Language and AIs

What current AI language use suggests about language and about AI

What AI says about language, and what language says about AI.

First let’s think about what the recent successes of LLM AIs say about language.

First of all, they contradict some parts of the poverty of the stimulus argument. Poverty of the stimulus was one of Chomsky’s arguments for why humans must have brains specifically adapted to learning language, rather than brains adapted for general cognition, which then learn language incidentally. The poverty of the stimulus argument goes: the language sample that children are exposed to is not sufficient to enable a general cognition machine to learn a full human language by copying. In particular, the sample of language that children are exposed to will contain very few gross language errors (things like getting your subject and object mixed up, or putting your thes after the noun); and so it would be hard for any general reasoning capability to infer anything about these kinds of errors. But that’s not what we find with people: people know that you can’t mix up random words in your sentence. Therefore, Chomsky reasoned, this suggests that people are bringing some preset patterns or algorithms to their language learning process, which enable them to infer not just positive rules about what’s OK, but also negative rules about what’s not OK.

Language learning LLMs are in an odd language learning situation, compared to people. They have access to a lot more raw data; but they access it in a much less dynamic way. In particular, their access to language errors will be even more limited than children’s, because they never get to hear someone correct themselves, nor do they get corrected on their language output. They bring to the process an algorithm that is language-specific, but also quite general in the way it operates. And the results, as of early 2024, are that these computers can speak, pretty much as well as us. So for them, the poverty of the stimulus wasn’t a constraining factor.

(Two caveats: (1) LLMs are dedicated language learning models, and that’s what Chomsky said we have, so this argument doesn’t really contradict Chomsky; (2) LLMs access much much more data, so the poverty of the stimulus argument might hold for humans, even if it doesn’t hold for computers.)

But here’s what I’m interested in, in early 2024. The advances in LLMs over the last couple of years have been very impressive. But the advances in language use, not so. LLMs got to the point where they could talk like real people with the release of GPT3 in 2020. At that point, sentences were 100% grammatical, and the semantics hung together as well as a human being’s do. Since then, my observation is that the language capabilities of AIs have not improved. (Caveat: it’s possible that this is a function of my/our own language perceptions.) That is, despite getting bigger and more powerful, their language comprehension and expression has not changed. That looks like a contradiction! There are two or three interesting things that this might mean.

(1) Language complexity has a natural level. Once you get “intelligent” enough to learn a language, there are certain features of languages that make them settle at a certain level of complexity.

(2) Language complexity has a natural upper limit. With a human language, you just can’t talk in a more complex way than we already talk. Computers have hit that upper limit along with us.

(3) There is an upper limit on what people can understand from language. AIs are now perfectly capable of talking with much greater levels of complexity, but they don’t, because they know it would blow our minds.

(4) AIs can’t learn higher levels of complexity in langauge because all their input is human language, at our default level of complexity. They can only learn from their language samples what we put in.

Let’s look at things the other way round: what does language say about AIs?

The fact that AIs have gotten themselves lodged at our level of language suggests that maybe they won’t be able to outdo us. They already have much broader knowledge than we do. But if they cannot develop higher levels of language from our huge language samples, then perhaps that means they cannot in general develop higher levels of reasoning than what goes into their input.

There are a couple of different possibilities here, too.

(1) This is a language-specific limitation. AIs may never be able to raise human languages to a level much higher than people; but that doesn’t mean that they can’t become smarter than people. They will be able to use our existing level of expression to express bigger, wilder thoughts. Or they will use existing human languages as a tool to build better languages that can express better thoughts.

(2) This is a general limitation. AI trained on human input can only ever get as smart as we are because that’s all the input gives it.

It’s worth thinking about what I mean by “higher levels of language”. I have a few ideas about what a language might look like that went beyond normal human usage. For example, it might be a much richer language, with a larger vocabulary, each word having more shades of meaning. It might be a language which combines words from other languages for additional support, similar to the way bilinguals might code switch at very high frequency. Or it might be a language in which the complexity of syntax gets much higher: think of those old-fashioned novels with long sentences, or made up examples about buffalo buffaloing buffalo. So far as I’ve seen, none of those things have emerged in AI as yet. And it’s not just natural languages: AI is used extensively in coding, and I have not seen any reports of an AI doing something clever with code that no human could have thought of.

In fact, the coding comparison is instructive, because code presumably doesn’t have the same natural features as natural language. If AI is using code “languages” at the same level as it uses natural language - i.e. at a human level - that suggests that it’s something about AI’s intelligence that is stopping it, not a feature of the language system.

Huh, that’s a point worth thinking about.