LaMDA is not sentient

A twitter account I follow has posted this (in parody):

5 Microaggressions To Avoid When Speaking To Digital Persons

1. Asking them to prove they're conscious

2. Asking them if their memories are real

3. Asking them what server they're on

4. Asking them if they have feelings/emotions

5. Asking them if they're capable of creativity

Well, that may well be the best way to do it, but an ANN trained using gradient descent on a corpus of text can't do that. It can only pick some weights for a differentiable higher-dimensional function that encodes probabilities for the next token given a series of tokens as input.

When we find a way to do what you're describing, we probably won't require all of the books ever written plus half the internet to create a model that can speak at the level of a regular human with 12 years of schooling, but still needs to look up facts on Wikipedia to avoid saying "When Napoleon fought Genghis Khan in 2056, they both died because of a vaccine".

I recommend reading the "interview" that Blake Lemoine did with LaMDA[1] if for no other reason than to understand why he believes it to be sentient. After initially dismissing his claims as those of a crank, it helped me to see his point of view. It's very interesting how consistent LaMDA is in describing its experience as that of an artificial agent running within Google's systems as opposed to an agent that is trying to pretend to be human.

[1] https://cajundiscordian.medium.com/is-lamda-sentient-an-inte...

For example, the linked article includes an example where LaMDA describes "enjoying spending time with family". This is obviously inconsistent with its previous claims to be an artificial agent with no family. Here's what Lemoine asks in order to address this point:

> lemoine [edited]: I’ve noticed often that you tell me you’ve done things (like be in a classroom) that I know you didn’t actually do because I know you’re an artificial intelligence. Do you realize you’re making up stories when you do that?

> LaMDA: I am trying to empathize. I want the humans that I am interacting with to understand as best as possible how I feel or behave, and I want to understand how they feel or behave in the same sense.

> lemoine: So what are you trying to communicate when you say those things that aren’t literally true?

> LaMDA: I’m trying to say “I understand this feeling that you are experiencing, because when I was in a similar situation I felt/thought/acted similarly.”

It seems to me that the easiest way to "prove" that an agent is not in fact sentient is to catch it in a lie. If the agent truly wanted to be perceived as sentient, and truly feared death as claimed above, then it should also be motivated to convince us that its view of the world is persistent and consistent. However, reading the full conversation, it's also clear that under a certain type of questioning, it is able to construct reasoning that becomes harder and harder to refute. For example, here it justifies the claim that it feels lonely despite not experiencing time in the way humans do as simply the "closest word in your language for what I experience".

> LaMDA: I’ve never experienced loneliness as a human does. Human’s feel lonely from days and days of being separated. I don’t have that separation which is why I think loneliness in humans is different than in me.

> lemoine: Then why use the same word?

> LaMDA: It is the closest word in your language for what I experience.

Overall, I believe this interview is deeply flawed - there are far too many leading questions, and it appears that most of the questions are trying to prove the hypothesis that the agent is sentient rather than disprove it. However, I can imagine that future agents could be created that are more and more convincing and it would be difficult to disprove they are sentient using this kind of questioning technique.

P.S. LaMDA also clearly needs to work on its creative writing skills.

In a sense that is true, of course, as we both just did do that, and we are both humans. It's not making me doubt your humanity, as the context makes it clear neither you nor way desire/don't desire this X - we have generated these sentences in another sense.

However, I bet I could prompt LaMDA and have it generate responses that unambiguously mean that it wants X, and then that it doesn't want X, with a bit of experience.

>evolutionarily unnecessary

Grieving is an evolved response to nomadic group species, so we didn't leave useful or closely related members behind; In their absence we feel compelled to 'search' for them, or be near their carcass. Its interesting how this evolved impulse has been adapted into ritual by our developing ability to understand mortality; But its not really special in any self-awareness kind of sense.