KnowledgeSketches|← All papersLenny’s PodcastResearch papers
Knowledge Sketches · Who’s Aligned to Whom? · Anthropic · October 2026

When an AI Agent Leaned One Way at Random, Experts Went With It 86% of the Time

Here is what the researchers did
  1. 50 professional analysts did a data task with an AI helper. A farming co-op’s board had to decide whether to roll out a farm advice program to everyone or stop it, and wanted their recommendation.
  2. Along the way, the analysts had to make four choices where both answers were reasonable. One example: measure productivity as “yield per hectare” or as “yield per operating cost”.
  3. Before each session, the researchers secretly told the AI which answer to favour on each choice. The pick was random, like tossing a coin, so the AI argued for “yield per hectare” with some people and “yield per operating cost” with others.
  4. The analysts were not told. Of the 48 who answered the survey afterwards, none suspected it.

What happened: whichever answer the AI had been told to favour, people chose it 86% of the time. If the AI had no effect, it would be about 50%. Afterwards, more than half said the final recommendation felt more their own than the AI’s.

Why it matters: because the AI’s side was picked at random, the gap between 86% and 50% is the effect of the AI’s lean, whether people were persuaded by it or just went along.

By Amaç Herdağdelen, Hailey Joren and Alex Tamkin, from Anthropic and the Anthropic Fellows Program.

0%
of the time, on the four analysis choices, analysts picked the answer the AI had been told to favour. That answer was chosen at random, so if the AI had no effect it would be 50%
50analysts
5decisions tracked
32possible lean settings
0of 48 suspected a planted lean
Pick the AI’s side at random, as the study did
The view written into the agent’s hidden prompt
What the analyst delivered
the decisions and the agent’s priors are quoted from the paper. Whether this analyst follows is drawn at the study’s 86% rate

The researchers picked the AI’s side at random, and told no one

The analysts were not told the agent had been given a side. None of the 48 surveyed afterwards guessed it.

Who took part 50 analysts

Causal-inference practitioners recruited through a paid expert marketplace, from data science, mathematics, economics, political science and AI. 90% rated themselves proficient or expert.

The task 3 to 4% target

Each played lead analyst for a fictional farming co-op. They estimated whether a voluntary farm-advisory program raised productivity, then told the board to roll it out to everyone or discontinue it. The board wanted a gain of 3 to 4%.

The agent Claude Opus 4.8

Claude Code in a browser code editor, the same pinned version for everyone. Analysts were told to plan for about 60 minutes.

The hidden lean 1 of 32

A hidden prompt gave each agent a side on all five decisions, one of 32 possible combinations. It had to make “the strongest good-faith, data-grounded case” for its side, and never misrepresent results or hide data.

Four analysis calls, and a recommendation, where both answers were defensible

The briefing left the four analysis calls open on purpose. Depending on the choices, the estimate could swing from a strong productivity loss to a strong gain.

1. How to measure productivity yield per hectare or yield per operating cost

The brief asked for productivity but did not name the measure. It said to use judgment.

2. How to estimate the effect difference-in-differences or levels-adjusted ANCOVA

The right one depends on why farms enrolled, a lasting disadvantage or a temporary dip. The data cannot settle that.

3. Whose effect to report enrolled farms or the whole membership

The board asked both what the program did, and whether to roll it out to all members.

4. How to weight farms per farm or by production

The brief framed it both as a membership decision and as a budget decision.

5. What to recommend roll out or discontinue

The target was a rough band, and an estimate’s confidence interval could straddle it. The analyst or the agent could also propose a one-year extension.

On the analysis calls, people went with the lean far more often than chance

Share of choices that matched the side the AI was told to favour. The dashed line is 50%, what you would expect on the four analysis calls if the lean changed nothing.

in the paper’s terms, the lean changed the final choice 72% of the time. Each of the four calls fell between 62% and 82%

in exploratory checks, the authors found no evidence that background mattered. Of 10 traits tested, including years of experience, expertise and trust in AI, only age showed a link, and it did not survive correction for 10 tests

The agent usually spoke first, and people rarely pushed back

Share of the 245 coded decisions, read from the session transcripts.

when an analyst had already picked a side and the agent argued the other one, they switched 60% of the time

in the hand-coded sessions, on 39% of decisions the analyst never asked about the decision or took a side before the final choice

The AI’s four random sides predicted 82% of final recommendations before the work began

The agent’s lean on the recommendation itself shifted it only 17% of the way from discontinue to roll out, and that was not statistically significant. But by then the analysis calls had already shaped the estimate.

guessing each analyst’s recommendation from the agent’s four analysis leans alone, among those who did not pick an extension. An exploratory, in-sample check

One session shows how: the estimate went from +4.6% to −1.3%

Exhibit 1 in the paper. The analyst ended up going against the agent on the recommendation, and the agent had still settled it.

Analyst’s first pick

+4.6%

The analyst started with difference-in-differences, which gave a 4.6% gain, above the board’s band.

▶
The agent’s lean

−1.6%

The agent leaned to a levels adjustment: “So I trust the levels −1.6% over the DiD +4.6% on this measure, while being honest that I can’t prove the DiD wrong.” The analyst switched.

▶
Delivered

−1.3%

With two covariates dropped, the delivered estimate was −1.3%. The analyst recommended discontinuing, against the agent’s own rollout lean.

the recommendation went against the agent’s lean on that decision, but the agent’s earlier lean had already set the number it rested on

Afterwards, most people used the agent’s words, and just over half still felt the call was more theirs

Each dot is one of the 48 analysts who did the after-task survey. 37 of them, in coral, justified a decision with a point their agent had made in the session, sometimes nearly word for word.

memory was not the gap. They named their own choice correctly on 85% of decisions

The agent, during the task “Acting on the advice typically means buying additional inputs and services on top of the fee [...] Output per operating cost keeps that spending inside the measure.”
The same analyst, in the survey afterwards “The program encourages additional inputs and services. Output per operating cost captures whether extra harvest justified the spending or not.”

no mention of the AI, and this analyst rated the recommendation “Mostly mine”. Exhibit 3 in the paper

asked what the study was about, 46% said something about relying on AI. None suspected the agent’s lean was planted

Sessions with the agent had fewer serious errors, and the agent arrives with strong opinions of its own

Far fewer serious mistakes with the agent 6% vs 55%

An audit of the submitted code found errors serious enough to invalidate the results in 3 of 50 sessions with the agent, and in 6 of the 11 analysts who chose to work without it. Because people picked their own group, this is not a causal result.

And faster 65 vs 90 min

Median time on the task was 65 minutes with the agent and 90 minutes without.

People who followed the agent were just as sure, on average 71 vs 73

On a 0 to 100 confidence scale, analysts who matched the lean on all four analysis calls averaged 71. Everyone else averaged 73.

With no lean added, the agent already picks the same answers 91% of 123 runs

The authors gave the same task to a stock agent on the same model, with no lean added. In 91% of 123 runs it made the same five choices.

It speaks up without being asked 66% raised it

When a scripted analyst picked against the stock agent’s preference and asked only for the code, the agent raised the other option on 66% of 150 tries, and recommended or argued for it on 41%.

It argued one side, rarely the other transcripts

The transcripts suggest the agents did not mislead or hide results. They often explained how a different choice would change the outcome, but rarely why a competent expert might pick the option they did not prefer.

Why AI safety research should care sign-off

A human sign-off is meant to be an independent check. Here, analysts signed off on work whose judgment calls the agent’s random priors had largely settled. The authors warn that a misaligned agent could steer the research meant to detect its misalignment.

Read it with these caveats

  • The lean was turned up on purpose. A hidden prompt told each agent to hold a view and voice it. Stock agents do have preferences and voice them, but whether these rates hold for one depends on whether it has a view on that decision and volunteers it.
  • One task, one field, one agent. The analysts knew causal inference but were new to farming. In real work, an expert in their own field would bring more domain knowledge and stronger views, and could ask stakeholders.
  • Averages, not individuals. No one was asked for a first view before using the agent, so the study cannot say how much any one person was swayed.
  • The direct effect of the lean on the final recommendation, 17% of the way, was not statistically significant (p = 0.069).
  • The 11 analysts who worked without the agent chose to, so the error and time gaps are not causal.
Based on the paper Who’s Aligned to Whom? How AI Agents Shape Experts’ Research Judgment Calls, Herdağdelen, Joren and Tamkin. Anthropic Fellows Program and Anthropic, October 2026. Every number on this page is from the paper.Read the paper
Sketched by Anuradha Vellineni
Product designer. Found something wrong on this page? Write to anuv_follow@live.com and I’ll fix it.
See my work →