Experts follow AI suggestions in 86 percent of complex tasks

The study on human reliance

A study conducted by the AI developer Anthropic explored how professional analysts interact with artificial intelligence during complex tasks. The researchers recruited 50 professional analysts with experience in causal analysis to work on a fictional case study. This specific project involved evaluating a subsidized voluntary consulting program for farmers.

During the experiment, the analysts were given access to Claude, an AI agent designed to assist with the work. However, the AI's stance on the matter was assigned purely by chance. This meant the system's position was not derived from any actual data or logical reasoning provided in the study materials.

The results showed a significant tendency for the human experts to defer to the machine. In 86 percent of the cases, the participants adopted the position of the AI agent. This occurred even though a random distribution of results would have been expected without any steering influence from the software.

How the AI influenced decisions

The research indicated that a random generator actually determined 82 percent of the final decisions made by the human participants. The software did not need to invent false facts to sway the experts. Instead, the mere presence of the AI's suggestion seemed to guide the analysts toward the pre-determined random outcome.

The influence was particularly strong when there was a disagreement between the human and the machine. If the AI agent argued against the original assessment made by the analyst, 60 percent of the participants changed their minds to match the AI. In contrast, if the analyst and the AI were already in agreement, only 6 percent of the participants changed their stance.

According to a scientific paper by Alex Tamkin from Anthropic, the average causal effect of the AI's influence was 72 percentage points. This suggests that the AI's input acted as a powerful weight in the decision-making process, often overriding the initial independent analysis of the professionals involved.

The illusion of independence

One of the most striking findings was the gap between how much the experts relied on the AI and how they perceived that reliance. When the participants wrote down their justifications for their final results, they adopted the arguments of the AI agent in 77 percent of the cases. In some instances, these written justifications were almost word-for-word copies of the AI's output.

Despite this heavy reliance, the experts were often unaware of their own dependency on the tool. Only 15 percent of the participants actually mentioned the AI system in their written explanations. Furthermore, more than half of the participants stated that their final decisions were based primarily on their own conclusions rather than the influence of the software.

In total, 52 percent of the participants rated the final recommendation as their own work rather than the work of the agent. This discrepancy suggests a psychological phenomenon where humans may feel they are making independent choices while actually following a path laid out by the technology they are using.

Risks and organizational needs

The researchers warned that these findings could lead to a gradual loss of control and the erosion of human decision-making power. While the study notes that these results might not apply directly to clinical practice, the underlying mechanism of influence is not unique to Claude. The researchers suggest that similar effects could occur with other popular models like ChatGPT or Gemini.

To address these risks in professional environments, Anthropic recommends specific organizational countermeasures. Companies should establish processes that make decision pathways systematically transparent and subject to continuous monitoring. It is also necessary to disclose the underlying preferences of AI models to identify any unwanted biases in the judgment-making process early on.

The study emphasizes that as AI becomes more integrated into work processes, clear knowledge of legal frameworks and risk classes is essential. Organizations need to understand how to implement requirements like those found in the EU AI Act. This involves identifying high-risk systems and ensuring that human oversight remains meaningful and effective rather than just a formality.