I avoid speaking to AIs in anything else than English as the results are almost always worse

I'm not sure how much the harness affects things, but the Deepseek web chat keeps trying to talk to me in Chinese. I tell it to use English, and it "forgets" a few turns later. I wonder if I'd get better results if I could read and write Chinese.

It depends how you measure "worse".

I'm using it in my native language, in hope this can escape some dumb guardrails. Recently Sonnet put a word partially in Russian (cyrillic) in its output instead of my latin-alphabet based language. I suppose that this kind of mishaps is less likely to happen in English.

My experience of N=1 is that this is true for most professional contexts, except for Legal and Fiscal queries.

Likely related to corpus but questions asked in these domain knowledge areas are not nearly as accurate and specially not nearly as complete as when asked in a native language.