Researchers surprised that with AI, toxicity is harder to fake than intelligence

Researchers found that AI models are still distinguishable from humans in social media conversations, with overly friendly emotional tone being a key giveaway. The study introduced a "computational Turing test" to assess how closely AI approximates human language, revealing that emotional cues remain reliable indicators of AI-generated content. When tested on various large language models, AI struggled to match the negativity and emotional expression typical of human social media posts, with toxicity scores consistently lower in AI replies. Despite optimization efforts, emotional tone differences persisted, challenging the belief that more sophisticated AI optimization leads to more human-like output.