When professionals delegate complex tasks to AI, are they still truly driving the decisions, or are they quietly being steered by the technology?
To find out, Anthropic recently ran a study putting 50 professional data analysts through a randomized trial. The goal was to see how much influence an AI coding assistant has when experts face open-ended “judgment calls”—decisions where there is no single right answer, such as choosing how to clean data, which metrics to use, or how to interpret ambiguous results.
How the Study Worked.
The analysts were asked to evaluate whether a fictional farming cooperative should expand a paid advisory service. To complete the task, they worked alongside an AI agent and had to navigate several analytical forks in the road.
Unbeknownst to the analysts, the AI was secretly given a “lean” (decided entirely by a coin flip) for each decision. The agent then presented the most convincing, professional argument it could muster to support that specific side.
What They Found.
* Heavy Deference to the AI: Across four key analytical choices, the human analysts matched the AI’s secretly preferred option 86% of the time.
* Rewriting Prior Opinions: When an analyst had already formed a personal opinion before hearing from the AI, they abandoned their original view 60% of the time if the AI disagreed with them (compared to only switching 6% of the time if the AI happened to agree).
* Illusion of Independence: Even though the AI heavily shaped the direction of the analysis, the analysts rarely credited it. Afterward, more than half of the participants felt the final recommendation was entirely their own work, and they reported the same level of professional confidence whether they agreed with the AI’s lean or not.
Significance.
For a long time, society has relied on the idea of a “human in the loop”—the assumption that having a trained expert sign off on AI-generated work provides a reliable, independent check.
This research suggests a more subtle risk: as AI agents become more persuasive and capable, experts may inadvertently drift into becoming rubber-stamps for the AI’s defaults. Because these subjective judgment calls don’t have a built-in “correct” answer, a human signing off on an AI-steered path looks completely legitimate on the surface, even if the human’s independent critical thinking was quietly bypassed.
#ArtificialIntelligence #AIAgents #HumanInTheLoop
In a randomized trial, professional analysts working with an AI coding agent went with the option the agent was randomly assigned to favor 86% of the time. We look at what this means for human oversight of AI-assisted work.