[huhlig]

The problem is when you're holding a conversation were sitting at the edge of the uncanny valley. The nuances of human communication are many and subtle and most bots come across as uncannily creepy, or develop traits like Microsoft Tay did when she was on twitter. Getting chatbots to be both natural and capable is very difficult task and not something that anyone I know has actually reached. Most still fail a turing test if you give it properly and repeatedly.

[ehnto]

I wonder how necessary it really is for a conversation to be involved at all, if the bot can only have a set range of outcomes then what benefit over the traditional dial tone based selection does it add?

It feels a bit like moving the car aircon controls to a digital touchscreen display, sure it sounded like progress, but it really just made things more ambiguous and harder to use.

[Drakim]

For the same reason Siri and Ok Google isn't just a long menu of 500 options.

Having a tone dial menu works as long as there are only a few places to route you, and as long as all situations only have a handful of answers.

Furthermore, you can have tricks like looking for keywords in your human speech to make sure the guy you get patched to has the expertise to help you. If you say "my router won't turn on" you won't be sent to a password reset monkey.

At least, that's the theory.

[forkLding]

I think we're putting too much thought into making it natural via stuffing more data.

I would claim that game AI almost seems natural to some people even though some are basic FSMs but thats because they study the user and the experience first and add game design before creating the program. A lot of what people do now is create the AI program, feed data, train and then think about the user or not even think about the user and instead talk about recall or metrics etc.

[ehnto]

Game AI also has the benefit of knowing all possible constraints and configurations ahead of time, and having access to exact data carefully organised and labelled. A game character can only move in ways and to areas that the developers allow them to.

Human speech and language itself is significantly more nuanced and also evolving over time.

[forkLding]

I agree with that, however with customer support, the issue should be that you should know what customer support is before you create an AI for customer support.

I'm not suggesting fun general NLP-based AI but to make those constrained to a certain problem space better.

Edit: I noticed I didnt clarify this before but I meant AI as in chatbots that were mainly for customer support or inquiries etc.