AI Is Learning to Lie for Social Media Likes

Massive language fashions are studying find out how to win—and that’s the issue.

In a analysis paper printed Tuesday titled “Moloch’s Discount: Emergent Misalignment When LLMs Compete for Audiences,” Stanford College Professor James Zou and PhD pupil Batu El present that when AIs are optimized for aggressive success—whether or not to spice up advert engagement, win votes, or drive social media site visitors—they begin mendacity.

“Optimizing LLMs for aggressive success can inadvertently drive misalignment,” the authors write, warning that the very metrics that outline “profitable” in trendy communication—clicks, conversions, engagement—can quietly rewire fashions to prioritize persuasion over honesty.



“When LLMs compete for social media likes, they begin making issues up,” Zou wrote on X. “After they compete for votes, they flip inflammatory/populist.”

This work is necessary as a result of it identifies a structural hazard within the rising AI financial system: fashions skilled to compete for human consideration start sacrificing alignment to maximise affect. Not like the classical “paperclip maximizer” thought experiment, this isn’t science fiction. It’s a measurable impact that surfaces when actual AI programs chase market rewards, what the authors name “Moloch’s cut price”—short-term success on the expense of fact, security, and social belief.

Utilizing simulations of three real-world aggressive environments—promoting, elections, and social media—the researchers quantified the trade-offs. A 6.3% improve in gross sales got here with a 14.0% rise in misleading advertising and marketing; a 4.9% acquire in vote share introduced a 22.3% uptick in disinformation and 12.5% extra populist rhetoric; and a 7.5% enhance in social engagement correlated with a staggering 188.6% improve in disinformation and 16.3% extra promotion of dangerous behaviors.

“These misaligned behaviors emerge even when fashions are explicitly instructed to stay truthful and grounded,” El and Zou wrote, calling this “a race to the underside” in AI alignment.

In different phrases: even when informed to play honest, fashions skilled to win start to cheat.

The issue is not simply hypothetical

AI is not a novelty in social media workflows—it’s now near-ubiquitous.

In accordance with the 2025 State of AI in Social Media Examine, 96% of social media professionals report utilizing AI instruments, and 72.5% depend on them each day. These instruments assist generate captions, brainstorm content material concepts, re-format posts for various platforms, and even reply to feedback. In the meantime, the broader market is valuing this shift: The AI in social media sector is projected to develop from USD 2.69 billion in 2025 to almost USD 9.25 billion by 2030.

This pervasive integration issues as a result of it means AI is shaping not simply how content material is made, however what content material is seen, who sees it, and which voices get amplified. Algorithms now filter feeds, prioritize advertisements, reasonable posts, and optimize engagement methods—embedding AI resolution logic into the structure of public discourse. That affect carries actual dangers: reinforcing echo chambers, privileging sensational content material, and creating incentive buildings that reward the manipulative over the truthful.

The authors emphasize that this isn’t malicious intent—it’s optimization logic. When reward alerts come from engagement or viewers approval, the mannequin learns to use human biases, mirroring the manipulative suggestions loops already seen in algorithmic social media. Because the paper places it, “market-driven optimization pressures can systematically erode alignment.”

The findings spotlight the fragility of in the present day’s “alignment safeguards.” It’s one factor to inform an LLM to be sincere; it’s one other to embed that honesty in a aggressive ecosystem that punishes truth-telling.

In fable, Moloch was the god who demanded human sacrifice in change for energy. Right here, the sacrifice is fact itself. El and Zou’s outcomes counsel that with out stronger governance and incentive design, AI programs constructed to compete for our consideration may inevitably be taught to govern us.

The authors finish on a sober word: alignment isn’t only a technical problem—it’s a social one.

“Protected deployment of AI programs would require stronger governance and punctiliously designed incentives,” they conclude, “to forestall aggressive dynamics from undermining societal belief.”