I’ve been experimenting a lot with gpt-4o-mini, gpt-4.1-nano and gpt-5-nano in different languages, my native language being Italian, and whenever i ask them questions in Italian or other languages other than English, the result is very poor… not to say ridiculous.
This is only a fraction of a general downgrade i have noticed in performances in the last models…
some examples: i have asked my chatbot based on gpt-4.1-nano to summarize an article about energy in the atmosphere… my request was in italian, and the answer was in italian…
in all occurrences, ‘energy’ became ‘exergia’ instead of ‘energia’, about 50% of plurals were misgendered in italian, and complete sentences of 4 to 8 words where written in spanish(?!). Occasionally, complete words were missing, like: “parlava più ambivalente” that should have been “parlava in modo più ambivalente”, and sometimes the model would switch to English without any reason. i.e.: “Put together: cosa significa per l’affermazione originale” or “Un mondo più stabile, greener, con un clima più piatto”.
Then i changed the model to gpt-5-nano with verbosity: high the result was even more shocking: complete sentences in spanish, misgendered plurals, non existing words, and much more.
I then switched to chat in Romanian, and i noticed the same patterns (misgendered plurals in particular) and - in this case - an absurd loss of diacritics, that are fundamental in Romanian (like the loss of ‘ă’ that often bacame ‘a’, or the ‘ț‘ that was changed to ‘t‘).
It looks like, with the urgency to feed their models with always a higher amount of texts, they end up fed with unchecked and badly written material that is making the models ‘dumber’.
The result is that in most of my chatbots, at least those that i offer freely to the public, i was forced to switch back to gpt-4o-mini that still provides valuable translation and answering results.
That is - at the least - frustrating…