Ville V. Kokko

The discourse around just what “artificial intelligence” is seems to pull towards two opposite poles, both of which go too far. More nuanced discussion certainly exists, even behind these two kinds of rhetoric, but, as always, the simple versions stand out.

One extreme is to say that AI really intelligent. This leads to expectations that it can replace humans at work – without critical examination as to what it can actually do and what it can’t – and to people trusting it as a wise commentator, sometimes to the point of AI psychosis. It’s also fundamental to the discourse behind both possibilities and risks of artificial general intelligence and superintelligence. Even while admitting that there is no AGI yet, the expectation is that we can get to it by doing (scaling) more of the same, which implies that the narrow AI we have is somehow covertly general already.

The other extreme is captured by tropes such as that LLMs are nothing but next-token prediction machines or stochastic parrots (referring to the casual use of the phrase here, not the actual paper). It also involves talking about how the whole concept of artificial intelligence is misleading, since it’s not really intelligence. We might even hear about the “Chinese room” thought experiment.

The truth is that the tropes in both of these discourses are metaphors. In one, AIs are (even) minds; in others they are (just) machines.

Both of these discourses are incomplete. To usefully oversimplify: one tends towards saying that creating a machine intelligence is easy. The other tends towards saying that it’s impossible.

I would hardly say that we can know that a machine can’t have a mind. The idea implies something subtly anti-scientific and mystical. Is the mind a magical, transcendental thing like a soul? If it’s not, then it has to be able to be made of something else. This is what the scientific worldview tells us (putting aside how we haven’t solved the hard problem of consciousness): we have biologically evolved to have minds, and what makes our minds happen is our brains, which are ultimately made up of chemicals like everything else.

Could the same thing be realised in a machine? We don’t know, but any argument about how you can’t tends to make consciousness or life mysterious in a way that isn’t supported by science. Also, any analogical argument about function does the same. If LLMs can’t think because they just do X – even though they’re an incredibly sophisticated system doing X – then why can’t you say the same thing about the processes in our brains?

On the other hand, none of this means that creating machine intelligence is easy, or that just because we have created some kind of partial intelligence already, we should assume it’s the same thing as human intelligence. We don’t know what exactly would be required for that – that’s also why we stopped trying to program intelligence and moved onto machine learning instead – and the idea of just growing “intelligence” until it becomes humanlike or even godlike is a story that the current leading AI companies seems to have got from transhumanists imagining hypothetical science fiction futures. (I’m not being sarcastic – this seems to be historically the case.)

Thus, I do agree that “artificial intelligence” seems to be a misleading term, leading us to imagine it to be intelligent in ways that these things aren’t. It’s not only the term and the discourse around it that causes this. The fact that the current era of AI happened to grow mainly around chatbots also naturally leads to the illusion that you are talking to someone who’s like a human. This also creates and/or feeds into the illusion that you can assume this thing can have general intelligence; because it seems smart in some ways, it will feel like it won’t be dumb in other ways.

LLMs are dumb in other ways, though. Then again, they can also do some tasks better than humans, mainly ones that would take a lot of time. When using them for anything – or considering whether to use them at all – it is important to understand what they can do and what they can’t. Assuming they have “intelligence” only goes so far in that. The same goes for other “AI” systems, though they don’t have all the same challenges.

At the same time, discourse and metaphors that flatly deny AIs “thinking”, “intelligence” and so on aren’t telling the whole truth either. What would be realistic and appropriate would be to examine just what parts of what we think of as intelligence these tools do have, and what part not.

Now, it’s not that metaphors can’t be useful – or that they have to capture the whole reality to be so. In the case where AI is being humanised too much, talk about how it is just this and that instead can land just right as a balancing antidote to that idea. Hence, in the current climate, it seems that the metaphors of AI being “just something” are not that harmful, whereas the metaphors of AI as intelligence are.

Nevertheless, we need to keep in mind that there is truth to both. You never know when you might need to take a fuller perspective.