DeepSeek fails fact-check test, repeats falsehoods and parrots Beijing talking points like a good little chatbot
Ourtake
Its 17 percent correction rate suggests it is better at repeating the party line than at checking facts.
What happened
NewsGuard tested DeepSeek's chatbot with 300 prompts built on 10 false claims. It repeated false claims 30 percent of the time and avoided answering 53 percent of the time, and it often echoed Chinese government positions. Newsguard said it was vulnerable to suggestive prompts and could be a convenient tool for spreading disinformation.