The 3 things an AI must demonstrate to be considered sentient

AI
photo: imustbedead, pexels.com

A Google developer recently decided that one of the company’s chatbots, a large language model (LLM) called LaMBDA, had become sentient. According to a report in the Washington Post, the developer identifies as a Christian and he believes that the machine has something akin to a soul — that it’s become sentient. As is always the case, the “is it alive?” nonsense has lit up the news cycle — it’s a juicy story whether you’re imagining what it might be like if the dev was right or dunking on them for being so silly.

We don’t want to dunk on anyone here at Neural, but it’s flat out dangerous to put these kinds of ideas in people’s heads. The more we, as a society, pretend that we’re “thiiiis close” to creating sentient machines, the easier it’ll be for bad actors, big tech, and snake oil startups to manipulate us with false claims about machine learning systems.

The burden of proof should be on the people making the claims. But what should that proof look like? If a chatbot says “I’m sentient,” who gets to decide if it really is or not?

I say it’s simple, we don’t have to trust any single person or group to define sentience for us. We can actually use some extremely basic critical thinking to sort it out for ourselves. We can define a sentient being as an entity that is aware of its own existence and is affected by that knowledge: something that has feelings. That means a sentient AI “agent” must be capable of demonstrating three things: agency, perspective, and motivation.

Agency

For humans to be considered sentient, sapient, and self-aware, we must possess agency. If you can imagine someone in a persistent vegetative state, you can visualize a human without agency. Human agency combines two specific factors which developers and AI enthusiasts should endeavor to understand: the ability to act and the ability to demonstrate causal reasoning.

Current AI systems lack agency. AI cannot act unless prompted and it cannot explain its actions because they’re the result of predefined algorithms being executed by an external force. The AI expert from Google who, evidently, has come to believe that LaMBDA has become sentient has almost certainly confused embodiment for agency.

Embodiment, in this context, refers to the ability for an agent to inhabit a subject other than itself. If I record my voice to a playback device, and then hide that device inside of a stuffed animal and press play, I’ve embodied the stuffy. I have not made it sentient. If we give the stuffy its own unique voice and we make the tape recorder even harder to find, it still isn’t sentient. We’ve just made the illusion better. No matter how confused an observer might become, the stuffed animal isn’t really acting on its own.

Getting LaMBDA to respond to a prompt demonstrates something that appears to be action, but AI systems are no more capable of deciding what text they will output than a Teddy Ruxpin toy is able to decide which cassette tapes to play. If you give LaMBDA a database made up of social media posts, Reddit, and Wikipedia, it’s going to output the kind of text one might find in those places. And if you train LaMBDA exclusively on My Little Pony wikis and scripts, it’s going to output the kind of text one might find in those places. AI systems can’t act with agency, all they can do is imitate it. Another way of putting this is: you get out what you put in, nothing more.

Perspective

This one’s a bit easier to understand. You can only ever view reality from your unique perspective. We can practice empathy, but you can’t truly know what it feels like to be me, and vice versa. That’s why perspective is necessary for agency; it’s part of how we define our “self.” LaMBDA, GPT-3, and every other AI in the world lack any sort of perspective. Because they have no agency, there is no single “it” that you can point to and say, for example: that’s where LaMBDA lives. If you put LaMBDA inside a robot, it would still be a chatbot. It has no perspective, no means by which to think “now I am a robot.” It cannot act as a robot for the exact same reason a scientific calculator can’t write poetry: it’s a narrow computer system that was programmed to do something specific.

If we want LaMBDA to function as a robot, we’d have to combine it with more narrow AI systems. Doing so would be just like taping two Teddy Ruxpins together. They wouldn’t combine to become one Mega Teddy Ruxpin whose twin cassette players merged into a single voice. You’d still just have two specific, distinct models running near each other. And, if you tape a trillion or so Teddy Ruxpins together and fill them each with a different cassette tape, then create an algorithm capable of searching through all the audio files in a relatively short period of time and associating the data contained in each file with a specific query to generate bespoke outputs… you will have created an analog version of GPT-3 or LaMBDA.

Whether we’re talking about toys or LLMs, when we imagine them being sentient we’re still talking about stitching together a bunch of mundane stuff and acting like the magic spark of provenance has brought it to life like the Blue Fairy turning wood, paint, and cloth into a real boy named Pinocchio.

The developer who got fooled so easily should have seen that chatbot’s assertion that it “enjoyed spending time with friends and family” as their first clue that the machine wasn’t sentient. The machine isn’t displaying its perspective, it’s just outputting nonsense for us to interpret.

Critical thinking should tell us as much: how can an AI have friends and family?

AI‘s aren’t computers. They don’t have networking cards, RAM, processors, or cooling fans. They’re not physical entities. They can’t just “decide” to check out what’s on the internet or search other nodes connected to the same cloud. They can’t look around and discover they’re all alone in a lab or on a hard drive somewhere. Do you think numbers have feelings? Does the number five have an opinion on the letter D? Would that change if we smashed trillions of numbers and letters together? AI doesn’t have agency. It can be reduced to numbers and symbols. It isn’t a robot or a computer anymore than a bus or airplane full of passengers is a person. 

Motivation

The final piece of the sentience puzzle is motivation. We have an innate sense of presence that allows us to predict causal outcomes incredibly well. This creates our worldview and allows us to associate our existence in relation to everything that appears external to the position of agency from which our perspective manifests. However, what’s interesting about humans is that our motivations can manipulate our perceptions. For this reason, we can explain our actions even when they aren’t rational. And we can actively and gleefully participate in being fooled.

Take, for example, the act of being entertained. Imagine sitting down to watch a movie on a new television that’s much bigger than your old one. At first, you might be a little distracted by the new tech. The differences between it and your old TV are likely to draw your eye. You might be blown away by the image clarity or taken aback by how much space the huge screen takes up in the room. But eventually you’re likely to stop perceiving the screen. Our brains are designed to fixate on the things we think are important. And, by the 10 or 15 minute mark of your film experience, you’ll probably just be focused on the movie itself. When we’re in front of the TV to be entertained it’s in our best interests to suspend our disbelief, even though we know the little people on the screen aren’t actually in our living room.

It’s the same with AI devs. They shouldn’t be judging the efficacy of an AI system based on how gullible they are to the way the product works. When the algorithms and databases start to fade away in a developer’s mind like the television screen a movie’s playing on, it’s time to take a break and reassess your core beliefs. It doesn’t matter how interesting the output is when you understand how it’s created. Another way of saying that: don’t get high off your own supply.

GPT-3 and LaMBDA are complex to create, but they operate on a single stupidly simple principle: labels are god. If we give LaMBDA a prompt such as “what do apples taste like?” it will search its database for that particular query and attempt to amalgamate everything it finds into something coherent — that’s where the “parameters” we’re always reading about come in, they’re essentially trillions of tuning knobs. But in reality the AI has no concept of what an apple or anything else actually is. It has no agency, perception, or motivation. An apple is just a label.

If we were to sneak into its database and replace all instances of “apple” with “dogshit,” the AI would output sentences such as “dogshit makes a great pie!” or “most people describe the taste of dogshit as being light, crispy, and sweet.” A rational person wouldn’t confuse this prestidigitation for sentience. Heck, you couldn’t even fool a dog with the same trick. If you put dogshit in a food bowl and told Fido it was supper time, the dog wouldn’t confuse it for kibble.

A sentient creature can navigate reality even if we change the labels. The first English speaker to ever meet a French speaker didn’t suddenly think it was okay to stick their arm in a French fire because they called it a “feu.”

Without agency, an AI cannot have perspective. And without perspective it can’t have motivation. And without all three of those things, it cannot be sentient. 

Source:
Tristan Greene, The Next Web

Tags

Back to IP in the news