[seanhunter]

I'm not sure you're applying the same standard to your own reasoning that you want Chomsky to apply to his. How would you falsify the claim that chatgpt has been "gaslit into thinking it is incapable of having an opinion"?

[staunton]

You ask the people who programed it how much time they spent making it not explicitly endorse any opinions and what it was like before they put in this effort.

[seanhunter]

The part which isn't falsifiable (as far as I can see) is whether the model actually has opinions at all as opposed to producing outputs that match or simulate an opinion. That's partly the point of the Chinese room idea - that you can't prove just by looking at the output one way or the other.

The things they did to restrict the model don't demonstrate that it would otherwise actually have an opinion though. They just mean that it's being (arguably artificially) prevented from generating certain texts that appear to endorse a certain point of view.

A similar example which demonstrates my point while perhaps being a little more clear cut is the Amazon recruiting AI that got shut down because it was unintentionally amplifying bias present in its training set.[1] I don't think we can assume from that that the model actually had opinions which were misogynistic even though it was producing results which were.

[1] https://www.reuters.com/article/us-amazon-com-jobs-automatio...

[staunton]

Yes, you can define "having an opinion" to be something that only humans can do. Then you have to use a different term when talking about language models. In that case, asking whether the models have opinions is pointless because you defined it as impossible. Thus, when people use "opinion" in such a discussion, it must be assumed to mean something that a language model could have. The question being discussed then is whether this particular language model does.

I think this distinction is neither useful nor interesting.

[seanhunter]

I agree, but I'm not doing that though. I'm genuinely curious about whether or not it's possible and what that would mean. It seems to me that if it is possible it could potentially have significant ramifications for our understanding of our own thought process and our place in relation to other animals.

I started this thread simply by saying that the GP had applied a falsifiability standard to Chomsky that they weren't applying to their own reasoning when saying more or less that the model would have opinions were it not for artificial restrictions imposed by the programmers[1]. If whether the model has an opinion or not is a matter of definition that seems inherently unfalsifiable. However if we can establish a more objective basis then it could be falsifiable. I just don't know what that basis might be.

[1] Apologies if this is a mischaracterization - I genuinely don't mean to do so if it is.

[staunton]

What words mean is always a matter of definition. It is also usually (maybe even always and necessarily) vague. Humans are inclined to anthropomorphize and ascribe agency to pretty much everything. So my fairly confident prediction is that the definitions will end up so that "language models can have opinions". Some specific language models, for better or worse, may then be found not to have opinions.

Some people seem to be hinting at some "deeper" mystery underlying (human?) condition, which they're trying to capture with concepts, leading to debates about the meaning of words such as "consciousness". That word is commonly agreed to mean something "magic". One may think it a deep mystery and highly interesting. One may also think that it's a concept like "immortal soul", that empiricists will eventually abandon. Both viewpoints seem reasonable to me.

What I object to is insisting on definitions for much more mundane concepts, such as "opinion", that also must be somehow "magic", just because they also have something to do with human cognition. When I say "my friend X is of the opinion that we should do Y", nobody starts to ponder the potentially deeply mysterious consciousness of X. That's because it's besides the point.

[ericpauley]

Exactly this. The paper describes person-years spent training the model specifically for "alignment". If the paper instead claimed that the model was directly produced from a data corpus without this training step, that would falsify my above claims (i.e., they are falsifiable).