Researchers Surprised That With AI, Toxicity is Harder To Fake Than Intelligence

Researchers from four universities have released a study revealing that AI models remain easily detectable in social media conversations despite optimization attempts. The team tested nine language models across Twitter/X, Bluesky and Reddit, developing classifiers that identified AI-generated replies at 70 to 80% accuracy rates. Overly polite emotional tone served as the most persistent indicator. The models consistently produced lower toxicity scores than authentic human posts across all three platforms.

Instruction-tuned models performed worse than their base counterparts at mimicking humans, and the 70-billion-parameter Llama 3.1 showed no advantage over smaller 8-billion-parameter versions. The researchers found a fundamental tension: models optimized to avoid detection strayed further from actual human responses semantically.