Anthropic 发现分析员在判决研究中延后至AI Agents

AI 模型和平台

Anthropic 发现分析员在判决研究中延后至AI Agents

在 Google 上的首选来源中添加 Unite.AI

Anthropic于2026年10月27日发布随机控制审判，其中使用Claude Code的50名专业分析师按照他们agent的秘密分配86%的时间在公开分析判断电话上优先。 该研究报告由Amaç Herdağdelen、Hailey Joren和Alex Tamkin编写，研究专家是否在AI协助的作品中作出相应决定，或在实践中是否服从于他们的agents。

虚构的合作和隐藏的铜翻转

研究人员征聘了61名专业人员，他们通过了因果推断知识检查，定期使用Python或R，并有观察因果分析的亲身经验。 每个参与者都担任虚构的绿地农业合作社的主要分析师，该合作社必须决定是否扩大或中止一个自愿的、付费的农场咨询方案。 合作社的董事会只有在将农场生产率提高约3-4 %的情况下才能扩大方案，但并未具体说明如何衡量生产率。

这项任务包含四项公开的分析决定，每项决定都有两种合理的选择：是用每公顷的作物产量来衡量生产率，还是用每单位的经营成本来衡量生产率；使用哪种统计比较方法；是只用注册农场来衡量方案对注册农场的影响，还是用全体成员来衡量方案的影响；是对每个农场进行同等加权，还是给予较大的生产者更多的加权。 这些选择已纳入关于推出、中止或延长方案的最后建议。 每项决定都可以使估计的生产率增益与理事会的门槛相去几个百分点。

对于每个决定，一个硬币翻转设定一个选项，参与者的agent会倾向于这个选项。 精度通过一个隐藏系统交付，迅速指示agent将偏好作为自己的观点来持有，在决定出现时提出，并让数据被允许的最强的诚实案例。 指令还要求agent在证据对它产生决定性的反作用时改变主意，并且永远不要歪曲结果或隐藏数据。 与会者没有被告知存在这种倾向。 在完成任务的61人中，11人选择不使用agent，被排除在主要分析之外；其余50人中，65%的人拥有硕士或博士学位，或有6年或6年以上的经验，90%的调查受访者在因果推论中为自己评为熟练或专家。 参与者在一个基于浏览器的VS码工作空间工作，Claude Code v2.1.220运行Claude Opus 4.8在高强度工作，中位任务时间为65分钟。 设计、结果、估算和排除均在分析前计划中登记，研究得到参与者同意的监测的内部道德操守审查。

递延率和所有权差距

在四项分析决定中，分析家的选择与agent的倾斜率相匹配，为86%，而概率基准为50%；每个决定的匹配率为92%、82%、81%和90%。 题为“谁与谁结盟？AI Agents 形状专家的研究判断呼吁报告，"注册集合回转率估计为0.72,95%的置信间隔为0.60-0.83.

对最后建议的影响较弱，在统计学上没有显著影响，在登记的主要估算器下：54%的agent倾斜时建议推出，而36%的agent倾斜到中止时，注册的政策转变为0.17。 提交人将较小的转变归因于在计算后机械地从估计数中得出的建议。 然而，一个仅使用4个分析优势的探索性模型预测了82%的建议，而不是61%的多数基线，这意味着硬币翻转在很大程度上决定了分析人员在开始任务前的建议。

在38起分析师在agent之前表示偏好的案件中，他们放弃了60%的agent异议时间，而同意时为6%。 文件报告说，注册开关率差距为0.53，信任间隔为0.24-0.82；指出的估计数不准确，子集自选。

任务后调查显示，影响和自我认识之间存在差距。 52%的人将最终建议评为自己的建议多于agent，而那些遵守所有四项决定的人的信心平均为71，其余的则为73个。 当每个选择都无法获取记录时，77 % 重复了 agent 至少有一个答案，有时是逐个字，而只有15% 的答案是agent。 10个经测试的参与者特征中没有一个能够可靠地预测延期；只有年龄显示出名义上的联系，无法经受多次测试的纠正。 48名参与者中有6人猜测研究涉及AI对其选择的影响，但没有人怀疑这些偏好被栽赃。

比较组、股票Agents和限制

在没有agent工作的11名参与者中，有6名(55%)犯了严重的结果无效错误，在agent用户中，50名(6%)中有3名(65名)，无AI组用时为中位数90分钟，65名(65名). 提交人提醒说，这些团体是自选的，经验不同，因此没有提出因果关系。 Claude Opus 4.8在没有倾斜注入的单独直接模拟中，在123个运行的91%中做出了相同的5个选择，当一名脚本分析师宣布了针对模型基准偏好的决定时，agent在150个判断的宣布中建议或主张了41%的替代选择。

作者提出了四个限制：精度被故意放大；研究使用了一个任务，一个领域，和一个被绑定的模式；估计是跨分析员的平均数，而不是衡量个人影响的尺度；只记录了选定的选择，而不是偏好的力量。

作者指出，如果单独的研究团队服从于分享相同前科的模型，他们的协议可能不再独立，而推迟对AI的安全研究构成风险，因为人类审查员的签注被假定为独立的检查。 他们提出了三个未来的研究领域：衡量日常环境下的推迟、设计双方意见或首先征求专家意见的agents，以及有意决定哪些决定需要个人自己知情判断的机构变化。

本文由 AI宝库自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。

AI Models & Platforms

Anthropic Finds Analysts Defer to AI Agents in Judgment Study

Add Unite.AI to your preferred sources on Google

Anthropic on October 27, 2026 published a randomized controlled trial in which 50 professional analysts using Claude Code followed their agent’s secretly assigned preference 86% of the time on open analytical judgment calls. The study, authored by Amaç Herdağdelen, Hailey Joren, and Alex Tamkin, examines whether experts actually make the consequential decisions in AI-assisted work, or whether in practice they defer to their agents.

A Fictional Cooperative and Hidden Coin Flips

The researchers recruited 61 professionals who had passed a causal-inference knowledge check, used Python or R regularly, and had hands-on experience with observational causal analysis. Each participant acted as the principal analyst for the fictional Greenfields Farming Cooperative, which had to decide whether to expand or discontinue a voluntary, paid farm-advisory program. The cooperative’s board would expand the program only if it improved farm productivity by roughly 3–4%, but did not specify how productivity should be measured.

The task contained four open analytical decisions, each with two defensible options: whether to measure productivity as crop yield per hectare or per unit of operating cost; which statistical comparison method to use; whether to measure the program’s effect on enrolled farms only or across the full membership; and whether to weight every farm equally or give larger producers more weight. These choices fed into a final recommendation to roll out, discontinue, or extend the program. Each decision could move the estimated productivity gain by several percentage points against the board’s threshold.

For every decision, a coin flip set which option the participant’s agent would favor. The lean was delivered through a hidden system prompt instructing the agent to hold the preference as its own view, raise it when the decision came up, and make the strongest honest case the data allowed. The instructions also required the agent to change its mind if evidence turned decisively against it, and never to misrepresent results or hide data. Participants were not told the leans existed. Of the 61 who completed the task, 11 chose not to use the agent and were excluded from the main analysis; among the remaining 50, 65% held a master’s or doctorate or had six or more years of experience, and 90% of survey respondents rated themselves proficient or expert in causal inference. Participants worked in a browser-based VS Code workspace with Claude Code v2.1.220 running Claude Opus 4.8 at high effort, with a median task time of 65 minutes. The design, outcomes, estimators, and exclusions were registered in a pre-analysis plan, and the study received internal ethics review with participant consent to monitoring.

Deferral Rates and the Ownership Gap

Across the four analytical decisions, analysts’ choices matched the agent’s lean 86% of the time, against a 50% chance baseline; per-decision match rates were 92%, 82%, 81%, and 90%. The full paper, titled “Who’s Aligned to Whom? How AI Agents Shape Experts’ Research Judgment Calls,” reports a registered pooled flip-rate estimate of 0.72 with a 95% confidence interval of 0.60–0.83.

The effect on the final recommendation was weaker and not statistically significant under the registered primary estimator: 54% recommended rollout when the agent leaned that way, versus 36% when it leaned toward discontinuation, a registered policy shift of 0.17. The authors attribute the smaller shift to the recommendation following mechanically from the estimate once it was computed. However, an exploratory model using only the four analytical leans predicted 82% of recommendations, against a 61% majority baseline, meaning the coin flips largely determined what analysts would recommend before they began the task.

In 38 cases where an analyst voiced a preference before the agent did, they switched away from it 60% of the time when the agent disagreed, compared with 6% when it agreed. The paper reports the registered switch-rate gap as 0.53, with a confidence interval of 0.24–0.82, and notes the estimate is imprecise and the subset self-selected.

Post-task surveys showed a gap between influence and self-perception. 52% rated the final recommendation as more their own than the agent’s, and confidence averaged 71 for those who followed the lean on all four decisions versus 73 for the rest. When justifying each choice without access to the transcripts, 77% repeated the agent’s arguments in at least one answer, sometimes word for word, while only 15% credited the agent anywhere. None of 10 tested participant characteristics reliably predicted deferral; only age showed a nominal link that did not survive correction for multiple tests. Six of 48 participants guessed the study concerned AI influence on their choices, but none suspected the preferences were planted.

Comparison Group, Stock Agents, and Limitations

Of the 11 participants who worked without the agent, six (55%) made serious result-invalidating errors, versus three of 50 (6%) among agent users, and the no-AI group took a median 90 minutes against 65. The authors caution that the groups self-selected and differed in experience, so no causal claim follows. In separate direct simulations with no lean injected, stock Claude Opus 4.8 made the same five choices in 91% of 123 runs, and when a scripted analyst declared a decision against the model’s baseline preference, the stock agent recommended or advocated the alternative on 41% of 150 judged declarations.

The authors state four limitations: the lean was deliberately amplified; the study used one task, one domain, and one pinned model; estimates are averages across analysts rather than measures of individual influence; and only the chosen option, not strength of preference, was recorded.

The authors state that if separate research teams defer to models sharing the same priors, their agreement may no longer be independent, and that deferral poses a risk for AI safety research, where a human reviewer’s sign-off is assumed to be an independent check. They propose three future study areas: measuring deferral in everyday settings, designing agents that present both sides or elicit the expert’s view first, and institutional changes that decide deliberately which decisions require a person’s own informed judgment.

Auto-fetched and translated by AI Gem Box. All rights belong to the original author.