4 ms·

This reminds me of the concept of semantic internalism vs. externalism, which most comments here seem to be misunderstanding. Most of the critiques of the view

This reminds me of the concept of semantic internalism vs. externalism, which most comments here seem to be misunderstanding. Most of the critiques of the view that AI art is still meaningful is based on either a hypothesis or empirical testimony of being moved by art without knowledge of the artists. Thus, because the artwork was causally responsible for engineering a mental state of aesthetic satisfaction, the artwork qualifies as being a piece of art. If that is the crux of the discussion then the conclusion is trivial. However, I think the AI art as pseudoart view is trying to make a statement on the external (i.e. ‘real world’) status of the artwork, regardless of whether viewers experience the (internal) mental state of aesthetic satisfaction. The line of thinking is that there is a difference between semantics (actual aboutness) and syntax (mere structure). The classic example is watching a colony of ants crawl in the sand, and noticing that their trails have created an image that resembles Winston Churchill. Have the ants actually drawn Winston Churchill? The intuition for externalists is no. A more illustrative example is a concussed non-Japanese person muttering syllables that are identical to an actual, grammatically correct and appropriate Japanese sentence. Has the person actually spoken Japanese? The intuition for externalists is that they have not. Not everyone is in agreement about this, although surveys have shown that most people agree with the externalist point of view, that meaningfulness does not just come from the head of the observer — the speaker creates meaning since meaning comes from aboutness (semantics). The most famous argument for semantic externalism was put forward by Hillary Putnam, I think, in the 60s. Roughly, on a hypothetical Twin Earth which was qualitatively identical to Earth, except which water was not composed of H2O but some other substance XYZ, an earthlings visit to Twin Earth and looking at a pool of what appears to be qualitatively identical to water on earth and stating “That’s water” is false, since the meaning of water (in our language) is H2O, not XYZ. To externalists, the meaning of water = H2O is a truth even before we’ve discovered that water = H2O. I think the argument for AI art being pseudoart follows a similar line of thinking. Even though the AI produces, say, qualitatively indistinguishable text from what would be composed by a great novelist, the artwork itself is still meaningless since meaning is “about” things. The AI, lacking embodiment, and actual contact with the objects in its writing, or involvement in the linguistic or cultural community that has named certain iconography, could never make (externally) truly meaningful statements, and thus “meaningful” art, even if (internally) one is moved by it. If one is to maintain the internalist position, that any entity that creates aesthetic mental states qualifies as art, then it seems trivial, since literally anyone can find anything aesthetic. Externalist intuition effectively raises the stakes for what we consider art, not necessarily as a privileged status available only to human creations, but by arguing that meaning, and perhaps beauty, does not only exist when we experience it.

mwigdahl 4y ago
Thanks for writing this -- it's very illuminating and made me think further about it (as someone who commented earlier taking the internalist position). I think there's going to be a lot of discussion of this as AI work proceeds, and the question of whether an AI can truly understand language in a sense that allows it to produce "aboutness" becomes more relevant. Could a human being, raised in a featureless box but taught English and communicated with using a text-based screen, produce text with semantic value? It seems pretty obvious that the answer is "yes". Will a synthetic mind developed and operated in similar conditions ever be able to produce text with semantic value referencing its own experiences? Probably not now, but at some point?
- wizzwizz4 4y ago
  > Will a synthetic mind developed and operated in similar conditions ever be able to produce text with semantic value referencing its own experiences? Probably not now, but at some point? Perhaps. But the GPT family of algorithms isn't a synthetic mind: it's a predictive text algorithm. It can interpolate, but it can't have original thought; it almost certainly doesn't experience anything; and if, somehow, it does? Its output wouldn't reflect that experience; it's trained as a predictive text algorithm, not a self-experience outputter.
- techno_tsar 4y ago
  Interestingly, I think a strong externalist would argue that a human being raised in a featureless box could not produce text with semantic value to the people outside the box. One upshot of semantic externalism are brain in vat type arguments, where statements such as “I am perceiving trees” (when they are simulated trees) is false, since the trees that the person is seeing is actually another concept, tree, while tree refers to real world trees. However, tree might be meaningful to the community of people also stuck in the simulation. So it might entail that AI art, in some sense, might be opaque to us but semantically meaningful to other AI raised on the same training data. That would require the AIs to be able to experience aesthetic states to begin with. More precisely, I think it would be akin to the person on the featureless box knowing all the thesaurus concepts to say, pain, but never actually experiencing pain itself. They might be trained to know that pain is associated with certain descriptions such as sharp, unpleasant, dull, heartbreak, and so on, and perhaps extremely complicated and seemingly original descriptions of pain. However, until the human actually qualitatively experiences pain, they only know the corpus of words associated with it. This would be syntactic but not quite semantic. It’s similar to the infamous Mary and the black and white room thought experiment, where even with a complete knowledge of physics, she still learns something new the first time she experiences the blueness of a sky, despite knowing all the propositions related to blue such as that it’s 425nm on the EM spectrum, or that it’s some pattern X of neutrons firing. That said, it’s not clear if this applies to statements other than subjective states. Qualitative descriptions of subjective states like pain, emotions, the general gestalt of the human condition might be empty of content, but perhaps certain scientific and mathematical ones pass the test, as they don’t need to be grounded in direct experience to be meaningful.

mwigdahl 4y ago

Thanks for writing this -- it's very illuminating and made me think further about it (as someone who commented earlier taking the internalist position). I think there's going to be a lot of discussion of this as AI work proceeds, and the question of whether an AI can truly understand language in a sense that allows it to produce "aboutness" becomes more relevant. Could a human being, raised in a featureless box but taught English and communicated with using a text-based screen, produce text with semantic value? It seems pretty obvious that the answer is "yes". Will a synthetic mind developed and operated in similar conditions ever be able to produce text with semantic value referencing its own experiences? Probably not now, but at some point?

wizzwizz4 4y ago
> Will a synthetic mind developed and operated in similar conditions ever be able to produce text with semantic value referencing its own experiences? Probably not now, but at some point? Perhaps. But the GPT family of algorithms isn't a synthetic mind: it's a predictive text algorithm. It can interpolate, but it can't have original thought; it almost certainly doesn't experience anything; and if, somehow, it does? Its output wouldn't reflect that experience; it's trained as a predictive text algorithm, not a self-experience outputter.
techno_tsar 4y ago
Interestingly, I think a strong externalist would argue that a human being raised in a featureless box could not produce text with semantic value to the people outside the box. One upshot of semantic externalism are brain in vat type arguments, where statements such as “I am perceiving trees” (when they are simulated trees) is false, since the trees that the person is seeing is actually another concept, tree, while tree refers to real world trees. However, tree might be meaningful to the community of people also stuck in the simulation. So it might entail that AI art, in some sense, might be opaque to us but semantically meaningful to other AI raised on the same training data. That would require the AIs to be able to experience aesthetic states to begin with. More precisely, I think it would be akin to the person on the featureless box knowing all the thesaurus concepts to say, pain, but never actually experiencing pain itself. They might be trained to know that pain is associated with certain descriptions such as sharp, unpleasant, dull, heartbreak, and so on, and perhaps extremely complicated and seemingly original descriptions of pain. However, until the human actually qualitatively experiences pain, they only know the corpus of words associated with it. This would be syntactic but not quite semantic. It’s similar to the infamous Mary and the black and white room thought experiment, where even with a complete knowledge of physics, she still learns something new the first time she experiences the blueness of a sky, despite knowing all the propositions related to blue such as that it’s 425nm on the EM spectrum, or that it’s some pattern X of neutrons firing. That said, it’s not clear if this applies to statements other than subjective states. Qualitative descriptions of subjective states like pain, emotions, the general gestalt of the human condition might be empty of content, but perhaps certain scientific and mathematical ones pass the test, as they don’t need to be grounded in direct experience to be meaningful.