More AI Isn't Always Smarter: Study Finds Too Many Models Can Hurt

A team of six researchers published a study looking at "multi-agent systems" — setups where several AI models work together on one tough question, like a panel of experts who each answer and then compare notes. They tested eight different ways of choosing which models to put on that panel, using hard science questions as the exam. The goal was simple: figure out which selection strategy actually produces better answers.

The intuitive answer — more models means better answers — turned out to be wrong. As the researchers widened the pool of candidate models, accuracy often fell below what a single top-performing model achieved on its own. The reason is that models trained by different companies think in different styles, so they disagree in ways that don't help. A weak or odd model out can drag the group down, or confuse the AI "judge" that picks the winning answer.

The strategy that worked best was narrower: pick your candidates from within a single model family, meaning several versions of the same maker's model, or models trained on similar data. They're consistent enough to work together but varied enough to add something. In other words, diversity for its own sake isn't a virtue — it's a risk. The authors describe adding random models as a source of "system instability," which is a polite way of saying the whole thing gets flakier.

Why should you care? You probably don't build these systems, but the AI assistants and chatbots you use may be assembled this way behind the scenes. If companies assume that stacking more models equals better quality, your tools could get slower, more expensive to run, or quietly less accurate — and "powered by many AI models" may be marketing rather than a real improvement. One honest caveat: this study used academic science benchmarks, so real-world results may differ, and the findings still need to be tested at larger scale before anyone redesigns their products around them.