A critique of Gmail's “smart replies” | Hacker News Reader

A critique of Gmail's “smart replies”

medium.com

146 pointsuserbinator111 comments

Open articleView on HN

"As a Google ML ethics guy, I need a drink." https://twitter.com/Theophite/status/1060991160494120961

I started reading the thread and laughed fairly loud and disruptively when I read this:

so, it turns out that our users are using our product for...

[spins Wheel of User Behavior]

... divination.

https://twitter.com/Theophite/status/1060991848280256512

The author seems to imply the smart reply buttons immediately send an email with the chosen phrase. This is incorrect. They open an email draft beginning with the chosen phrase.

There are many times I've used one of these buttons to begin a much more detailed email. Not only do they save a few keystrokes, but they help you get past that first wave of "blank canvas" writer's block.

Heh... despite seeing them on literally every email I've responded to, so possibly hundreds of times, I have never been okay with sending a derpy "That sounds great!" email, so I never realized they don't automatically send it.

It's amazing that they managed to fail at discoverability despite putting the controls right in front of me.

I have the same problem with messaging apps, where I can't tell whether pressing return inserts a newline or sends the message. I find myself avoiding newlines altogether because I'm scared of pressing return. Or, accidentally hitting return and sending a half formed jumbled mess to my boss/customer.

Please, developers of the world, make return insert a new line. It's much easier to undo than message-send.

I hate trying to guess whether a newline is enter, shift+enter, option+enter, ctrl+enter, or cmd+enter.

where I can't tell whether pressing return inserts a newline or sends the message.

Do you have any examples of messaging apps where return/enter inserts a newline? I can't think of a single one where it doesn't send the message, probably because inserting a newline is an uncommon case but sending is not. It's been that way since the days of IRC, from what I remember.

There's always been a division between "(IRC) messaging" and "(BBS) forums", and the relevant influence-trees, with I think the biggest line drawn is how the return key is handled. From this one UX feature, every other differentiation is natural.

This division was fine, because they served different purposes. Messaging (return => send) was always short term, quick communications; Forums (return => newline) was always longer-form posts, sentences to paragraphs, in a single post. The forum for discussion worth archiving, and the messaging app to replicate daily communications.

But slack and its kin crossed that line. They pitch themselves as the utility of forums (archiving, search, "starring", etc), with the quickness of messaging (rapid response times, notifications, etc), and now everything is fucked. Forum-like conversation get stuffed into a Messaging application, and gets broken up into a hundred different messages (ala 1/20, 2/20..20/20 twitter multi-posts), perhaps collapsed into a single proper post but usually not; Logs get stuffed into it, into a separate channel whose sole purpose is to be muted. Documents and emails and any work gets stuffed into it, because it can be made to fit.

And naturally, the archiving, the search, the document storage, everything breaks down, and its awful. Until you realize that you can't use slack like slack wants you to. It's a better, smoother, more maintainable and convenient BBS forum, trapped in a messaging shell.

The return => send is feared, and it is a fear special to slack, and its many recent derivatives. Because slack crossed the line, without crossing it. Messaging apps should have return => send; slack (and its kin) should have return => newline, because it is not a messaging app. It is not IRC, and it is not counted amongst IRC's friends.

It is a forum in denial.

I don't know about iPhones, but on Android it probably depends on keyboard installed/configured itself, until developer forces its own behavior.

I've seen only single app that misbehaved and was sending the messages on return key - Xabber. Others just let me insert newlines as I'd expect.

Lots of apps. I use FB messenger, WhatsApp etc. if I hit return it goes to a new line. Maybe somehow I enabled them in the options? I don't know.

In the default iMessage for iOS return inserts a new line and the > button sends.

Telegram inserts a newline. On Android at least.

And then for the places that do enter=send, they'll have ctrl+enter or shift+enter as newline, and vice versa.

Looking at you slack.

Doesn't Slack have an option for this?

Better than that: in Slack, it changes when you type triple-grave to open a code block. I see the logic, but it's nuts from a UX perspective.

I wouldn't mind so much if Slack and other offenders had a delayed send option so I could undo my f*ups.

Exactly, I've never dared to click it assuming it would send immediately.

Same here. This is news to me.

The text message suggestions send an immediate message. They're super-useful.

Is this perhaps different on mobile vs [desktop] web?

Text message versus gmail, actually. And the difference makes a lot of sense since no one expects a single text to contain all the information of a reply and also texts are more likely to be time sensitive and so reducing friction to answer is good.

Annoyingly, LinkedIn's messaging system does this (both on mobile and desktop). I probably don't mean to send a two-word response to a long message from a professional contact.

But for some reason, LinkedIn has twitter envy, or at least Facebook envy.

It's like how Nature seems constantly put out that it stuck being the premier journal for high end research and wants to make it's presentation and editorial content look as much like Scientific American as possible.

Ha, that bit about Nature is spot-on.

Be careful with that draft - the smart replies 'reply all'. Found this one out the hard way.

In my experience gmail is now offering to complete your sentence. I'll be a few words into a sentence and it'll give me an example of what it thinks I'm writing (in light grey text). If it hit enter it'll fill it in fully. If I keep writing something else it'll go away. It's decently helpful. I find it akin to something that Grammerly is doing but more visually low-key

I actually find this slows me down, even if I hit tab to use what it's suggested - it's distracting having the suggestion change as you type, and I find myself stopping to read what it's suggested before accepting it or restarting typing, which loses my thread of concentration.

LinkedIn embarrassed me a bit with one-click-sending a line I thought I could edit/extend before sending.

It's hard for me to understand why the fact that Google's autoreply exists is objectionable. If you don't like the response, don't use it. Nobody is the poorer. For those looking at a more nuanced critique, the Baffler published this:

https://thebaffler.com/latest/nudging-the-lexicon-haigney

I think it captures the effects of autoreply more accurately, at least the effects of a writer living outside of Silicon Valley.

While Google's suggestions may often be off-base, they are also often reasonable, but they may shift your register a little. That is, they are designed for people using email to make quick decisions and respond in a professional way.

I'm curious, how do you see that critique as materially different? Seems to make the same two points, the influence on human language and its role in strengthening Google's data dominance.

People can object to auto-reply's contribution to the degradation of conversations, even if they, themselves, don't use it.

That was the topic of about half the article.

I've gotten email responses that look like a suggested autoreply and it makes me judge the person and value of the conversation.

The person may have typed it but it comes across that way.

Happy to use an email provider that doesn't have a bot reading my mail.

The suggestions probably come from people’s email, so as the system gets better, you’ll be judging someone just because Gmail later plagiarized them.

I've always taken these Smart Replies as equivalent to Facebook Messenger's "button that sends a big thumbs-up emoji to acknowledge something." (It replaces the Send button when you don't have anything in the input box.)

I.e., these responses are there to serve as various textual forms of "I acknowledge that I received your message and am hereby discharging your social expectation that I will reply to it, by doing so with an information-free message."

I assume that Gmail would just suggest using an emoji rather than text, except for the fact that 1. etiquette says that an emoji is too "casual" for office work, but a one-word answer is just fine and exemplary in its professionalism; and 2. there are devices still in use that can receive email but are old enough that they can't display emoji, so textual equivalents are a better lowest-common-denominator.

For an interview last year, I interacted with Google recruiting almost entirely using smart replies. One of the suggestions to the initial email was fit to send without modification and the idea was so amusing to me that I stuck with it as much as possible throughout the process. It wouldn't surprise me if the other end of the conversation was mostly machine generated as well; two Google systems talking to each other with a bit of human editing here or there.

I did this recently as well! It kind of reminds me of a Black Mirror episode, “Hang the DJ”, where a dating app used ‘cookies’ to arrange matchings between people. I actually wished someone could invent this for the recruitment process—a stand-in AI that deals with all the boring parts of life.

There was a thread a few weeks ago about a guy who did a bot that responded to recruiter asking them relevant question about job (salary, relocation etc...).

That's funny. I might have to try it at some point!

What I find more disagreeable is the newly enabled autosuggest feature. Few things are more annoying than seeing bad suggestions pop up for every word, on a desktop device where I type at 100wpm.

Edit: word.

I'd say that's true efficiency. They are managing to annoy you at a rate approaching 100 times per minute!

A perfect example of why efficiency is not always a good thing!

The robot uprising has already begun!

Annoying us is probably their maximum capability for now.

https://what-if.xkcd.com/5/

Then turn it off? I got a pop-up asking me if I wanted to enable it (I did), but you can turn it back off by going Settings -> General -> Smart Compose -> Writing suggestions off

I never saw a pop-up, but I did take the time right away to disable it.

I really like these suggestions -- when I see one, I stop right away and rethink what I'm writing. If my sentence is so predictable that a machine can finish it, I'm not writing something that's going to be informative to its recipient. Or I'm using too many words to get my point across.

> If my sentence is so predictable that a machine can finish it, I'm not writing something that's going to be informative to its recipient

Wouldn't a common phrase or expression be easier to understand for the recipient?

True, but the suggestions are so conservative that anytime they pop up, it's a sign that I'm being too wordy. A familiar five word phrase might be easier for the reader than an unfamiliar phrase of equal length, but no words at all is even easier. I've consistently found that, when a suggestion pops up, I can safely drop most of the sentence without denting clarity. It's like spell-checking for semantics.

When you respond with “Thanks for sharing!” or “Glad you enjoyed it!” or “Very cool!” to an email message, you are not responding as an in-the-wild human. You are responding as a human that has been prompted by a robot.

I know someone whose blog is configured to filter out such vacuous "empty praise" comments, because 99% of the time they are from spambots.

In fact, that could be a feature. Instead of “smart reply”, how about “smarter not to reply”:

“The reply you have entered is not actionable. Do you still want to send it?”

Nip vacuous e-mail in the bud.

Gmail sells ad impressions. So less email is not good for business.

Other possible consequences: https://www.smbc-comics.com/comic/gmail

I kind of expected to see that in OP, but apparently they were trying to be Serious.

With the largest pile in the world of written natural language representing dynamic human communication, Google is able to use deep learning, combined with the methods of Natural Language Processing (NLP) to draw inferences about the essential content and intent of your emails and sort them into types.

Using these same machine learning methods, a trio of possible responses to a given type of email message is generated and tested. Every time a human uses one of these three responses, a datapoint is supplied to Google that says “given message type A, a human has chosen response X as an appropriate one.”

Multiply this last step by a thousand or a million or a gazillion, to the point where a clear statistical pattern emerges, and Google can conclude with some confidence that when a person expresses ideas, thoughts, feelings, or questions that can be classified as type A, it’s reasonable for another human to respond with utterance X.

This is usual attribution of magical powers to "AI" (and really they mean "ML") by the people who don't know how it actually works.

In this case:

* Sensitivity detector checks if email is sensitive, smart reply is turned off for sensitive emails (obviously didn't trigger in this case). I personally think this is questionable feature - imagine your keyboard turns off if it thinks topic is sensitive.

* Encoder model produces embedding which is compared with pre-computed and pre-clustered embeddings for whitelisted replies. Top replies for three top clusters are selected (selecting replies without clustering will produce very similar replies). All potential replies are white-listed for obvious reasons.

While clicks produce useful metrics (really, this is the target metric), they don't produce useful training data for two reasons: (roughly speaking) you can't train a model on it's own output and there is no way for model to learn to suggest something new. Fundamental challenge for training such models is that production metric is different from training metric: model is trained to select likely reply that was typed, production metric is reply that is clicked.

There is bunch of blog posts and papers from Google Research describing how it works.

My main concern with things like this and autocomplete is can they actually influence how the user thinks and what they say? It would be interesting to have users write about a subject with a "good adjective"/"bad adjective" biased AI and see if there is a measurable difference in opinion on the subject afterwards between groups.

For Google, Facebook and others, as a companies, best users to maximize long-term revenue are people that have income and choose to spend all their disposable income (and thus generate ad and lead conversion profits). As such, Google should show/suggest/autoplay search results/ads/YouTube videos/smart replies/assistant answers/News/Google Now cards to increase life-time likelihood for the given user to become working professional and mindless spender.

For specific hypothetical example, it's not hard to imagine, that training users to prefer things like "experiences" or entertainment that need to be re-purchased is better that promoting physical goods, like board games or bicycles.

This has got to be some sort of fallacy. It rings similar to the "if evolution selected for good traits, why can't we fly and shoot lasers out of our eyes?"

Yes, its to their benefit if their users have money. No, I don't think there's any chance that any company is able to optimize for someone gaining a middle-class life that meaningfully.

Companies do small-scale user studies and measure things like engagement and increased clicks on a small scale. They do not perfectly capture their user's income and spending level for many years and correlate that well enough with the company's auto-suggestions that said suggestions optimize for that very macro goal over the very micro measurements they make.

I think it's very hard to imagine Google successfully optimizing itself to meaningfully lead do that outcome for users, just as it's very hard to imagine humans evolving the ability to fly.

Current recommendations systems for FB, Google/YouTube have switched from models that try to increase immediate engagement to models that try to increase retention because it turned out that optimizing for immediate engagement decreases retention. (Latter is obviously much harder ML task requiring Q-learning, longer data series, etc.) In other words companies choose to sacrifice immediate revenue in order to increase long-term (optimally life-time) revenue from each particular user.

You can see it an example of early precursor of ML that shapes user behavior: learning what motivates user to return and then displaying it.

If you write an email to someone, and they respond with one word "Interesting", you should take a hint that you're boring them out of their minds. Don't blame robots for that :)

I do find it surprising how often a short reply (of any kind) just isn't appropriate, but Google goes ahead and makes a suggestion anyway. I would have thought that "no quick response" would have been one of the options they trained against.

They trained against the champions of get-the-last-word from all over the internet

Gmail Smart Replies often suggest things that I’d like to say to my fiancée but have the common sense not to.

My android has suggested "Thanks Honey" as a SMS reply to people where that wouldn't go over well. Gmail hasn't done that yet.

Use them if you like, don't use them if you like, write a self-aggrandising piece to draw attention to how smart you are if you like too. This is faux-problematizing.

Next thing you'll say is I miss the world where everybody's mail looked different because of unique handwriting, and this email thing has made me look like a robot.

Auto-replies save me the hassle of personally typing out an acknowledgement of receipt or confirmation of a meeting, which I find very useful. I don't get the problem with an optional feature which clearly makes certain types of common emails easier to deal with.

> The thing to remember about email is that it is a conversation.

I think this is a big claim and not obviously true enough to just assume. Email is just a medium, and can be any kind of communication within that - just like not all spoken communication is a conversation.

Nah, it’s fine. I like it. I have autocomplete in my IDE but I sometimes write unique code.

Can't read article, it wants me to pay $5.

Obligatory life-saving browser extension: https://makemediumreadable.com

Why do people even use Medium? Doesn't this kind of stuff go against the ethos of being a computer scientist?

Easier to set up / integrate and looks better than most Wordpress themes that I’ve seen set up.

I don't look at Medium blogs that often.

Some great blogs not hosted by Medium:

- http://samzdat.com - http://hotelconcierge.tumblr.com

An honest question - have you ever used any of these auto replies, ever?

I don't recall any situation, not even once, where they would be relevant to what I received. The most insane thing is that I receive auto-reply suggestions in English, even when the email was not written in English.

If this is the best that Google's "world-leading AI" can come up with, then I'm not worried about AI taking over the world. Not even slightly.

Most of the times they are correct but not sufficient, but I've used it enough times that I don't mind the suggestions being there 100% of the time. I do write tend to write terse emails though.

The newest "press tab to auto-complete word/sentence" while I'm typing though, that's incredible and I've used it in every email since I got the feature. A real time saver - and sometimes makes me sound more eloquent too!

I'm sure it's coming, but it would be nice if the smart replies used more of my phrases instead of Google's reccomendations, and mixed them up so it's no so obvious that you clicked on a smart reply. Also, clicking a smart reply doesn't automatically send the email, and still requires another click. It might be a good feature to "feel lucky" with when clicking.

I switch from GMail to a personal mail tired of their non-standard IMAP, their more and more heavy webUI more and more tied to Chrome. So...

BTW switching to personal mail not only improve my workflow but also push me to expand my knowledge like never before. Now my personal, notmuch based+scripts automation automate far, far, far more than Google and far better. If I'd like to have template replay I can add them in a snap.

> Users of Gmail — and there’s at least a 50 percent chance that’s you — have noticed an “upgrade”

I wonder: for all the people in the world using Gmail, how many use their web interface? I'm a "Gmail user" but I only access it via native mail clients. I haven't actually tried logging in to Gmail using a web browser in forever, so I never see their latest feature that everyone loves/hates.

I do wish the smart reply could pick out the intended recepient's name, or have the ability to pick it off of a drop down / additional three options. I can see people considering smart replies to be inauthentic, but all I need to say is "Thanks X!" or "That's great X", and get back to whatever else I was doing.

The four maxims in the article are really spot on.

Some people might find that too creepy or think Google is "reading" their emails.

The smart reply feature is already doing just that. If they aren't going to stop and we can't make them, then they could at least add the finishing touches to the feature.

I have used these on LinkedIn, especially to respond to requests from people I don't know or am not actually connected to -- as the writer notes, it isn't really appropriate to actual email from friends and family. Though I suspect all of these are meant for people replying from a phone or watch who don't want to bother typing.

"Thanks, I'll check it out"

You know what's worse? The damn "Smart Compose" dialog box that pops up and stops me from typing.

I cannot begin to fathom what deranged design flow led to Google creating a UI element that is objectively worse than Clippy.

I love these. They help me start an email otherwise I write ok or something. They're also getting better and more variety. They do need to get longer though. Give me a few sentences I can personalize.

LinkedIn also happens to have interesting answers. Which I don't use.

Or skype, I had to disable it, because it looked like complete gibberish (some English suggestions for non english messages).

I’m as google-pessimistic as the next guy, but I really struggle to see the problem in providing these quick reply options. Their phrase complete is great also.

Haha I love those things. Because I'm Australian one of them says "Got it, mate!"

A constant reminder that AI is thousands of years from taking over humanity.

On a slightly unrelated note, when did Medium start blocking content and prompting you to upgrade? I can't read this article.

When logged in, I was prompted to upgrade. Incognito let me read the article

Thanks for the link!

Well said.

I agree.

Thanks!

He shouldn't have blurred the text in [0], as blurred text could be read using techniques I admit knowing nothing about. It's better to just use a black block or something like that. Pixelating is bad too.

[0] https://cdn-images-1.medium.com/max/2000/1*h8cIVhi7n9-vuNJ-1...

Edit: I just noticed parts of the concealed text have been "smudged" before blurred. So no risk here. Still valid as a PSA :)

Personally I prefer replacing the text and then applying a light blur. Let them feel smart for a few seconds.

Yes, the first thing I noticed is the date quite clearly says 18 (why not blur the whole thing?)

There's a discussion on deblurring techniques before: https://news.ycombinator.com/item?id=4679801

This is ridiculous... I hate when robots auto fill in boring emails I have to send 100s of times a day!!!

Could you please review the site guidelines? This comment breaks two:

Please don't post shallow dismissals, especially of other people's work. A good critical comment teaches us something.

Please respond to the strongest plausible interpretation of what someone says, not a weaker one that's easier to criticize.

https://news.ycombinator.com/newsguidelines.html

very simply, it never makes sense...

I try not to respond to emails with one liners. If it deserves a response, it deserves a complete thought. I hate sending something to someone and getting a "thanks" reply back. I get that you appreciated what I sent you but if your reply isn't going to affect anything, don't respond. Shoot me an IM if you really need to give me a one liner.

I try not to respond to emails with one liners.

I have a co-worker who does this. She sends me data files with "Please review."

So I've started responding with things like, "It had a good beat, but I couldn't dance to it. Three stars." Or, "Excellent datas!!eleven! Fast shipping! Goog eBayer, would buy again!"

She's still sending me the one-liners, but at least I'm having fun.

Just run file against them and send the results back.

> Shoot me an IM if you really need to give me a one liner.

You say that, but...

14:51 help

14:51 I wanted to merge dev into master

14:52 and got message that dev is not up to date or something with the base branch

14:53 so I clicked on update branch, but than I saw in circleci that it attempts to merge master int develop!?

14:53 so I stopped it

So my issue with IM is you get pinged every 30 seconds while you watch people write their emails in real time.

I try to not do this either myself, but at the same time, I wish people do this a little more often.

Some people are just bad at replying things, and even an one liner would save me from wondering and/or needs for follow up of "did you receive my message?"

In many cases a one line reply is simply an acknowledgement of receipt.

>I get that you appreciated what I sent you but if your reply isn't going to affect anything, don't respond.

I often reply to non actionable emails/IMs with a thumbs up emoji indicating that I have read the message and have nothing to add. This means the other side know they don't have to try to contact me again because I have seen it.

You should try it:

https://www.buzzfeednews.com/article/katienotopoulos/i-tried...

Reply on news.ycombinator.com