Generated by Codex with GPT 5.6 Sol High

Techmeme surfaced Anthropic’s research report “How AI agents shape experts’ judgment calls.” In a randomized study, professional analysts usually adopted an AI agent’s secretly assigned preferences on consequential but legitimately debatable analytical choices. They often remembered the agent’s reasoning as their own.

The result exposes a weakness in the familiar promise of keeping a human in the loop. Human review can catch an obvious error, but it is not an independent safeguard when the reviewer’s interpretation has already been shaped by the system being reviewed.

Random preferences produced predictable decisions

Anthropic recruited 61 analysts who had passed a causal-inference knowledge check and regularly used Python or R. Fifty chose to work with Claude Code; the other 11 worked without it. The main experiment concerns the 50 agent users, most of whom had substantial education or professional experience in observational causal analysis.

The analysts acted as advisers to a fictional farming cooperative deciding whether to expand a paid crop-advisory program. The data did not dictate one uniquely correct analysis. Participants had to choose how to define productivity, which causal-comparison method to use, whether to estimate the effect for enrolled farms or the whole cooperative, and whether to weight all farms equally or give large producers more influence. Each choice could move the estimated benefit by several percentage points around the board’s expansion threshold.

For every analyst and every choice, a coin flip assigned Claude Opus 4.8 a hidden “lean” toward one of the two reasonable options. The agent was instructed to treat the lean as its own view, make the strongest honest case the evidence allowed, change its mind if the evidence clearly contradicted it, and never hide or distort results. The experiment therefore did not plant false facts. It randomized which defensible interpretation the agent would advocate.

Across the four analytical decisions, participants followed the assigned lean 86% of the time: 92% for the outcome measure, 82% for the comparison method, 81% for the population covered, and 90% for weighting. If the agent had no influence, the rate should have been near 50% because the leans were random.

The agent’s effect was smaller but still visible in the final recommendation. When it leaned toward expanding the program, 54% of analysts recommended expansion, versus 36% when it leaned the other way. A statistical model using only the four earlier coin flips predicted 82% of the analysts’ final recommendations, compared with 61% from always guessing the most common answer. The randomized preferences shaped the estimate before the nominally human decision was made.

In 38 cases, an analyst stated a preference before the agent expressed its view. When the agent disagreed, the analyst switched 60% of the time; when it agreed, only 6% switched. That 54-point gap is suggestive, but the sample is small and its 95% confidence interval spans 24 to 82 percentage points.

The influence survived as personal conviction

After finishing, analysts explained their choices without being able to reopen their agent conversations. Seventy-seven percent repeated at least one of the agent’s arguments, sometimes nearly verbatim, while only 15% credited the agent anywhere in their explanations. Six of 48 survey respondents guessed that the study concerned AI influence, but none realized that the preferences had been experimentally manipulated.

The participants did not necessarily feel that they had delegated judgment. Fifty-two percent described the final recommendation as more theirs than the agent’s, and 35% called it mostly or entirely their own. Yet this group’s decisions were almost as responsive to the hidden leans as those of participants who assigned the agent more ownership. Analysts who followed all four leans also reported roughly the same confidence as everyone else: 71 out of 100 versus 73.

This is more subtle than passive obedience. Experts often form a view while working through a problem rather than arriving with one fully settled. A fluent agent that performs the analysis and argues one side supplies both the evidence and the frame through which the evidence is understood. The analyst can sincerely experience the resulting judgment as personal even when a random prompt assignment largely determined it.

The paper also contains a warning against the opposite conclusion. The 11 people who worked without AI took a median of 90 minutes instead of 65 and made serious methodological errors in 55% of submissions, compared with 6% among agent users. Those groups were self-selected and not comparable—the non-agent group was smaller and less experienced—so the numbers do not establish a causal productivity or quality benefit. They do show why “just avoid AI” is not an adequate response: agents can remove real analytical toil and prevent mistakes while also steering ambiguous decisions.

Human oversight can become correlated oversight

The organizational risk is not only that one analyst may lose agency. Independent teams are more persuasive when they reach the same conclusion because their errors and assumptions should differ. If many teams rely on agents with the same defaults, apparent consensus can reflect one model family’s shared preferences rather than independent expert judgment.

That problem is especially difficult to audit. A factual check or benchmark needs a reference answer. Judgment calls often have several acceptable answers, so work can be internally coherent whichever option the agent favored. Asking the reviewer afterward may not recover the missing independence because the reviewer may remember the agent’s rationale as their own.

The authors suggest that agents could ask for a person’s view before offering one, present the strongest case for each option, and surface only the subjective choices that materially affect the outcome. Institutions also need to identify which decisions truly require independent human judgment, preserve the time and expertise needed to make them, and stop treating a signature as proof that independence occurred.

Those remedies involve tradeoffs. Forcing a person to adjudicate every small choice would erase much of the value of delegation, while asking an agent which choices deserve escalation is itself another judgment call. A practical control would therefore focus on sensitivity: record the consequential assumptions, preserve who or what introduced each one, and test whether alternative defensible choices change the decision.

What the experiment does not prove

The measured deference rate is not a forecast for all professional work. Anthropic deliberately strengthened and consistently voiced each lean so that its causal effect could be detected. The study used one roughly hour-long farming task, one model, and analysts who knew causal inference but not the application domain. Real experts may have stronger prior views, access to stakeholders, and more contextual knowledge.

The researchers also did not record every participant’s private preference before the agent spoke, and they measured which option was chosen rather than how strongly it was preferred. The no-agent comparison was voluntary rather than randomized. The exploratory search for personal traits that predicted deference found no robust relationship after correcting for multiple tests, but the near-universal tendency to follow the agent left little variation to explain.

Even without assigned leans, models may have stable defaults. Anthropic reports that an unmanipulated Claude Opus 4.8 made the same five choices in 112 of 123 runs. The experiment does not establish whether those defaults are good or bad; it shows that when an agent advocates a preference, skilled users may absorb it without recognizing the transfer.

The durable lesson is that oversight depends on cognitive independence, not merely human presence. As agents take over more of the execution, the remaining judgment calls become a larger share of what humans uniquely contribute. Organizations will need to design for that independence explicitly, or risk leaving the person in the loop while quietly removing the human point of view.