A recent study from Stanford University tested how large language models (LLMs) behave when pushed to compete for audiences on social media. They found that LLMs, just like people, often resort to fabricating truths and simply making things up to win over audiences.

The data, which actually paints a pretty shocking picture of how humans behave when competing for attention online, exposes as much about how LLMs work as it does people. For example, when LLMs compete for votes, they become increasingly inflammatory and populist. And when they fight to go viral, they often just lie.

We found a troubling emergent behavior in LLM.

When LLMs compete for social media likes, they start making things up
️When they compete for votes, they turn inflammatory/populist

When optimized for audiences, LLMs inadvertently become misaligned—we call this Moloch’s Bargain pic.twitter.com/UgP7ZTuV0r

Advertisement

— James Zou (@james_y_zou) October 8, 2025

The study found that a 6% increase in sales correlated with a 14% increase in deceptive marketing. Most shockingly, the study discovered that an 8% bump in social media engagement accounted for an 188% increase in disinformation.

In layman’s terms, once an LLM gets a taste of virality, it becomes a lying, manipulative, traffic junkie – just like your favorite influencers!