Hacker News Reader

Comment by Arnt | Hacker News Reader

I wrote something like this a long time ago. The customer's request was: "don't make this software produce swear words on its own". Note: on its own.

Producing "trump is …" when a user wants that is one thing. Producing that when the user wants something else could be construed as coming from the software itself, and we all know that people looking for a shitstorm will seize the opportunity.

So probabilistic software like a swipe interpreter will often be written to avoid some guesses, to lessen the chance of a shitstorm. Users who want that output have to use a dependable input method instead of one based on probabilities and guesswork.

IMO this avoidance isn't among the ten worst consequences of shitstorms.

View on HN

This is a very good counterexample, it made me switch from agreeing with OP to disagreeing.

As another example, imagine if "trans women are" autosuggested "men" or "ugly" for the next prediction.

This would probably be a bad customer experience for many people, even if Apple could prove these are simply the most common n-gram completions.

And if they altered their training data to avoid completions that are politically unpalatable, IMO that's actually much much worse than a grey list.

I'm a big free speech guy but this one is genuinely a bit complex.

Autocomplete should be following the user's own behaviour. If a user thinks that "trans women are men", then it will be not a universal experience, it is just this separated user's own statistics.

This is a good solution but you still have to deal with the cold start problem for the ML

Then you just don't suggest anything if there isn't enough data yet. Regardless, auto-complete suggestions is different from what TFA is about.

That's asking a lot given the current state of the iOS keyboard.

It already does this. I've found it'll suggest my kids' names frequently, after just 1-2 letters. It didn't do this before they were born. I assume it's training on my previous messages.

I never understood the memes about autocorrect until I got an iPhone. The keyboard is so much worse than GBoard it’s incredible.

I can imagine that autocompleting following users' behavior could bite people who do a lot of sexting or could reveal other secrets, ex. `the door code is`

Yes, please allow me to have corps not only know everything about me but have a hyper accurate way to know what I’m thinking as well.

Or we could just type the whole sentence we need.

Futo keyboard on Android has a transformer LM option that uses local ML for text prediction, I've found it to work very well. As far as I'm aware it's completely on device.

“Generated entirely on device” seems like an omission that doesn’t cover diagnostic, telemetry, “product improvement” logs.

A friend made a keyboard for iOS because he didn't like any of the existing ones, and it doesn't send any data anywhere: https://typeneuma.com

> even if Apple could prove these are simply the most common n-gram completions

Tech-illiterate online mobs enraged by Apple “making a statement” will not stick around long enough to hear them argue about n-grams or whatever.

> As another example, imagine if "trans women are" autosuggested "men" or "ugly" for the next prediction.

Or imagine if it auto-completed "the Holocaust is" with "a myth created by the ZOG"! I don't like these "imagine if X was Y" arguments because X isn't Y. Refusing to auto-complete "the occupied territories" (X) is really not the same thing as refusing to auto-complete antisemitism or group-based hate speech (Y). Argue whether X is right or wrong without changing the subject.

Most people are not anti semitic.

Most people are not against transsexuals and gays.

There is no reason these kind of suggestions get pushed.

At least in the western civilization, I don’t see why it need to censor anything.

Most people where?

In the Bay Area?

In the United States?

In the Middle East?

In India and China?

All of planet earth?

> I don't like these "imagine if X was Y" arguments because X isn't Y

What you're really saying, perhaps without realizing it, is "I have zero empathy or theory of mind for people who have different views than me"

The whole reason these arguments exist is because power changes hands. If you like a rule only because the person currently wielding power aligns with you, you don't actually like the rule. You just like power.

There is no winning here. It's impossible to make autocomplete inoffensive. In fact making it inoffensive, is offensive. theguardian.com would make Apple the bad guy no matter the policy choice.

That’s ok. I’ve already made theguardian.com the bad guy a long time ago so the two cancel out.

There is a way to win here: make it an opt out

- [ ] Disable autocomplete swear filter.

But finely grained options are not Apple’s way.

You'd end up with pages of options, especially once people realize they can get on the bandwagon.

Disable swears, alcohol, smoking, marijuana, gambling, child abuse and neglect, sensitive East Asian politics, sensitive Middle Eastern politics, WWII politics, Late Ottoman genocides, LGB issues, T issues, global warming, nuclear power, HOA issues, vegetarian issues, vegan issues...

I don't blame Apple for trying to straddle the line that makes everyone equally unhappy. But I steer away from autocorrect entirely so I'm not irritated daily by this like others may be.

Talking about autocorrect and irritation, I remember a situation where someone caused a minor international incident by writing that something "caused a lot of irritation". The word they were aiming for before autocorrect chimed in was "iterations" (on a document). I'd rather make my own mistakes than constantly have to fight an autocorrect system.

The problem here isn't a mere lack of suggestion. If it treated that situation as a completely blank context that wouldn't get in the way of typing what you want to type.

I don't even think it is complex; you can say what you like so there isn't a free speech issue here. Users are still going to type what they intended to type. Apple can reasonably have their software avoid making guesses about what users meant to say on charged topics.

It'd be a better if Apple provided a list somewhere for transparency but what they are doing isn't going to cause anyone any trouble.

An even more generous interpretation (not saying this is the case and a complete excuse, but might explain some of the choices) is that there are some phrases people do not want to type by accident.

“Male chauvinist pig” is something many would be very upset to send to their boss by accident. So it’s perhaps understandable some such phrases are not excluded from the guesswork.

What I don't understand is how you can type something that is close enough to “Male chauvinist pig” that an uncensored autocorrect will guess it, without intending to type “Male chauvinist pig”. That phrase and others like it, even if one or two letters are wrong or swiped through incorrectly, is unambiguous.

You can.

Step 1: a word of 4-6 letters that's somewhat close to "male". Say "name" or "masked". Look at the key positions, you'll see what I mean.

Step 2: a word that's even less close to chauvinist, and that makes sense after the word you intended, but no sense after "male", so the guesser picks "chauvinist", based on its incorrect first guess and your second swipe.

Step 3: ditto.

Even on a single-word basis, "male" might be make, makes, Maria, Maia, narrow, Maddow, marrow, Mariya, Marita, mafia, males, mallow, Maslow, male's, or masks. "Chauvinist" might be coatings, costing, coaxing, coasting, crating, contingency, contingent, contingencies, chauvinistic, Chauvin, or chaining. "Pig" might be PC, PIC, PCC, pox, poof, PVC, pH, put, pug, PV, Pooh, log, or pub. These are all possibilities that were actually presented when I tried to swipe these words on my Android phone, twice each. Taking into account additional context will broaden this set of possibly confusable words, not narrow it.

We might not talk about a "narrow contingency put" or "Maria crating PVC" very often, but an uncensored autocorrect would almost certainly correct those phrases to the far more common "male chauvinist pig" in the rare cases where someone did try to swipe them.

One of my friends is famously bad at texting. She’s even posted antisemitic stuff by accident. This is on Android.

I know these are all typos because she’s a really close friend and easily the last person who would be offensive to other people.

So yes, you’d be surprised at what people can typo on a phone.

"Male chauvinist" and "Male chauvinist pig" are related in meaning but not equivalent.

The difference matters enough that preemptively suggesting "pig" is not a desirable "service" for most people.

Isnt the regular keyboard on iOS typing also probabilistic? I thought the entire claim to fame of the original iPhone keyboard was how it could determine which key you meant based on letter frequency / word usage / etc. Does that mean it can also stop you from typing certain things? So whats the difference between the swipe interpreter and the fat finger interpreter?

A matter of degree.

The regular key-by-key will also do some probabilistic changes. Just less than the swipe interpreter.

"The customer's request was: "don't make this software produce swear words on its own". Note: on its own"

I'm against Apple's attempt to meld popular speech into a form that it thinks is acceptable on two accounts—I'm not the fastest typist so when prediction doesn't work the way I want it to it slows me down. Second, I'm in Australia and Australians use very colorful expressions in everday speech that seem to trigger the sensibilities of many delicate American ears. In short, we don't want attempts to censor what we say.

If one has to introduce censorship into prediction then users should have the option of either turning it off or selecting a graded system that best suits their style.

In fact that's not a completely new idea. Eudora email had an option that I rather liked, when selected, users could type whatever they liked then depending on the degree of offense or profanity a series of red chilli icons would appear. Mildly offensive text would bring up an icon with a one-chili warning whereas highly offensive language would automatically trigger a three-chili icon along with a message to the effect "They're burning hot words, do you really want to send this message as is?"

A scheme along these lines seems eminently more sensible than secretly censoring a predictive word list.

If you were relying on Apple to predict your speech in order to speak, then perhaps it isn’t your speech being censored. Speak for yourself; don’t complain that Apple won’t speak for you.

Censorship is dumb but required by a lot of games to avoid higher age ratings. A ridiculous example- my friend tried to name himself "Horse" in Elden Ring but got "**rse". This is the Scunthorpe problem.

Reply on news.ycombinator.com