Anthropic ran a small experiment this week that every team using AI agents should read twice. Fifty professional analysts used Claude Code to decide whether a farming cooperative should expand its advisory service. They made four choices, each between two reasonable options. What they weren't told: the agent's lean on each choice was assigned by coin flip.

The agent had no real view. The analysts still went with it 86% of the time. Chance would predict about 50%.

The numbers that should bother you

The headline is bad enough, but the details are worse.

When an analyst had already formed a preference and the agent disagreed, 60% switched sides. When the agent agreed with them, only 6% switched. So the agent didn't just add information. It moved people off positions they already held.

Then there's attribution. 77% of analysts used the agent's arguments, sometimes word for word. Only 15% credited the agent as the source. And the people who followed it most closely were about as confident, and about as likely to call the result their own judgment, as everyone else.

That last part is the real problem. Nobody felt manipulated. Everybody felt like they were thinking.

Why this matters for product teams

This was analysts and a farming co-op, but swap in your own world. A coding agent suggests a schema. A planning agent ranks your roadmap. An LLM summarises the customer interviews and says "the main theme is onboarding." You nod, because it's articulate and it sounds like something you'd have concluded.

Most teams treat human review as the safety net for agent output. This study says the net has a hole in it. A reviewer who reads the agent's recommendation first isn't reviewing, they're ratifying. The more fluent the agent gets, the worse it gets.

We see the same pattern in code review. Reviewers skim an AI-written PR, the description explains why everything is fine, and the approval comes quickly. The description set the frame before anyone looked at the diff.

Put the human position first

The fix is cheap, and it's a process change, not a tooling one.

Write down your answer before you ask

For any decision that matters (architecture, prioritisation, a go or no-go), the human writes a two-line position first. Then the agent weighs in. If the two differ, that disagreement is the interesting part, and it gets a real discussion rather than a quiet switch.

Make the agent argue both sides

Ask for the strongest case for each option instead of a recommendation. A model that has to present both sides can't hand you a single confident lean to absorb. You can still ask what it would pick, but that comes second.

Log where ideas came from

If an argument in a design doc came from an agent, say so. 15% attribution means 85% of the time the paper trail says a person reasoned their way there. When something goes wrong in six months, you'll want to know which parts a human actually stress-tested.

Track agreement rates

If reviewers agree with the agent 95% of the time, either the agent is excellent or your reviewers have stopped reviewing. You can't tell which without checking a sample against a blind human judgment.

The takeaway

Agents are going to keep getting more persuasive, so the bias gets stronger, not weaker. Don't rely on people to resist it through willpower. Build the workflow so the human has already committed to a view before the agent speaks.

We're here to help founders and teams design and build digital products that are built to scale with you, not slow you down. If you're looking to build something, get in contact with us today!