I've noticed a major gap in forums when we post in other languages. The detectors seem to give completely different results for a Spanish translation of the same technical post. Are these tools just English-centric right now?
3 answers
Most tools and their corresponding detectors are heavily optimized for English because that’s where the largest training datasets reside. When you switch to Spanish or Mandarin, the underlying "n-gram" probabilities change. A detector might not have a nuanced understanding of what "natural" burstiness looks like in another language's syntax. This leads to higher error rates. In fact, some studies show that non-native English speakers are often falsely flagged more because their English writing is more structured and follows "textbook" patterns, which look like AI.
Do you think as becomes more globalized, we will see language-specific detectors that are actually trained on local dialects and slang?
I’ve found that code snippets within posts confuse the detectors even more, regardless of the language used for the prose.
I've seen that too, Steven! The rigid structure of code in posts is almost always flagged as AI because code is, by definition, highly predictable.
We are seeing some progress, Brian. However, the investment is still mostly in English-language models for documentation. For now, if you are working in a multi-lingual environment, you have to take these "likelihood" scores with a massive grain of salt as they simply aren't calibrated for the linguistic diversity of a global team.