Want Better AI Results? Make the Machine Disagree With You

Human-AI teams that argue produce better results than humans alone, AI alone, or collaborative pairs that simply agree. That is the central finding of research on loan evaluations cited by University of Massachusetts (Amherst) Professor Monideepa Tarafdar, writing in The Conversation.

Her argument lands as evidence mounts that heavy reliance on generative AI erodes critical thinking and job skills. The American Psychological Association has flagged the problem of "cognitive offloading"—outsourcing mental workload to AI too quickly—which can produce shoddy, subpar work.

But Tarafdar's research suggests knowledge workers who use AI correctly can still gain a significant edge. The key, she writes, is creating "friction" by having AI challenge your ideas rather than execute them.

Why disagreement matters more now

Researching opposing viewpoints has always been sound practice. It matters even more today because AI models tailor their output based on users' previous prompts and responses. That tendency creates an "echo chamber" effect, leading both the models and their users to overlook information that doesn't fit an established pattern.

To counter this, Tarafdar says users should ask AI directly for opposing perspectives.

The approach is not theoretical. In her study, a marketing professional asked AI to create virtual customers who disliked the product being marketed. The model's negative feedback included new product ideas the designers had not considered.

Another research subject, an attorney, asked an AI model to find obscure legal loopholes and game out how they might be exploited. The prompt pushed the model to provide "surprising but realistic" scenarios of unethical worker behavior.

The evidence for friction

Tarafdar cites outside studies backing the friction thesis. In the loan-evaluation study, published in Management Science, human-AI collaboration beat both human-only and AI-only work—and collaboration involving disagreement produced the best results of all.

A second study, also in Management Science, showed creative writers produce better work when they use AI interactively as a sounding board rather than to ghostwrite copy on its own.

The management play

Tarafdar says leaders should instruct employees to use "friction-generating queries," drawing on their own knowledge and experience to make the AI question its outputs.

"In these efforts to support knowledge workers, an important detail is pointing out that adding friction can appear to make an AI system work against you, but—as new research shows—it actually helps it work for you," she writes.

For companies deploying AI across their workforce, the implication is direct: the competitive edge may belong not to those who automate fastest, but to those who engineer productive disagreement into the loop.