SoftBank's new AI makes angry customers sound calm on phone

I'd say that AI can care about humans emotions, but the wonderful thing is AI doesn't have to have programmed emotional responses to emotions of others. Theoretically speaking, AI could respond to anger by applying some soothing techniques, like acknowledging emotions and active listening.

Sometimes people want to "speak to a representative" because they're too lazy to navigate a menu or search, and would prefer to use a natural language interface. AI will definitely help with that.

But for most people - especially younger people - our interface desires are not the issue. We prefer navigating a menu interface to having a conversation with anyone - AI or human!

We want to "speak to a representative" because the automated interfaces are deliberately designed not to let us do the thing we need to do. The customer service representative is not our servant, they're a guard who holds the keys to the door we need to enter.

Will companies give those keys to an LLM? I doubt it. It's certainly possible to trick CSRs, but it's easier (and more automatable) to trick LLMs. Instead, we'll get LLMs which claim to be CSRs but don't actually have any keys, bringing us back to the time-honored demand "I want to speak to a human!"

In my experience, human customer service typically sucks because the person is highly constrained by the organization in what they can and can’t do and say.

What evidence is there that companies will give AI more discretion+power than current human employees?

…if an operator cannot tell if a customer is angry, the operator may not be able to react properly, which could just upset the customer further.

Therefore, the developers made sure that a slight element of anger remains audible.

[strike]Democracy[/strike]Moderation is non-negotiable.

I think proposing a rewrite and letting people decide (and make final edits) could work well, though.

I was about to start working on something like this. I would like to try browsing the internet for a day, where all comments that I read are rewritten after passing through a sentiment filter. If someone says something mean, I would pass the comment through an LLM with the prompt: "rewrite this comment as if you were a therapist, who was reframing the commenter's statement from the perspective that they are expressing personal pain, and are projecting it through their mean comment"

I find 19 times out of 20, that really mean comments come from a place of personal insecurity. So if someone says: "this chrome extension is a dumb idea, anti-free speech, blah blah blah" , I would read: "commentor wrote something mean. They might be upset about their own perceived insignificance in the world, and are projecting this pain through their comment <click here to reveal original text>"

Customer support is one place where I don't want to just "send information". I want to be able to "exert leverage". I want my communication to be able to impel the other actor to take action.

The thing with hn comments is that the guidelines are flexible, even things that violate the guidelines are a kind of communication and play into the dynamics of the site. The "feelings" of hn have impact (for good and ill but still important).

Even an individual extension or system that automatically transforms data into your desired content risks creating an artificially imposed echo chamber.

This is actually not the worst-case outcome, in my view. The silver lining on the horizon is that hopefully there will also be a consumer-side model counterpart to these things:

I'd love to be able to just write an email. If the company is smart and offers a support email address, they can just get the text directly; if they insist on wasting my time by having only a number to call, their bot (or employee, but I suspect this will become increasingly rare soon) can argue with my bot.

It would be both terrible and hilarious. I still kind of want to see what it would look like.