Auto-generated description: Four people on a circular bicycle-like apparatus engage in conversation, with pixelated speech bubbles above them.

The concept of multiplayer AI chat is interesting. The problem, though, as Matt Webb states is succinctly boils down to:

If you’re in a chatroom with >1 AI chatbots and you ask a question, who should reply?

And then, if you respond with a quick follow-up, how does the “system” recognise the conversational rule and have the same bot reply, without another interrupting?

So what are we to do?

You can’t leave this to the AI to decide (I’ve tried, it doesn’t work).

To have satisfying, natural chats with multiple bots and human users, we need heuristics for conversational turn-taking.

It’s worth reading the post in full, but to summarise and pull out the relevant quotations, in his work with glif, Matt found three approaches that don’t work: (i) context-based decisions by an LLM as to whether to reply, (ii) a centralised ‘decider’ on who should reply next, and (iii) attempting to copy conversational turn allocation rules from the real world.

Fortunately chatrooms are simpler than IRL.

They’re less fluid, for a start. You send a message into a chat and you’re done; there’s no interjecting or both starting to talk at the same time and then one person backing off with a wave of the hand. There is no possibility for non-verbal cues.

Ultimately, Matt found that a series of nested rules worked quite well:

  1. Who is being addressed?
  2. Is this a follow-up question?
  3. Would I be interrupting?
  4. Self-selection

My premise for a long time is that single-human/single-AI should already be thought of as a “multiplayer” situation: an AI app is not a single player situation with a user commanding a web app, but instead two actors sharing an environment.

Although I haven’t cited it here, Matt’s post is infused with academic articles and references to communications theory. It’s a good reminder that “natural” interfaces don’t happen by accident. Human-computer interface design needs to be intentional, not accidental, and the best examples of are a joy to behold.

For example, I’m reminded when stepping into other people’s cars just how amazing the minimalistic approach to the Polestar 2 is. You literally just get in and drive. That’s how everything should be in life: well-designed, human-centred, and respectful of the environment.

Source: Interconnected

Image: Nadia Piet & Archival Images of AI + AIxDESIGN / Infinite Scroll / Licenced by CC-BY 4.0