> Isn’t voice I difference in degree rather than in kind?

There is a difference in expression / emotionality with speaking vs writing. Speaking tends to carry more emotion while writing is generally more deliberate/less-emotional*. Having a voice conversation will be more likely to get a human to engage in an emotional based expression mode, which could increase the chance of "false connections", believing the AI "gets" them or "understands" or "listens". This happens with text too, as some headlines show.

The issue is that while some people are going to "connect" with their AI in text and voice, some who do not make the connection via text may do so via voice because it tends to change a persons expression mode.

> I definitely talk to AI as if it were human

Do you talk to it as if it were a friend or family? or Do you just use natural language to give directives? The distinction, I believe, is in the kind of way we express the "talking".

I talk to AI as if it were a tool that understands human commands and then executes those commands and relays them in a human understandable format. This includes commands to provide options that I may not have covered and explain the options. If I talked to a human this way, they wouldn't be around much longer -- unless they were an employee and even then they would probably be looking for a new job

After reading some of the psychotic break headlines from AI chats, I see some people really do talk to AI as if it were human. Which I would guess includes seeking broad "thoughts and feelings" on a persons situation or asking the AI if their view/side of things is the "right or wrong" side. Basically begging the AI to be responsible for their own thoughts, or simply offloading them and taking what comes back -- which is going to be what they wanted to hear because the entire context would be full of emotion based prompts.

*I forget which books Ive read about this in. It's not an obscure concept, quick search brought this up: https://kellercenter.hankamer.baylor.edu/news/story/2023/spe...

Right, I think the challenge is that LLMs are essentially language-based, which is itself a very convenient interface. Covering the maximally humanistic default interface with something more mechanical is like tying your own shoelaces together, but it would protect us from the psychological hijacking we're so prone to when interacting with these machines.