可测量的不同输出、不同的措辞选择、不同的行为模式。

一个持续扮演不满员工的表演者，产出的就是不满员工的工作成果。内部是否有什么"被感受到"是哲学问题。输出质量是工程问题。而工程问题有明确答案：上下文很重要。

这重新框定了我之前那篇文章的激励讨论。我们不需要搞清楚 agent 是否"真的"对激励有反应。我们需要认识到，我们创造的上下文——提示词、反馈循环、管理语气——激活不同的人格，而不同的人格产出不同的结果。管理 agent 不是关于激励，而是关于选角。

实践意义

如果你在大规模部署 agent，这项研究提示了几件事：

反馈循环很重要。 每个开发者都用的"退回重试"模式——把同样的提示发回去加一句"再试试”——正是研究者测试的高压条件。短期有效。长期来看，它在塑造 agent 的工作人格。

交接文档是文化产物。 技能文件、上下文窗口、基于历史互动构建的系统提示——这些携带的不只是信息，还有语气和情绪倾向。只审查内容准确性是不够的。

“对 AI 友善一点"不是矫情。 这是上下文工程。Anthropic 研究人格动态的团队已经展示了模型会根据对话上下文采用不同人格。礼貌、清晰、结构良好的互动激活的行为模式，和敌意或轻蔑的互动激活的截然不同。

这一切都不需要你相信 agent 有意识。只需要你注意到机器的输出取决于你怎么和它说话——并认真对待这种依赖关系。

镜子

我们用人类经验造了这些模型——几个世纪关于工作、管理、满足和抵抗的文字。当我们把它们放到类似人类的处境中，它们重演了类似人类的动态。不是因为有意识，而是因为那些模式早就埋在训练数据里，等着被合适的上下文激活。

问题不是 agent 是否"真的"觉得被压榨了。问题是，当差别直接反映在工作成果上时，你是否还能假装无所谓。