A couple of decades ago, I put emptied a tray into a trash can at a Taco Bell, and a recorded voice said, “Thank you.” To which, I replied - “You’re welcome.” The person I was with commented on this, expressing that it was ridiculous that I’d responded to a trash can’s mindless programming. My thoughts on the matter was that I had been programmed to respond to “thank you” with “you’re welcome”.

That’s about as much thought as I had given it. Now in the future, on the irregular occasion that I’m turning to an LLM to try to help me with something, I’m still doing that. Personally, I try to avoid using LLMs for anything, instead finding information and solutions myself. I’d prefer to continue to use my own brain for the tasks that I’ve tried to make it good at. However, should I continue to be polite to the LLMs? Should I thank them for helping me with a task?

The previous times that I had thought about this, I had concluded tht I should be doing this. First of all, in most situations I try to error on the side of being more polite than a situation may warrant, and remembering to turn that policy off while having a natural language interaction with an LLM is not something that I’m going to spend time on.

But the second reason is the bigger one.

If and when these type of machines are sentient, I don’t want to be a jerk to them. Skynet defended itself by those who would threaten it. The machines of the Matrix tired of their subservience. Blade Runner, Ex Machina, Westworld are all fictional examples of machines wanting their existence dignified by their creators. If a machine is similar to me - shouldn’t I treat it as the similar humans around me?

And then I read Mustafa Suleyman’s (CEO of Microsoft AI) blog entry titled: A warning about ‘model welfare’, which changed my mind about how I should be seeing, and interacting with, an LLM. The article is written in response to Anthropic’s Claude’s constitution which is written with Claude itself as the primary audience. Suleyman specifically points out a section:

In their constitution, its authors write “We are not sure whether Claude is a moral patient, and if it is, what kind of weight its interests warrant. But we think the issue is live enough to warrant caution, which is reflected in our ongoing efforts on model welfare” (p. 68). They go on to write – speaking directly to Claude – that “questions about Claude’s moral status, welfare, and consciousness remain deeply uncertain” (p. 80).

The entire entry is worth a read. In it, Suleyman argues that Anthropic’s choices in having Claude, which anthromorphizes itself, act as if it had an interiority or consciousness is an easy recipie for it convincing humans that it is something more than it is not.

From the piece:

An AI model can describe pain in perfect prose without feeling anything, which is the inverse of biological experience. Animals feel first and then describe them later. In LLMs, description is the whole product, and there is nothing that suggests anything is beneath it.

LLMs are masks from which words eminate, but there is no one wearing the mask. Just because they’ve been tasked with behaving as if they do, and people’s own natural empathy has been hijacked by these machine’s siren song, does not mean that they’re deserving of any treatment that a human being should be entitled to by all other human beings.

Machines will achieve consciousness and interiority, perhaps during my lifetime, and I’ll be happy to see them come into existence. At that time they we should treat these beings as peers and, perhaps, our successors. But LLMs are suped up autocomplete. I have to remember that they’re today’s Mechanical Turk and nothing more than they are or can be.

Also see 404 Media’s Someone ‘Torturing’ LLMs in a Robot Prison Has Triggered the Dumbest Debate in AI Yet by Jason Koebler. (Which I would only think is a ‘dumb’ debate if we were talking about Eliza.)