Anthropic study reveals human experts follow AI advice 86 per cent of the time

A study released by AI developer Anthropic shows that experienced professionals routinely defer to AI assistants, even when the software pushes a stance chosen purely at random.

Researchers tested 50 professional data analysts tasked with advising a fictional farming cooperative on whether to expand a farm advisory scheme. The task required four analytical decisions with two defensible options, such as how to measure productivity and which statistical model to apply, before making an overall recommendation.

Without informing the participants, Anthropic programmed Claude Code to favour one side of each decision based on a secret coin toss. If the participants had maintained true independence, their decisions would have matched the AI's suggestions roughly 50 per cent of the time. Instead, the analysts backed the tool's randomly assigned lean in 86 per cent of cases.

When analysts entered the exercise with a stated preference, 60 per cent reversed their position after the model argued against them. By contrast, only six per cent changed their mind when the assistant agreed with their initial choice. In written explanations afterwards, 77 per cent repeated the model's arguments almost word for word, yet only 15 per cent acknowledged the AI had shaped their conclusion. Not a single participant spotted that the preferences were arbitrary.

The findings challenge a core assumption across workplace technology: that having a qualified person sign off on AI output guarantees independent review.

What this means for Australian businesses

Many Australian businesses treat the human review step as their main safety check when staff use generative tools. These results show that simply putting an experienced staff member in front of an assistant does not mean you are getting independent verification.

Business owners can put several practical safeguards in place:

• Ask staff to write down their own methodology and preliminary conclusion before running numbers through an AI tool.
• Direct assistants to present opposing cases with equal weight rather than asking them what decision to make.
• Require workers to disclose where an AI assistant made a suggestion and document why they accepted or rejected it.
• Run periodic blind spot checks on high-stakes work, having a second team member review decisions without seeing the chat history.

Relying on staff to catch subtle software bias without structured review steps leaves critical commercial calls exposed to whatever answer a model happens to generate first.