I was the PM on this feature (and these are my views not Googles, as per normal). The truth is actually much more banal. Most people are just really positive over email.
The model is trained to offer reply suggestions that have the highest chance of being accepted, from a large whitelist of the most common short replies. The whitelist contains many negative options. We're optimizing for click through rate. That's it. There's no editorial judgement, definitely not "‘no’ struck someone as too negative and they had to take it out."
We actually experimented with intentionally inserting more negative options to increase diversity. Doing this reliably causes a hit to our metrics.
Discovering this made me pretty happy about the world. Most people are generally pretty friendly to one another (at least over email!).
> We're optimizing for click through rate. That's it. There's no editorial judgement...
I think Google in general needs to start taking seriously the idea that this is an editorial judgement.
There’s nothing necessarily wrong with optimising for click-through rate, but doing that will skew the options you offer to your viewers in specific ways. Those outcomes are still Google’s responsibility, even if they are the result of choices which are not under Google’s direct control, because Google has made the editorial choice to use click-through rate as their primary metric of choice.
(It’s relatively innocuous in this particular case, but when Google uses click-through to choose between content offered up by other people, Google is setting up feedback loops between consumers, content producers and the algorithm that mediates between the two which can trap the system as a whole in pernicious local attractors in the phase space of possible choices that may be impossible to escape.)
Precisely. Worth noting that the same phenomenon of "pernicious local attractors" is responsible for a whole slew of discriminatory behaviors. Blacks are more likely to go to prison! Therefore we'd better frisk this black guy extra thoroughly...
The truth may be banal, but the impact is not. That's the problem with technology. So often, there's no ill intent in the design decisions, but at scale, the effects can be harmful, even massively harmful.
> We're optimizing for click through rate. That's it. There's no editorial judgement...
Something that all of us as technologists need to learn is that this IS an editorial judgement. We do not get to disclaim responsibility just because we delegated that responsibility to an algorithm. It is we who delegated it, we who chose the algorithm and the metrics, and we who are responsible.
"We're optimizing for click through rate" is simply not good enough in 2019.
We could claim to be naive in 2000, perhaps even 2010. But today? We all know we're playing with fire now, and it doesn't much matter that of course we didn't MEAN to burn the house down. What matters is that we don't give lit matches to children and we know how to not set the house on fire.
I agree and good point on the broader meaning of editorial judgement. If we had seen any negative impact such as the FB/YouTube examples we would absolutely take that into account. I haven't seen anything remotely like yet with regard to smart reply but would want to be the first to see it if it exists.
How would you know? And what would your threshold be?
For example, suppose you did a controlled study and found that the new autocomplete UI was presenting users with only "yes" options, X% of the time. Is there a threshold beyond which you would unlaunch or adjust the feature?
What if the new UI was actually changing user behaviour, causing them to answer "yes" more often than they otherwise would have? Is there a delta (positive or negative) beyond which you would unlaunch or adjust the feature? If you had to adjust it, how would you know you were adjusting it correctly?
What sort of telemetry would you need to put in the product so that you really would know if this was happening?
I'm curious if your team thought about these questions before launching, and what ideas came up.
Your willingness to use "we're optimizing for click through rate" as a defense is quite frightening to me, though.
Someone who thinks "We're optimizing for click through rate" is morally neutral is not qualified to be making these decisions, just as someone who thinks giving lit matches to children is morally neutral is not qualified to be a fire marshal.
It's a pretty long journey from 'our algorithm chooses which auto-replies to offer you based on which have the highest probability of being selected' to giving lit matches to children.
I'm afraid to ask your opinion of SwiftKey's autocomplete.
That's not quite what I'm saying, though. My concern is not with stating the fact that the algorithm uses this metric. My concern is with presenting the decision to use this metric as morally neutral.
If you think that giving lit matches to children is morally neutral, you probably either don't understand how dangerous fire is or don't understand how unpredictable children can be. That's the point of the analogy.
I don't know what could be more morally neutral than 'we offer the suggestions that you've shown us you're most likely to want to use.' Anything else would be trying to put _their_ words in people's mouths.
And let's not lose sight of the fact that this is a tool that offers an automatic one to three-word email response for when you're too lazy to actually reply. The stakes are about as low as they come.
Pretty much the entire history of progressivism is a long, slow dawning realization that it's not harmless to treat individuals on the basis of group statistics, for one group category after another. Part of the problem is feedback loops - you can inadvertently amplify subtle perturbations by feeding them back to people. For example, it might seem harmless to subtly nudge a woman away from a career in STEM, on the basis that women are rare in STEM and therefore they probably won't enjoy it. But this action, repeated across millions of people, creates the pattern that you are reacting to. I think the principle probably generalizes to "everyone", not just "women" and "black people".
In the case of autocomplete suggestions, they can still cause harm even when they're statistically likely. What do you think Google would autosuggest for "Blacks are...", if it were around in 1950? What would the statistics on the completion of that sentence look like? And would the effect of someone seeing that list be 100% neutral - or would it subtly nudge them?
There's nothing obviously harmful with the specifics of GMail's auto-suggestions now. But the principle of 'we offer the suggestions that you've shown us you're most likely to want to use' is not morally neutral.
We're talking about auto suggestions along the lines of 'Yes, I have', 'no, I haven't', 'sounds good!', 'yum!'.
I'm not arguing against your principles necessarily, but deploying the argument where it's really not warranted is a form of crying wolf and turns people against it.
And as an aside, at my university the only time I ever saw any recruiting material for STEM, it was along the lines of 'Scholarships for women in STEM'.
But the point is, whether this particular case is a problem or not, the general attitude that led to its development is a problem, and we're using this opportunity to discuss that (even though the original article is not very coherent).
I think the general attitude of being willfully blind to fundamental distinctions between things that are only superficially similar and shoehorning issue debates in where they don't belong is ultimately unhelpful for the debate over what is, elsewhere, an important issue.
"As low as they come"? You might be forgetting that Gmail has about a billion monthly active users. Scale matters.
I'm sure you can think of many products that have fewer users, and features in those products that are less central to them than e-mail composition is to a mail client.
> Your willingness to use "we're optimizing for click through rate" as a defense is quite frightening to me, though.
I would suggest that if you are frightened by what is essentially an auto-completing on-screen keyboard being optimized for usefulness over making some random person on Hacker News feel it was more nuanced than just trying to be useful... there's probably no way to implement a feature like this, and make it useful, which wouldn't have you shaking in your boots.
Don't be frightened by what is essentially a riff on T9.
> If we had seen any negative impact such as the FB/YouTube examples we would absolutely take that into account.
Maybe? Maybe not? There’s visible and tangible negative impact from Youtube algorithms (owned, incidentally by the same conpany). And Youtube (or Google? Or Alphabet?) rarely if ever acknowledges and fixes those problems because their algorithms are also optimised for click-through.
I read GP as saying they watched the system's behavior and decided that going by click rates gave the best results. I'd say that is the editorial judgment, in the sense that you're talking about.
If they just set things up to go by click rates and assumed that whatever results occurred must necessarily be preferable, then that would be disclaiming responsibility in the way you describe, but GP's version doesn't sound like that.
It's not editorial it's meta-editorial (metatorial?). They're picking the method that picks the content, not the content itself.
(But this isn't blind -- they're picking the method while watching the output, so it may be a fancy way of picking the exact content you want plus other similar content plus adversarial noise).
I think this is exactly right, though it sounds pithy. The same process produces food that is unsubtle, fatty, and highly sugared. It sells really well! But in the long run, it's terrible for us, and also boring - all fast food tastes more or less the same, because the more adventurous it is, the worse it sells. The term frequently used in discourse is "lowest common denominator".
It's a question of capitalism optimizing for local optima versus global optima. What were seeing with the internet is, what happens when we apply this process not just to food, but culture itself?
It is good enough, 2019 doesn't mean we don't reflect user behavior.
Not every recommender is a moral agent. Auto-suggest and conspiracy videos aren't the same thing.
And technically, yes, there's not such thing as a non-editorialized recommender, but this feels like a pointless nit since there's a world of difference between trying to be a reflection vs intentionally not reflecting based on a human-made moral judgement.
What even is the moral editorialization in auto-suggest? It's obvious we don't want conspiracy videos pushed front and center of educational topics, but how on earth do you balance the number of "yes" vs "no"? That's clearly going to be a completely made up number. If you don't have a clear problem to fight, it's probably best to not editorialize just to say you did.
Pretty sure Gmail is a product not a life giving service, they still do tend to optimise for people continuing to use their service. The societal effects of X technology seems to be the new scare tactic buzzword. It's an automated email response used for emails that don't merit an actual response, the slippery slope fallacy doesn't apply here.
I'd apply that instead to their language autocomplete feature(the grayed out text that appears when you start typing and that you can just press tab to insert into your email) which is scary in the way it tries to shepherd your language
From my personal experience I think the metrics produced the correct outcome here. I'm not going to click one of those things unless it gets the tone right. So, if you present one 'yes', one 'no', and one 'maybe' (for example) then you only have one chance to nail the tone (because obviously the only candidate is the one that is giving the answer I want to give) and the odds are pretty low. If on the other hand 95% of all replies are a yes, giving three 'yes' options gives you three chances to right the right tone, 95% of the time.
Or you could take this as an opportunity to present first choice of content, then choice of tone. That's two clicks, but it would be a lot more useful if I could pick between yes/no/maybe, and once chosen it would present me with a list of phrases for that choice.
One thing I can definitely understand is that most negative responses have to be carefully crafted, usually polite, and highly dependent on the relationship you have with the other person. One-liner positives are often considered polite since you're just accepting their request; one-liner negatives are almost never polite.
While it's exceedingly commonplace to fire off a one-line e-mail such as "Yes I can" or "Sure that works" it's very uncommon to respond with just "No I can't" or "No that doesn't work". More likely you'd respond "That doesn't work but how about ..." or "That doesn't work because ..."
> Discovering this made me pretty happy about the world. Most people are generally pretty friendly to one another (at least over email!).
Another possibility is that modern wage laborers are forced to wear a mask of positivity and say "yes, absolutely!" or "awesome!" to everything, because they are terrified of losing their jobs and not being able to keep their necks above water, in a world where fewer and fewer people own most of everything.
Perhaps you yourself are wearing this mask of positivity, when you come here and say that this makes you "pretty happy about the world".
The road to hell is paved with good intentions (and blindly following simplistic metrics).
> wage laborers are forced to wear a mask of positivity
RSA Animate made a short animation about how a culture of forced positivity in white color jobs, and how it contributed to the late-2000s financial meltdown.
I can't even figure out what author thinks is wrong with the feature.
The post is mainly about services that push users into accepting things (like data sharing or microphone access), but the only example raised is your "suggested replies" feature, which has no apparent connection to what the post criticizes.
I'm sure you considered whether optimizing for click-through rate at the expense of diversity of response "moods" (or whatever) was the right tradeoff. Why did you ultimately decide it was?
We also take qualitative metrics pretty seriously and found that this experience was well received on a more subjective level, high csat in particular. It surpassed the success criteria we set out for the feature and I'm happy about where it landed. Of course it's not going to be ideal for everyone, so we have a setting to disable.
To be clearer on our metrics: In addition to qual there's also a precision/recall tradeoff, so saying ctr alone was an oversimplification. We have coverage targets, but that doesn't really impact the positivity of the suggestions.
I make absolutely no claim of accuracy to your particular situation here, but the unique speaking style you use (unique to a layman, but quite common I suspect among people who find success in SV) made me immediately recall a Noam Chomsky interview I watched quite some time ago:
The key line being: "I'm sure you believe everything you're saying, but what I'm saying is, if you believed something different you wouldn't be sitting where you're sitting."
As someone who tends to write overly verbose emails, I find this feature to be useful, in that it reminds me how often "Will do." is an entirely appropriate, succinct response.
Is your model trained globally, or does it learn individual users' styles?
I could imagine that globally the most common replies are three forms of "yes" but for any individual user the most common replies are two forms of "yes" and one "no", simply because there's less diversity in one user's replies than in the entire population.
There is important variation. Some of us looked around and found how to turn it off, immediately. We are the ones who get a different experience from the rest.
That this the only other choice available is also significant, particularly in reference to other cases where they have chosen to make turning something off impossible.
Taking away the extremely popular "classic layout" is an example. It wasn't costing them anything, but getting people used to accepting arbitrary changes forced on them has worked out very well for Apple, and Google has to want some of that. Each passive acquiescence makes the next easier, until it becomes automatic.
Facebook deemphasizing clickbait content reliably hurts their metrics. YouTube deemphasizing extremist videos reliably hurts their metrics. This metrics-driven product development was totally understandable (and I even tried to do it) up until a few years ago when we simply got too good at it and it ripped the world apart precisely along the seams where the human brain has bugs that cause these things to increase metrics.
Email prefill replies are a silly and relatively harmless example to make this case, but the point is every last button on every product we use is caked in this and we're realizing too late that metrics identifying what humans want often does not coincide with what humans want being good.
People may not say no as often, but one no can be very powerful. It's harder to say no. Billions of people nudged towards saying yes is not nothing.
This whole feature is fucking terrible. I hate it. Please provide a way to turn it off.
For that matter, everything about the Gmail redesign is terrible. It has gimmicky crap like this I don’t want, nothing new I do want, is slow as molasses, and is extremely buggy and unreliable. It is constantly making me wait (sometimes minutes on a slow connection) to do actions which used to be fast, constantly misinterpreting my inputs and weirdly scrolling my page around, it takes unreasonable amounts of CPU/memory, and I have lost text I was typing several times. I frequently end up wanting to punch my computer when using the current Gmail, something which never used to happen.
Sometimes I get fed up enough to use the plain html version. But that is not really satisfactory either.
Bringing back the Gmail of circa 10 years ago would be a huge improvement.
Edit: apparently the suggested responses feature can be disabled in settings. That’s something at least.
This model could exploit known or unknown bugs in the human psyche, optimizing for click through rate today, or something else tomorrow, while bypassing a conscious human decision, like "yes" and "no".
Imagine a light switch in your house that doesn't go ON an OFF but has a bunch of different options every day depending on what the model is trained to do or who the PM for this "feature" is.
It's ultimately manipulation, a loss of control, an unwinable fight against a learning machine that gets what it wants.
For me anyway, most of my gmail emails are sent to people I am very familiar with (my wife, friends etc.) So the language I use will often contain personal jokes or colloquialisms unique to us or our community. This feature would be so much more useful if it was trained on my own email. The click through rate would be much higher then, at least for me.
> [...] experimented with [...] negative options [...] causes a hit to our metrics.
I have very bad opinion on this feature overall, but now i am curious; how did you measure people wanting to say No and not finding a quick reply? And then how did you measure the quality of the No available on the test set?
and so was Hitler in ~1923. Optimizing something that affects how other people act (same happens on YT Veritasium:My Video Went Viral https://www.youtube.com/watch?v=fHsa9DqmId8) leads to bad outcomes (radicalization, race to the bottom, catering to lowest denominator etc).
The model is trained to offer reply suggestions that have the highest chance of being accepted, from a large whitelist of the most common short replies. The whitelist contains many negative options. We're optimizing for click through rate. That's it. There's no editorial judgement, definitely not "‘no’ struck someone as too negative and they had to take it out."
We actually experimented with intentionally inserting more negative options to increase diversity. Doing this reliably causes a hit to our metrics.
Discovering this made me pretty happy about the world. Most people are generally pretty friendly to one another (at least over email!).