In our op-ed for Tech Policy Press (“We Need to Talk About How We Talk About ‘AI’”), we made the case against the anthropomorphizing language that makes it harder to have clear discussions of what so-called “AI” technologies actually do, and when and whether to use them. But these ways of speaking are deeply ingrained at this point, and it takes work carve new conversational and writing habits. That work involves at least three steps:
- Noticing which word choices are anthropomorphizing
- Finding alternatives
- Getting in the habit of using the alternatives
Just normalize calling it “clanker” and be done with it.
I still prefer wireback
Now now, clankers are still humans.
I have to be honest, reading this article gave me a bit of a bad mouth taste, but let’s see if someone here in the comments sees it otherwise. We anthropomorphise so many things as part of language; not because we think they are human, but simply because it eases the use of language. When I say that a vending machine gave me an error code, I don’t literally mean that it handed something to me… Critisizing “AI” as a piece of terminology is not critisizing AI, it is critisizing natural language. Changing language around “probabilistic automation” (the proposed terminology to replace “AI”) will not change anything about the moral, ecologic and legal problems that people in this lemmy channel take issue with.
I think you’re missing the point. The issue with anthropomorphic terms is in that it reinforces an incorrect understanding of a capability, or rather, a lack thereof.
This is entirely intentional. That’s why it’s called “agent”, “prompt”, “skill”, “hallucinate”, “lie”, “harness”.
For people who understand how transformer based LLMs work, these terms are fine. But, most people don’t. Heck, most developers don’t.
An LLM cannot hallucinate for the same reason it cannot lie, or for that matter reason. Skills are just pre-appended prompts, and prompts are just a series of tokens that set up a parrot to give you the next likely word, and an agent is just giving that parrot a loaded gun, and harnesses is trying to vacate the room in case the parrot decides to shoot.
People who don’t understand the underlying mechanisms, are then misled by the existing connotations of these words. They hear “hallucinate”, and think that the rest of the time “it works without that kink”. It’s insane how it normalised shaking a magic 8 ball, and referring to one of the answers as “it hallucinated”. And as for lying, one can only do so with an understanding of truth and reality. LLMs don’t have this mechanism. Skills makes it sound like you’re imbuing a capability, while it is exactly just a template input for the prompt. I could go on.
The number of dangerous uses. And I mean this in the sense of people dying, or other similar negative consequences. All stem from a public misunderstanding that LLMs can produce correct answers to questions. It cannot. That it sometimes does, is by coincidence. People have died as a result. It would be nice if the language we use can help educate the limitations of their capabilities, rather than serve as a fucking sales pitch.
Feed a picture of a mushroom to an LLM, but don’t trust it to decide whether to eat it, but maybe use the suggested name to look up the mushroom in a book about it. Just make sure the book was written before 2022, because fuckwit editors have already used LLMs to generate images for books about mushrooms…
I absolutely agree. This will change nothing at best, and come across as sanctimonious BS that drives people away from what you’re trying to say at worst. Whatever problems exist with ai, they come from funding models, incentives, negative externalities (power, water), inequity (pay to play, unauthorized use of training data), and the way they allow the least moral among us to poison the commons with unprecedented speed and efficiency.
Very little of this is affected by what we call it.
We anthropomorphise so many things as part of language; not because we think they are human, but simply because it eases the use of language.
The point is attributing capabilities to algorithms which they are not capable of, like “reasoning” or “lying”. Such sloppy use of language affects negatively our ability to talk and think about reality.
Well, I see two categories of langague that are being taken under a single umbrella here:
- Natural every-day language, to which my point still stands. “Lying” definitely falls into this category. I can only repeat here the above, we simply talk like this. You cannot expect human language to be 100% formal precise all the time in this sense.
- More formal definitions. Yes, here I agree a bit that these should be as precise as possible. But when it comes to definitions, I also don’t think this is something that affects our ability to talk about it. When a paper or a dictionary defines what a term means, that’s the meaning we imply by using the term. Not the parts the term is made up from. There are plenty of latine names for animals and plants that we use, even though the individual parts of the name are not actually true about that animal and come from misdocumentation or a misconception. Nazis are not socialist, despite their name conatining the word. What I do agree with is that the imprecise naming might cause confusion when being new to the term, but once the talking and thinking starts, you would already be at a point where you know what the term means.
I tend to use:
- Assumed Intelligence
- My Cat
- Unsupervised Text Extruders
I don’t let software agents generate text that uses first person pronouns or that addresses the user directly. I think the largest problem AI presents is the psychological abuse of the user and the alienation from social and physical reality that it creates.
I’m thankful that I rarely talk about AI IRL except to say that it’s shit and I fucking hate it
These are really good points.
AI agent → probabilistic, unverified software manipulator
My absolute favorite piece of the post
Jesus Christ. This article wants me to memorize 200 concepts for an unknown benefit. The simple reason we anthropomorphize is because as social animals we learn the value of being nice when asking for things.
And no matter how advanced the AI, there is a human at the end that can decide to turn off the switch if he/she doesn’t like my chats.
I am not changing to accomodate software that it is supposed to be designed to understand whatever I throw at it.
And no matter how advanced the AI, there is a human at the end that can decide to turn off the switch if he/she doesn’t like my chats.
You should be more concerned with Roko’s Basilisk.
https://en.wikipedia.org/wiki/Roko's_basilisk
Eternity is infinitely long.
What can be more anthropomorphised than generative AI models, they are trained only on the outpourings of humanity. You are asking to describe orange juice without saying its quite orangey.
Anthromorphising LLM ‘behavior’ is like saying oranges juice themselves.
