“Phonetic Overlap”
Tripp Mickle and Eli Tan (Hacker News, MacRumors):
While using Apple’s automatic dictation feature to send messages on Tuesday, some iPhone users reported seeing a peculiar bug: the word “racist” temporarily appearing as “Trump,” before quickly correcting itself.
The message blip, which was replicated several times by The New York Times, provoked controversy after appearing in a viral TikTok post, raising questions about Apple’s artificial intelligence capabilities.
An Apple spokeswoman blamed the issue on phonetic overlap between the two words, and said the company was working on a fix.
[…]
“This smells like a serious prank,” Mr. Burkey said. “The only question is: Did someone slip this into the data or slip into the code?”
Speaking the word racist with dictation doesn’t always show “Trump” first, though it did show up more often than other words in our testing. We also saw “Rhett” and “Rouch” appear before the iPhone corrects to racist.
The New York Times story cleverly omits the fact that other words like “rampage” trigger this glitch, not just “racist.” This leads some credence to Apple’s explanation that this is due to “phonetic overlap.”
Except that there is overlap with “rampage.” There are a variety of possible explanations, from a prank to some sort of semantic adjacency in a training dataset, but I think the idea that it’s because the phones are similar is laughable. I wonder why Apple PR felt the need to give this specific explanation instead of just saying that it was a bug.
Previously:
5 Comments RSS · Twitter · Mastodon
Unless there are a lot more examples of apparently unrelated words appearing, I'm going to go with "95% likely internal sabotage." I'm giving it 5% credence because "ra" in "racist" and "Tru" in "Trump" do sound somewhat similar, but I'd like to see other similar examples of this happening before believing this.
@Plume Yeah, if you look at the sounds from Apple’s own dictionary, the [r] is the only real overlap between [trəmp] and [ˈreɪsəst]. The vowels are different. The [m] and [s] phones are about as different as can be. To me, phonetic overlap would be something like Trump vs. Drumpf, where most of the phones are the same and the [t] and [d] are both alveolar plosives.
@Plume: yup, the spectrograms should be completely different.
Some systems do sometimes mistranscribe, because they are minimizing the likely error over a full sentence, and the LLM which is modeling the probability distribution wasn't trained on the right materials. E.g. medical transcription. But this seems very odd.
My guess is that they use 2 models (a simple one to provide immediate feedback, perhaps on device, and then another more complex one, perhaps off device, to improve accuracy), and the immediate feedback one has in its dictionary a decoding from the sound "Trump" to the word "racist".
Sorry, I got it the wrong way around: the sounds "racist" and "rampage" encoded as the word "Trump".
Reading the linked previous posts, this whole thing reminds me that autocorrect seems to be getting worse again in iOS 18 after hitting its peak in iOS 17.
When they first revamped it, it's like they cleaned out all the wrong things it had learned and started fresh and it worked great. It learned from me and finally worked well enough for me to leave it on by default after having it off for several years prior.
But now it's going downhill again for the same reasons it did before. It's not paying attention to me again, it's paying attention to random "corrections" inserted into the dataset by god knows who.
Worst of all, it's back to changing words I typed correctly into incorrect ones. And now this example of obvious sabotage, topped off by a nonsensical official response from Apple. When any response at all is rare.
What is going on in that Spaceship?
