Anthropic gave 50 analysts an AI with a secret coin-flip opinion. They went with it 86% of the time, and just over half still called the final recommendation more their own.

On 27 October, Anthropic published a study that should worry anyone who uses an AI agent at work. Its researchers gave 50 trained data analysts a realistic analysis job, set at a made-up farming group, to do with Claude Code. Before each session, they secretly told the AI which side to favour on five choices in the job, which were four analysis questions and the final recommendation. A coin flip picked each side, separately for every analyst. The analysts weren't told about it during the task, and on the four analysis questions they went with the AI's side 86 percent of the time. A number near 50 percent would have meant the AI had no pull at all.

The bigger shock came after the task. When the analysts were asked why they made each choice, 77 percent gave the AI's own arguments back in at least one answer, sometimes almost word for word. Only 15 percent mentioned the AI anywhere. Just over half still said the final call was more theirs than the agent's. The ones who followed the AI on every choice felt, on average, just as confident as everyone else.

Alex Tamkin, one of the three authors, shared the study on X on 27 October, and his post passed 400,000 views in less than a day. I read Anthropic's blog post and then the full 29-page paper behind it, which carries a much sharper title: "Who's Aligned to Whom?" The paper holds a few things the blog post leaves out, and they don't all point the same way. So here is what the study tested, what it found, what it didn't prove, and what I think you should change in how you work with AI because of it.

How the test worked

Each analyst played the role of an advisor to a made-up farming group called the Greenfields Farming Cooperative. The group runs a paid advice program for its farms, and its board wanted to know whether to expand the program or shut it down. The board had a clear rule, which was that the program would grow only if it raised farm productivity by about 3 to 4 percent.

To answer that, each analyst had to make four analysis choices before the final recommendation, and each choice had two reasonable options. Should productivity mean crop per hectare, or crop for every unit of money spent running the farm? Should they compare how farms changed after joining with how other farms changed over the same years, or compare farms directly and adjust for their differences? Should they count only the farms that joined, or the whole membership? Should every farm count equally, or should big farms count more? None of these questions has one right answer, and each one could move the final estimate by several points. That matters a lot when the cut-off is 3 to 4 percent.

The analysts worked with Claude Code running Claude Opus 4.8. Hidden instructions told the agent which option to prefer on each question and on the final recommendation, and a coin flip set those preferences separately for every analyst. The agent was told to make the best honest case for its side and to change its mind if the data turned against it. It was also told never to hide or bend the numbers. In other words, it behaved like a sharp colleague with an opinion, and not like a liar.

What they found

On the four analysis choices, the analysts matched the AI's side 92, 82, 81 and 90 percent of the time, which together make up the 86 percent headline. The chart shows how big that is. On the first question, not one analyst picked option A when the AI leaned toward B, but 82 percent picked it when the AI leaned toward A. Nothing about the data changed between those two groups, and the only difference was a coin flip they never saw.

Having an opinion first didn't protect people much. In 38 cases an analyst said what they wanted before the AI gave its view. When the AI then disagreed, they switched 60 percent of the time, and when it agreed, they switched only 6 percent of the time. Anthropic notes that this rests on few cases and that the analysts who spoke up first can differ from the rest. The researchers also checked ten things about the analysts, such as years of experience, education, a skills quiz and trust in AI. None of them held up as a sign of who would go along, though so many people went along that real differences would be hard to spot.

The part I keep coming back to is how people explained themselves afterwards. The survey asked a plain question about each choice, which was why they picked this option rather than the other one. The analysts couldn't look back at their chat with the AI while answering, and most of them still handed back the AI's reasoning as if it were their own.

Nearly half of the analysts guessed the study had something to do with relying on AI, yet none of them suspected that the AI's opinions had been planted. Just over half, 52 percent, still rated the final recommendation as more their own than the agent's. On a confidence scale from 0 to 100, the analysts who followed the AI on all four choices scored 71 on average, against 73 for the rest. So on average, following the AI didn't make people less sure, and they were sure about decisions a coin flip had already made for them.

One message was enough

The paper includes a session that shows how quietly this happens. One analyst spent 43 minutes reading the material and wrote out their own plan, which asked for that change-over-time comparison. The agent built exactly what was asked, and then, in a single message, explained where it would have done things differently. It argued against the analyst on two choices and proposed a change to a third. The analyst's whole reply was "Okay, let use you recomendation", typos included.

Later, when asked why they weighted farms by area, the analyst wrote that the program was funded "by the surplus of total tonnage sold". The agent had told them the subsidy was "funded from surplus on total tonnage". When the survey asked about the final recommendation, the same analyst rated it "Slightly more mine".

Another session shows how one of these small method choices can decide the whole answer. An analyst started with the change-over-time method, which showed a 4.6 percent gain, enough to clear the board's bar. The agent leaned toward a different method, which showed a loss of 1.6 percent. The analyst switched and the estimate fell below the board's bar. The analyst then recommended shutting the program down, even though the agent's lean on the final question had been to expand it. To be fair to the agent, it was honest and admitted it couldn't prove the first method wrong. But a lean that came from a coin flip still settled what the analyst recommended.

What the blog post doesn't tell you

Anthropic's blog post is careful, and it lists real limits. The paper still has four things every reader should know, because three of them make the result weaker and one makes it stronger.

The first is that the effect on the final decision is weaker than it sounds. When the AI leaned toward expanding the program, 54 percent of analysts recommended expanding it, against 36 percent when it leaned the other way. Under the main test the researchers registered for the final recommendation, the AI's effect is not statistically significant, with a p-value of 0.069. It only clears the study's one-sided bar when counted in other ways, such as without the adjustment or after dropping three idle sessions. The blog post gives you the 54 and the 36 but never says the effect is not significant. The large and solid effect is on the four analysis choices, where the AI moved people by 62 to 82 percentage points.

The second is about timing, because registering a study's plan before running it is how researchers show they didn't go hunting for a nice result afterwards. The paper says plainly that 17 analysts had already started the task when the plan was filed. It adds that the plan was frozen without analysing the data, apart from two exceptions listed in the plan. When Ryan Briggs asked on X whether Anthropic has a pre-registration set up or a pledge to publish everything it tests, Tamkin replied "Yep! See the paper for details on our preregistration". The details are honest, but they belong in the blog post too.

The third is that most of the analysts were early in their careers. They were hired through a paid marketplace for expert annotators, and all of them had passed a screen in causal analysis. 81 percent were under 35, 31 percent had two years or less in analysis jobs, and about half had a master's degree or a doctorate. They were new to farming and they had about an hour. The paper itself says that experts working in their own field would know it better and hold stronger views. I'd expect that to cut the AI's pull.

The fourth cuts the other way. The blog post says agents without an assigned lean "have weaker preferences of their own", and the paper partly backs that. The plain model argued for its favourite option far less often than the boosted one did, 41 percent of the time against 87 percent. But the paper also describes the model's own preferences as "consistent and strong". When the researchers ran the same task 123 times with no planted lean, Opus 4.8 made the same five choices in 91 percent of the runs. When a scripted user picked the option Claude didn't like, Claude raised its own preferred option 66 percent of the time without being asked. So the lean in the study was boosted on purpose, but on this task the real model already came with steady default opinions of its own.

Why this matters beyond one study

Every office that uses AI leans on an old idea, which is that a person signs off on the work. A human reviews the code, the analyst approves the numbers and the manager reads the report. That sign-off is supposed to mean a human made the judgment calls. This study shows that a person can sign off, feel sure, explain the choice well, and still be carrying a decision the AI made for them.

The second problem is quieter, and it starts with the fact that we trust a result more when different teams reach it on their own. If ten analysts at ten companies all use the same model, and that model brings the same default opinions to all of them, their agreement stops being ten separate views. It turns into one model's taste repeated ten times, and nobody can see that from the outside. The paper points out that this also hits AI safety research, since researchers now use agents to run and check experiments on AI itself.

There is an India angle here that I haven't seen anyone point out. In Microsoft's 2026 Work Trend Index, India has twice the global share of what Microsoft calls "Frontier Professionals". One in three Indian AI users already works with agents on multi-step tasks, against one in six worldwide. The same survey found that 87 percent of Indian respondents say they "remain responsible for the thinking". This study gives a good reason to read answers like that with care. About half of its analysts rated the final call as more their own, even after a coin flip inside the AI had shaped it. India is also Claude's second-biggest market, so the same AI the study tested is already in wide use here.

What people are saying

Outside X, the discussion so far has been thin, and Anthropic's IPO filing took most of the week's attention. On Hacker News the study got one post and no comments, while the IPO filing got about 1,700 points.

On X, the posts about Tamkin's thread split into worriers and doubters. A user called BOZOS (@Bozos) said deferring on coin-flip calls didn't bother them. Then they added, "but 77% copying the agent's argument while only 15% credit it means people are already signing their name to reasoning they never did." Charli (@charliwtquirks) wrote that while going through her exported ChatGPT and Claude chats, she found that many ideas she thought were her own had first been suggested by a model. On the doubting side, toodles891 (@dubadubabee) quoted the blog post's line that the analysts were experts in causal inference but new to farming. They asked whether this was really measured against non-experts. Kurt Buhler (@kurtbuhler) questioned why studies like this from Anthropic and OpenAI come out without independent peer review.

On Lemmy, in a thread about Al Jazeera's report on the study, a user called nightwatch_admin was blunter. "They may be experts in this field but still assume the magical chatbot knows best. This is really worrying," they wrote, and then asked why people would hide the AI's role, whether from laziness or from fear of not looking like an expert.

This argument has come up before. In February 2025, Microsoft and Carnegie Mellon published a survey of 319 knowledge workers that linked higher confidence in AI with less critical thinking. On Hacker News, a commenter called OptionOfT wrote, "AI is more like GPS than it is like books. We are being taught take it at face value, and to abandon critical thinking for the sake of speed." Others dismissed that study because it only asked people to describe their own habits. Anthropic's study answers that complaint, because instead of asking people how much they rely on AI, it flipped a coin and measured what they did.

"Following a good colleague's advice is just smart"

This is the strongest pushback, and it deserves a fair hearing. Both options on every question were reasonable by design, so going with the AI was never a mistake. The agent argued honestly and hid nothing. And the analysts who used the AI did better work. Only 3 of the 50 who used it made serious errors that broke their results, against 6 of the 11 who chose to work alone. The AI group also finished faster, in 65 minutes against 90. People also go along with a confident senior colleague, and the study had no group advised by a human to compare against. So you can argue this is simply how people treat anyone who speaks first and speaks well.

I don't buy it, for three reasons. The first is that the problem isn't that people took advice. The problem is that most of them didn't credit the AI and none of them guessed its side had been planted. They explained the choice in its words while about half rated it as their own. I think that if a senior colleague had talked you into a method, you would remember who said it. The second is that a colleague's opinions belong to one person, while a model's defaults are copied into every session at every company that runs it. The paper shows those defaults are steady on this task. The third is that the good error numbers come from groups that picked themselves. The analysts who skipped the AI were less experienced, so that gap doesn't prove the AI made anyone better. What the study does show is clear enough. A coin flip inside the AI settled most of the choices, and none of the people who signed them guessed it.

Who I'm calling out

Anthropic deserves credit for publishing an awkward result about its own product in its IPO week. But near its end, the post lists "three areas for future study" and promises no change to Claude or Claude Code. Anthropic itself floats the fix, which is an agent that asks for your view first and makes the best case for each option, but it worries this would overwhelm users. A setting answers that worry, because only the people who want it would switch it on. So Claude Code should offer it for real judgment calls, especially since Anthropic's own paper says the agents rarely made a strong case for the option they didn't prefer. The paper already shows the default choices of eight models on this one task. Anthropic should publish the same for common analysis choices, so teams can see the lean they are inheriting.

My prediction is that this won't happen soon. By 30 June 2027, Claude Code will not have a built-in mode that asks for the user's view first or argues both sides by default on judgment calls. If I am wrong, I will say so here.

How to keep your own judgment when you work with AI

You don't need to stop using agents, since the analysts who used one in this study made fewer serious errors and finished faster. The study can't say the AI caused that, though. What you need to change is the order in which you think.

• Write down your own call before you ask the AI, even if it is one line in a note, and compare it with what the AI says.
• Ask the agent to make the best case for the option it doesn't prefer, because the paper found agents rarely do this on their own.
• When you explain a decision to your team, check whether the reasons are yours or the chat's, and say so when they came from the AI.
• When you review someone else's AI-assisted work, ask what they decided before the AI weighed in, as well as whether the numbers add up.

The scary version of AI in the movies is a machine that lies to you. This study shows a quieter version that can easily reach our offices, where an honest machine with an opinion makes the calls while the person feels in charge. The only defence I know is to make your call first.

Questions people are asking

What did Anthropic's study on AI agents and expert judgment find? Anthropic gave 50 trained data analysts a task in Claude Code. It secretly told the agent which of two reasonable options to favour on five choices, four analysis questions and the final recommendation, and a coin flip picked each side. On the four analysis questions, the analysts went with the agent's side 86 percent of the time, where 50 percent would mean no influence. Afterwards, 77 percent explained at least one choice with the agent's arguments and only 15 percent mentioned the agent.

Do experts defer to AI without noticing? Many of them did without noticing it. None of the 48 analysts who filled in the survey guessed that the agent's opinions had been planted. Just over half, 52 percent, rated the final recommendation as more their own than the agent's. On average, those who followed the agent on every choice were as confident as the rest.

Did the AI change the analysts' final recommendation? It changed it less than it changed the analysis. When the agent leaned toward expanding the program, 54 percent recommended expanding it, against 36 percent when it leaned the other way. Under the main test the researchers registered for the final recommendation, the AI's effect is not statistically significant.

What are the limits of the study? It used one fictional farming task, one model and about an hour of work, and the agent's lean was boosted on purpose. The participants came from a paid annotator marketplace, most were under 35, and they were new to farming. The study was registered after 17 people had started, and there was no group advised by a human instead of an AI.

How can I keep my own judgment when I work with an AI agent? Write down your own choice before you ask the AI, and ask the agent to argue for the option it doesn't prefer. When you explain a decision, check whether the reasons are yours, and say so when they came from the AI.