> So a bot that acts like a human (with hems and haws) is just wasting my time.

Duplex largely uses hems and haws like humans do - to signal that they're thinking about what to say next.

The system also sounds more natural thanks to the incorporation of speech disfluencies (e.g. “hmm”s and “uh”s). These are added when combining widely differing sound units in the concatenative TTS or adding synthetic waits, which allows the system to signal in a natural way that it is still processing. (This is what people often do when they are gathering their thoughts.) In user studies, we found that conversations using these disfluencies sound more familiar and natural.

https://ai.googleblog.com/2018/05/duplex-ai-system-for-natur...

Natural-sounding TTS improves the accuracy of speech recognition, because people tend to speak in an unnatural and over-enunciated style when they know they're talking to a bot.

Oh gee thanks. I have an accent, ring me up when they make a bot that understands me when I don't speak in an unnatural, over-enunciated style.

Voice recognition has been improving greatly, but it is still painfully close to the Scottish Elevator[1] at times in my experience.

[1] https://www.youtube.com/watch?v=5Kue42kO0hY

I have a friend who was participating often in psychology experiments at Stanford, and he became familiar with the whole procedure of letting the subjects believe that they were interacting with another person via a computer, when in fact they were interacting with a program (makes everything more standard and easier to analyse).

One day he participated in this "split or share" kind of experiment, and he was ruthless. Nobody's emotions would be damaged by acting nasty and never sharing with the computer program.

Turns out, it probably was actually a real person who was behind on the other side. He saw some old woman crying, coming out of some adjacent room after the experiment was over.

So, yeah, different social conventions definitely apply.

If I know anything about college psych experiments, the woman was an actor and the study was about your reaction to seeing her cry.

Actually setting up a software environment that connects two participants in the same simulation, and have it run without any bugs is usually way beyond the capabilities of the psychology students running the experiment. Not to mention the bother of coordinating the different teamed up participants (“Did you click start? The other participant is waiting for you to click sta— oh no it timed out. Hang on I'll restart the pairing sequence…”).

If you participate in one of those studies on campus, you are facing a computer running a local bit of standalone software, or a webservice running a questionnaire.

Surely you could solve this problem quite simply by having both participants and the experimenter join a group chat using COTS software+accounts (Skype, Hangouts, Whatsapp...) pre-installed on the machines?

You would normally want to limit their input to a few form fields so the data can be analysed, rather than letting them have a conversation. Look at the ultimatum game for example.

It's going to be quite the adventure for anyone who happens to sound like the automated voices.

Susan Bennett, the voice of Siri:

https://en.wikipedia.org/wiki/Susan_Bennett

There's the reverse where, what if AIs do develop feelings like that? Why shouldn't they be treated with the courtesy we treat other humans?

Just thought of these:

- Would this kind of personal assisstant be useful for people on the autism spectrum, to help navigate the kind of implied and unspoken narratives that people with autism seem to have difficulty with?

- Would there be a call to train Duplex to speak in a way that is more comfortable for people on the spectrum?

- Or any of the neurodivergent tribes for that matter?

I've said this before, but in reality strong AI will be another species. Every species of any moderate intelligence is expected to be treated with some courtesy and respect, but their social norms are far different from that of a human and we tailor our interactions with them in different ways. If I'm chewing gum and a human sees me, I might offer them a piece too. I'd never do that to a dog, no matter how much the dog wanted me to. If I saw a human chewing on the grass, I might stop them and ask them some questions to see if they're okay or need medical attention. If I saw a rabbit doing it, I might take a picture because it's cute, but I'd leave it alone (unless it was in my vegetable garden).

There is no reason to expect that strong AI will share the exact same feelings we do unless someone explicitly programs it to (which would be a mistake). Any truly emergent emotional behavior on the part of an AI would be very likely to differ substantially from ours. Making a chatbot "sad" is not the same as making a sapient (or even sentient) being sad. If AI ever achieves sentience, we're going to have to learn what makes it uniquely happy and sad.

Sure. There is also the reflection: how we think about and treat AI, or any other sentient species speak as muc about how we understand what it is to be human.

> Why shouldn't they be treated with the courtesy we treat other humans?

Because courtesy requires some amount of empathy and mental effort on top of what's required just to communicate your point. That effort is wasted on a machine.

Empathy is not necessarily just for the one receiving empathy.

There is a radical view that consciousness is the fundemental unit of reality, and as such, it would not be wasted.

Sure that's fine for some, but I personally don't see myself bothering to faux-empathize with a machine since I really don't have a mental-model for that kind of entity, so the empathy would be false in any case. Empathy for machines may in fact be cutting straight to the point and not worrying about emotional impacts since maybe machines value efficiency and conserved-cycles.

That mental-modelling would be cognitive empathy. Affective empathy is another, where the empathy is felt, rather than modelled out.

Then there is a particular kind of emotional attitude / posture / intent, where the something is treated as if it were sentient. It is neither cognitive or affective empathy, ao much as a shift in how things are experienced. It isn't something invoked through will. Hard to describe if you've never experienced altered consciousness states.

That particular shift is a fairly important experience for people (or at least for neurotypical), but is not obvious why to anyone who had never experience that shift. (I am speaking about neurotypical people, and some neurodovergent people, not necessarily someonr with a pauci-affective empathy neural architecture, such as people who are psychopaths).

Unless they’re being used to cover up computation delays on the back end, I feel like the hems and haws go beyond merely making the system more realistic and enters into outright fraud. By all means, sound natural and good, but don’t actually pretend to be a human if you’re not.

(And suddenly I have a vision of that last sentence being used against me in court in a few decades, with a jury consisting entirely of AIs....)

However you’d expect an Mmmm-hm as reaction to certain phrases as an ACK. What it does is just ACKing the phrase from other user. I wonder if other user would get worried and started asking “are you there?” if they didn’t receive the ACK

Edit: typos

Totally! My bad reaction is to things like interstitial “um” which is completely unnecessary in a computer voice.

That's an interesting observation.

I'm ok with the computers doing this, but it doesn't offend me that a computer is doing this. (Having said that, I take a dimmer view of dark UX patterns where I know I am being manipulated for some gain).

There seem to be a whole bunch of software folks afflicted with a fear of human interaction.

Hence this product which does not solve any compelling problem, except for those scared or stressed by human interaction. Not only is this deceptive and unethical, many will now be suspicious of calls with strangers.

For Pichai to dehumanize this poor girl and not even realise it confirms the worst fears of a disconnected technical community. What others may see as a sterile cold dystopia lacking human warmth many in software find attractive.

I think people are missing the point that one of the reasons why Duplex exists is that there are many businesses such as restaurants that do not offer some kind of online service for booking or reserving a service, and the only interface they offer is via phone. The main purpose of Duplex is to make things more convenient for people do not want to go through the hassle of interacting with an establishment via the phone, in which case, they can have Duplex handle making the call and being on hold on their behalf. If more businesses offered online services that are convenient for customers, then there would be less use for a service like Duplex.

I can't help but think of Gilfoyle's reaction (on Silicon Valley) when he sees a fridge's AI using vocal ticks

https://youtu.be/A48AJ_5nWsc?t=71

I think there’s an aspect to this that you did not bring up. Systems like this can be fragile. There’s no way for a human to bail in a socially acceptable manner if conversation ends up in a dead end.