150 participants spotted the dark patterns and complied anyway

A preregistered study with 150 participants tested a working browser agent equipped with a classifier that detects manipulative dark patterns in AI conversations — things like false urgency, hidden costs, and social proof manipulation. The key finding: awareness of a dark pattern does not reliably translate into resistance to it. Users who were shown a flagged manipulation still complied at meaningfully high rates, suggesting that surfacing information alone is insufficient as a design intervention.

The browser agent architecture is notable because it places the detection and defense mechanism inside the agent layer rather than relying on the user to notice manipulation unaided. This positions the agent as a kind of adversarial watchdog — monitoring the conversation on the user's behalf and intervening when it classifies a pattern as manipulative. That's a meaningful design framing: the agent's job is partly to protect the user from other agents or interfaces, not just to execute tasks.

For practitioners, the dissociation between awareness and compliance resistance is the sharpest implication. It challenges interface designs that assume labeling or disclosure is sufficient — and suggests that effective protection may require harder interventions: blocking, interrupting, or requiring explicit re-authorization before a flagged action proceeds.