LLMs' Non-English Text Improves, Still Detectable by Nuance
TL;DR. AI-generated text in non-English languages, while improving significantly, remains distinguishable to native speakers due to specific word choices and linguistic patterns. - Native Japanese speakers report AI models frequently use overly formal or unusual vocabulary, unlike human writing. - AI-generated English has advanced, making detection more difficult for native speakers compared to other languages. - Model performance varies widely across languages, with Anthropic's Opus and Google's models noted for strong non-English output.
- Native speakers can still identify AI-generated text in languages like Japanese by specific word usage.
- AI-generated English is becoming increasingly harder to distinguish, suggesting rapid improvement.
- The quality and detectability of AI text vary significantly depending on the specific model and the language.
Sources
- Ask HN: Can you still tell AI-generated text apart in your own language? — news.ycombinator.com