Thanks for elaborating on this. I couldn’t find the right words, but as you said, if a user is using a different variety than the target variety, it already becomes primed for that variety of the user, not the target variety.
Which is why, until we see actual input/output pairs in full here, it’s tough to assess what is going on.
On the flip side, if you do take a moment to think about it, I’m sure the vast majority of training data from the internet uses American English, and there’s likely a lot more data in American varieties than others. If we think about how these systems truly work, where it predicts the most likely word that comes next, then choosing “apartment” instead of “flat” may be as simple as that word being more probable to occur next according to the AI and its training data. The other thing to notice is that any guardrails or manually implemented text of any degree are likely using American English, because OpenAI itself is an American company. Again, it all depends on what the responses and prompts were that’s causing the undesirable behavior.
The kinds of queries and what requests are being asked may also play a role here. If people are getting frustrated because it’s using “color” instead of “colour” when prompting (or generating prompts) for DALL-E, that is something more out of our control because of the model quirks and how particular image gen models need to be prompted to elicit a desirable output. If people are trying to generate a story or some form of writing that requires british varieties of english, that becomes a different problem.