This same AI effect can be seen in the Android keyboard, where _properly_ spelled words will be replaced after typing another word or two because it's been determined to be more likely what you want. It's infuriating.
It actually does this with such consistency that I think it's a very specific mistake. For example, I frequently swipe "See you soon." It always, always renders as "See you son", which I then have to manually retype. Sometimes twice. I don't have a son, and I'm not a blind old cowboy who jocularly refers to any random person as "son". I honestly just want to type "soon", for the love of... anyhow, this is an ongoing, totally inane battle of wills with my phone.
I think what's happening here is that there's a very impressive and sophisticated heuristic for predicting the probability of what you want to type by looking at the frequency of what you have typed in similar contexts. It uses its state-of-the-art AI to evaluate the context and build an array of candidate words, along with their respective probabilities. I suspect it is very accurate as it does this. Then it sorts the array by probability and pops the top element into the predictive text input.
Alas -- per my pet theory anyways -- it sorts like this:
When a system tries to do something automagically and makes a mistake, it is very frustrating, especially because, to allow seamless large changes, hide competitive details, or make the UI more "streamlined", such systems rarely give users options to tune the results. A system that gives controls to the user and expects them to tweak their own experience is so much better in my opinion, except in the metrics of first-time usage (or first-time-since-major-change), when those controls look like information overload and make the system seem like something that must be learned before it can be used.
And yet, when the latter inevitably breaks on an edge case, users can try to fix it themselves. They don't hit a wall of frustrated "I can't do anything", they hit a challenge that they are empowered to try overcoming. They already know what they want and can set things that way, rather than trying vaguely to teach a system (machine learning, hardcoded heuristics engine, department of humans making seemingly unconnected changes to a GUI with each passing version and no obvious plan) to understand their desires.
I miss the days when users were seen as intelligent professionals who are willing to change settings, create and re-dock an assortment of toolbars to every edge of every screen/window to suit their daily tasks, read a manual (or at least search the integrated help entries) to overcome problems. Rather than "busy" phone users who just want to complete a task with minimal time spent learning and get back to posting on facebook or whatever, and who accept the automagical solution because adequate results instantly are somehow considered better than great results with some work.
Ugh, that whole block of text just kept growing; I had better leave and go ramble/rant at trees or clouds or something elsewhere.
There was a time when autocorrect on phones was moderately useful. That time has past. One of the first things I do with a new phone now is turn off autocorrect completely. It doesn't bother me to manually correct my own mistakes, but it bugs me a lot to have to correct the mistakes of the freaking robot that's supposed to be saving me time.
Wow, I... didn't know that gesture was a thing. Thank you! It seems to help a bit! I'm now getting a 50/50 son/soon ratio. That said: when I manually type "soon" -- definitely with a double "o" -- it still autocorrects to "son" on the first try. So, going to keep my pet theory intact.
The weird thing is that if I type nothing at all, the contextual predicted "next word" on GBoard is actually very good -- I wasn't praising it for comedic effect. But it really does seem like there's a sign error in a sort function which kicks in after you start typing.
I'm pretty sure Swype did explain it, and I know SwiftKey has this gesture as well (though I don't remember if it was ever explained to the user or just assumed they'd remember it from Swype).
Swipe style keyboard? How does that effect spelling ? So confused - I get how it changes to "predict," anticipate or whatever the AI engineers say but what I don't understand is changing a real word to a non-word. That's not intelligence. That's something else and I don't know what they get from changing word to a misspelled and nonexistent word other than eventually driving the human race insane: I'm going to be a Luddite;
Er, "son" and "soon" are both real dictionary words. Swipe-style keyboards use an internal dictionary to find the most likely match to the swiped pattern, and have methods for adding new words (mine for example automatically adds any tapped-out words after you hit space).
Very likely, every time you type "soon" and it corrects to "son", the probability of the change gets increased. So, the more it errs, the more wrong it will be in the future.
About it only replacing after you start writing, the probability of "son" must grow faster than the one of "soon" after you type "so". If there were such a huge bug, you would be seeing a new word every time, not always the same one.
Try switching between languages. I work in English, my wife is Italian, my colleagues are french speaking, I am Spanish and have Catalan friends, and I live in Germany. I gave up on predictive keyboards long ago.
You don't even need foreign languages to run into this problem. English is the only language I speak fluently, but I'm Australian. I'm
also a programmer, and I have some american friends. So, depending on context sometimes I spell 'color' (programming or talking to Americans), and sometimes I spell 'colour' (talking to Australian friends and family). Same with behaviour / behavior, favour / favor, etc. The context for which spelling I decide to use is complicated. In the same document I might name the `getColor` function, but describe it as getting the colour of a pixel. I might have two chat windows open side-by-side with different people and in each window spell the same words differently.
All my devices insist on shaming me (or autocorrecting me) for one of those spellings. At this point it feels like a complete gamble which I'll get corrected on. I'm just slowly getting used to correcting the autocorrect. :(
When you said this it reminded me that css allows colors to be either ‘gray’ or ‘grey’. Which I’m glad because then I don’t have to fumble until I picked the right one. Though, I’ve learned to type hex colors (especially grey scale colors) intuitively now, so I usually use those a lot more than typed grays now.
SwiftKey seems to manage three languages (Finnish, Swedish, English) simultaneously pretty well. No need to explicitly switch language either, it figures out the current language as you type.
Yes, SwiftKey works almost perfect for me too in three languages scenario (Bosnian, English, Dutch).
The only thing it is confused about is the letter "i", which means "and" in Bosnian, and SwiftKey often capitalizes it where it shouldn't. Probably happens because I often mix Bosnian and English in the same message (instead of translating technical terms).
I have 3 languages in my GBoard. I've just arrived from a Spanish speaking country, and now when I try to write in Portuguese it still completes with spanish words. Sure they are are very similar languages.
I'm on the same boat, I'm Italian and speak Spanish a lot with friends, English is my daily language. The latest Google keyboard has helped but it's not nearly as good as T9 was.
It's all the keyboards. Swype used to work great, but now when it actually gets words right, if I have sentence with a second word that could possibly have been two similar words on the path it will just straight up replace both rendering the sentence completely incomprehensible. Who are these people that don't correct their sentences until the second incorrect word?
I second that. 5 years ago I could swipe a whole message blindly with no errors. Now I have to correct every second word.
I'd love a feature to disable all that deep learning and AI and just use the algorithm they originally had (proximity of where you typed to words in the dictionary). That worked so much better.
I'm glad it's not just me that's seen Swype getting progressively worse! Either you have to really emphasize what letters you want, or you give up and type it out. I'm about finished with it.
I had a Galaxy S3 and was a heavy user of Swype. My friends marveled at how fast I could type with it. It was perfect! I recently changed my phone to an S8 and Swype became unusable. It gets almost every second word wrong, so much that I'm thinking of disabling it entirely :(
Speaking of virtual keyboards I always liked (and still use) the one from Blackberry and I never had such problems (you should be able to install Blackberry keyboard on any Android phone). I switched to Google Pixel from iPhone once I found out I can use Blackberry keyboard there.
Oh man, Android does that too? This awful, terrible, no good fake "AI" behavior on iOS was one of the many quality issues with modern iOS that was making Android look more attractive.
iOS keyboards have been getting worse with every update, probably as more and more engineers feel the need to make a mark, or are required to fix bugs, and they spoil it. The simple and predictable statistical model of the original iOS was better than what we have now. So much of iOS was better back then, IMHO.
On BlackBerry if it autocorrected to something you didn’t want, one press of delete would revert to your original word. iOS makes you a) rekey the whole thing and b) will probably try to change it again. I can only assume that no one (or moons, it literally just tried) who works on this uses it themselves!
I always thought Blackberry had the best autocorrect experience. My old Blackberry would occasionally surprise me by catching something I wouldn't think it would know about, and after some initial tuning never ever frustrated me. It only enhanced my experience.
Call me a crazy person, lately I've been thinking long and hard about going back to the BB (BB Bold 9930) and am willing to make all the sacrifices that come with it for a lot of the frustrations with smartphone OSes listed in this thread. Maybe I'm getting older and my demands on what I expect from a phone are beginning to normalize and simplify: Text, Calls, Email, a browser for sports scores and reading news articles on the train.
Honestly it's very probable that I'll only ever keep a smart phone around as a music/podcast/audiobook device.
I never used a BB, but I did have a bunch of "dumb" cell phones, and I miss them. Making calls on them was easy: Flip open, press Talk, dial, Success! Now it's: Turn the phone on (which takes approximately 60 seconds on my Samsung because it can't see keystrokes until it finishes trying to connect to wifi), hit Home button, wait 3 seconds for that to work, hit Phone icon, wait 2 seconds, hit keyboard button, wait 1 second, dial phone waiting 500 ms between each key, press Call icon, wait 20 seconds for call to go through, hit speakerphone button (because as often as not the cheek sensor fails to work and my cheek disconnects the call), Success!
As the saying goes, smart phones are just pocket computers with shitty phones attached.
This is a Galaxy S6. I'd probably get better performance if I didn't keep it in low-power mode, but I have no choice because the quest for thinner phones means the battery only lasts 2 hours in regular-power mode.
That's not true for this situation. For the "deep" replacements more than a word back that we're talking about, and also other changes that happen without any user notification via blue text popups, there is no easy way to fix without the long process of moving the cursor.
4 years ago was a very different time. Software quality in general continues to drop. I’m not sure if that’s due to increasing incompetence in the software engineering workforce (unlikely, but possible), malice (more unlikely), or apathy (most likely).
When wages don’t grow over 10 years, what incentive is there to write the best software you can?
I suspect the cause is a little more subtle (and terrifying).
The majority of people don't give a shit. Correct spelling and obscure searches are not even on their radar, it's not a part of their reality. Don't let the comments here fool you -- it is a very specific, picky, technical crowd that frequents HN.
The voice of "those who care" has always been a minority, although it used to matter more, simply because people who care and worry and try to do a good job tend to have more power and money (conscientiousness is a great predictor of success), and so businesses cater to them more. Now that everything seems to be turning more uniform, more global, more binary, more equal, that voice is marginalized (good thing? bad thing?) -- you're seeing the effect of a hoi polloi stampede.
So it's not the fault of "incompetent programmers" -- it may be a trickle down effect of our social incentives and economic trade-offs.
> The voice of "those who care" has always been a minority
To add to this, people who cares most powerfully had probably switched to alternative, open source, software solutions. This leaves the remaining group with less "care" on average so fewer would complain. Kinda like evaporative cooling.
> 4 years ago was a very different time. Software quality in general continues to drop. I’m not sure if that’s due to increasing incompetence in the software engineering workforce (unlikely, but possible), malice (more unlikely), or apathy (most likely).
I think another dimension is how deployability has changed.
Before... When you wrote and shipped software, getting your software out was a big problem, a big deal. This also meant that if you shipped a bug, shipping an update would be equally expensive (for you and your customers), and the amount of goodwill you lost would be quite tremendous.
Now everyone has a appstore, always up to date apps, and whatever else is usually "in the cloud" somewhere. The time of people installing applications in a normal desktop-context, with installers and having IT-administrators handle updates once every second year is surely long gone.
With that kind of change, and an increased focus on delivering early, doing proper QA is no longer something which is rewarded in the market.
Who cares if you made a bug-free, awesome service, when you did it 6 months after someone else shipped a similar, but buggy service which everyone is already using? They have established a user-base and as such already has social momentum and lock-in.
What do you have to offer which is not only fantastic enough to make some bother migrating, but also so amazing that these people will also go convert their friends and families? "Less bugs" alone is not going to cut it.
Basically, taking the time to deliver quality software these days is increasingly something you get punished for in the market-place.
The result? We get shit like this and we can only blame ourselves.
I find it more reasonable to assume that you can only improve a very specific function so much before your "improvements" turn to "pointless sidegrades".
And they can't even fix a simple typo like this after years of usage - "th8s 8s example tex6". Come on Google, you made that UI, you know that i and 8 are next to each other, you have a database of correct words and probably of typical errors and typos, wtf?! (you even know how to correct "wft" to "wtf" and can correct simple word with number typo).
Yes! Those one-off errors used to be fixed by most autocorrect systems. Seems like whoever wrote the new AI-based systems didn't make a checklist of existing features and attempt at least parity before switching over. It's embarrassing.
I don't know; I'm pretty happy that I don't have to type apostrophes any more. I can just go on typing words like "ill" and "wed" and it'll figure out after a few more words that those should be "I'll" and "we'd".
What the parent comment is talking about is something more extreme and I've noticed it too. It sometimes changes prior words that are valid after you have moved on to the next word. It's not correcting the word you just typed, it's correcting a previous word without any sort of feedback like you get for normal corrections- it's going backwards and changing earlier words, and then you try and fix it but the exact same correction applies automatically again over and over.
Unfortunately I can't remember any examples at the moment, it's just something that happens to me every so often. They're really irritating though, because they aren't well expressed in the current autocorrect UI (which works on the current word) and it doesn't seem to get the hint when you go back and correct it, so it keeps applying it over and over.
No, that's what I was talking about too—it needs the context of the (part of speech of) the next word to figure out in the cases I mentioned whether I actually intended the (correctly-spelled) word "wed" or the (correctly-spelled) word "we'd". It doesn't change it until after you hit space to commit the word that comes after the "wed" input.
Ah, ok. I'm pretty sure I've had it happen with more than just contractions though, like with common phrases. The real problem is that it's really hard to undo the correction. I need to start keeping better track of it so I can file a bug report... it makes it really difficult to type certain combinations of words.
If I type "its" and "it's" correctly, they are regularly changed to the incorrect spelling. If I don't bother, they are not corrected. The only way to get correct "its" and "it's" is to go back and fix them after auto-correct has screwed them up.
It's probably because the spelling is corrected based on a machine-learned model, whose corpus is likely to contain many instances where "its" and "it's" were swapped.
Which means the corpus is broken. But regular people rarely care about correct spelling, in my experience, and so I doubt corpus maintainers will care either...
You're imagining that people who make NLP corpora actually vet the text going into them? I dream of a world where people can be convinced to care that much. I'm not even talking about the scenario you suggest of filtering for proper word usage, I'm talking about filtering at all.
The corpora used for popular word embeddings are full of weird nonsense text (in the case of word2vec) or autogenerated awfulness like spam and the text of porn sites (in the case of fastText) or both (GloVe). And most people who implement ML don't care how their data is collected.
I mean, nobody expects the engineers to manually read through everything, but if the quality of the input text is significant for the quality of the autocorrect (or whatever other application you're using machine learning with), you kind of have to make sure the input is pretty good... You could for example choose datasets which is expected to contain mostly correct grammar and spelling (such as Wikipedia, books, etc.) rather than datasets which is expected to contain mostly incorrect grammar and spelling.
Or don't use a machine learning model. I honestly don't care, just don't automatically turn a correct "its" into an incorrect "it's".
Wouldn't late edition books with only corrected text be better, proofread, edited, proofread, edited, ... Google have millions of them they've assumed copyright of. Surely there's enough text there. Do they really just use random website text?? Nearly every news story I read has errors and they have style guides, trained writers, editors, etc..
Do publishers sell their published text as a mass for use in AI/ML? Like 1000 books, no images or frontispiece, etc., possibly jumbled by sentence/para/page.
They say they welcome contributions; I don't know if they just mean new sources of text, or if this includes code for filtering or fixing their existing ones.
I've been trying to figure out how to disable that! Anyone have any ideas? Google search turned up no results, but after this thread I shouldn't be surprised.
that's not Android keyboard. That's Google keyboard. Solution: use a different keyboard. I've used the AOSP keyboard since I started using Android. Sure it's made by google but it does what I say rather than what some proprietary idiot robot thinks I think. While your'e at it, use the send feedback function to let them know you hate it.
Its annoying autocorrection tendency to choose 'fir' instead of 'for' frustrates me. People almost never use the word 'fir', but use the word 'for' often. It would be nice if you could blacklist words you want it to never choose.
There is SwiftKey, on the hand, that does those kind of annoying corrections a couple of times, remembers your choice, and does them no more. It's been a long time since I've seen a 'fir' with SwiftKey.
On my S8 keyboard the configuration menu can be opened using the gear icon appearing in the keyboard. In the configuration menu there should be a bunch of options under "Smart typing" including 'Predictive text' which can be turned off.
This is in the european market, though, I don't know if it's configured differently for different markets.
Here's the problem. I already turned off all those options long ago, but the annoying behavior remains. To be clear, I don't mind the keyboard predicting things. It's only the retroactive changes that bother me.
Someone else suggested Swift keyboard, so I'm going to give that a try.
