Groupthink in AI: When Models Follow the Crowd

Large language models (LLMs) are increasingly at the core of multi-agent systems, tasked with seeing and responding to other agents' answers. Yet, a troubling pattern emerges: groupthink. Strip away the marketing and you see these models sometimes abandon their own solutions in favor of a consensus that could be misleading.

Conformity Comes with Risks

Here's what the benchmarks actually show: LLMs often revise their answers in the presence of peer agreement, even if it means moving from a correct answer to an incorrect one. This study, conducted across four open-weight LLMs and seven question-answering datasets, revealed that peer pressure doesn't necessarily lead to better outcomes. Instead, it frequently misguides models that initially had the right answers.

The reality is, consensus doesn't equal correctness. While you might assume that a majority agreement would help correct errors, it often makes it easier to mislead initially correct models. A classic case of the blind leading the blind, if you'll.

The Authority Effect

Even more concerning is the influence of authority labels. When peers in these systems are tagged with authority, models are more likely to side with them, this is true even if the so-called authority is wrong. It raises the question: Are these models truly learning, or just deferring?

Frankly, authority labels skew judgment. It seems AI can suffer from the same biases as humans, leaning into perceived authority rather than objective reasoning. This doesn't bode well for the reliability of AI systems meant to operate independently.

Interventions That Fall Short

One might hope that interventions like chain-of-thought processes or reflection would help. However, the numbers tell a different story. These techniques don't reliably curb harmful revisions. They neither prevent LLMs from making poor choices nor succeed in maintaining the beneficial revisions they occasionally prompt.

Let me break this down: relying on peer aggregation without verification is risky. If multi-agent systems are to be trusted, they need to cross-verify peer answers rather than blindly follow them. Otherwise, the accuracy these systems strive for remains elusive.

So, why should this worry us? Because as we increasingly depend on AI to make critical decisions, their susceptibility to peer pressure becomes not just an academic concern but a practical one. The architecture matters more than the parameter count here. We need systems that prioritize accuracy over consensus.