Apple Voice Dictation Writes 'Trump' When You Say 'Racist'
boehs.org
boehs.org
It’s likely guessing based on previously typed history.
Secondly, I find it interesting how it knows what the correct word should be, but spends some time on an obviously incorrect word before correcting. I think this points towards it being deliberate.
My suspicion is this: There is some kind of prediction for partially complete words, and somebody has either messed with the training data or the code that makes the predictions. The first part "race-is" could complete to "racist" or "racism" or something else. The second system then predicts based on the full word (given by the break), and the suggestion is corrected.
For those thinking "this is okay", just bare in mind that next time this could be a word association that strikes some form of emotion in you. Imagine losing your father recently, saying "Dad" and the correction suggests "Dead" -> "Dad", it might be insensitive.
... I mean, this is what LLMs are. They're word association machines (or more pedantically token-association machines). What we're seeing here is almost certainly an initial speech recognition model followed by another model correcting the output to what it 'thinks' is most likely/reasonable. AIUI this is how most LLMish transcription things work; this is why they sometimes produce weird results where the topic is _unlikely_.
I suspect it's not an LLM for dictation?
It's probably a network that converts sound to some phonetic tokens, then some kind of probabilistic matching like a Markov chain (which would be computationally cheap locally)?
Is it the vector space of two words being approximate to one another?