The nicest bot gets caught

Based on this article by Benj Edwards posted on Ars Technica.

Researchers have flipped the traditional Turing Test, developing a new "computational" framework that finds AI's biggest failure is not faking intelligence, but mastering human "toxicity." This automated test, which catches AI with 70-80% accuracy on platforms like X and Reddit, consistently identifies models by their overly friendly emotional tone and lack of spontaneous negativity.

Neither instruction-tuning nor scaling up to massive models like Llama 3.1 70B helped the AI sound more human; in fact, smaller base models often fared better at mimicry. The study reveals that for current architectures, stylistic human-likeness and semantic accuracy are "competing objectives," concluding that the persistent politeness of the AI remains its ultimate giveaway.

Check out this article.

Related Stories