Changing the bot's locale from en_US to en_GB results in a MASSIVE drop in accuracy - nonsense sentences are being matched to intents instead of to the fallback intent

I've recently tried to change my bot's locale from en_US to en_GB as my bot is being used in the UK.

With en_GB as the locale, nonsense sentences such as "khk djh kdjhc lkjf" are being matched to real intents rather than (as I would expect) the fallback intent. The training data is identical and there is no particular regional language being used. It's a v2 bot with around 100 intents.

In fact, it's virtually impossible to trigger the fallback intent with any combination of input text.

Has anybody else experienced this? Is there a workaround or is the en_GB locale just significantly worse at recognition?