Can an AI Chatbot Ever ‘Get’ a Joke?

“A prevalent malady of our age is to think of the mind as a computer” – Professor Sir Anthony Kenny.

The late science editor of the Daily Telegraph, Adrian Berry, used to get a laugh by telling the following apocryphal anecdote.

An AI is asked to write a short story with the popular themes of religion, aristocracy, mystery and sex:

“Good Lord!” screamed the Duchess, “I’m pregnant! Whodunnit!?”

First time I heard this I was working on my PhD on the metaphysics of consciousness and the viability of conscious machines. I wondered at the time if the hypothetical silicon wordsmith would understand why she’d been so inadvertently and very mildly amusing.

Can a machine, even in principle, ever get a joke? Not tell one, for clearly it can do that, but get one?

Jokes have not been taken seriously enough by the philosophers. Neither have the deeper essences of amusement, humour and laughter more generally, come to that. There is an entanglement of distinctive concepts to be tidied up there and it’s strange that we haven’t had more fun doing just that – conceptual housekeeping being very much our thing.

Which is not to say these things have not been discussed at all. Kant offers us a theory of the joke in his Critique of Judgement. Other thinkers, not all of them philosophers, have written about their moral and political significance. C.S. Lewis in the Screwtape Letters has his elder devil explaining how certain jokes encourage disorder in the psychology of humans and thereby generate entry points for the opportunistic younger demon to walk through at leisure.

Kant is interested in the phenomenology of jokes, Lewis their moral anthropology. What is the relevance of all this to the issue of whether Claude knows what it is up to when we think that he’s thinking? Perhaps this: taken together they show (a) that jokes don’t work without consciousness of self; and (b) that if machines can’t banter along with the rest of us then they are not integrable into the moral order and cannot therefore be persons.

Kant suggests that we laugh at a joke because it causes “an affection arising from the sudden transformation of a strained expectation into nothing”. To put it simply, we laugh because of distraction or because there is incongruity between set-up and punch line.

If I had the space, I’d give you an example of a Kant “joke” – all of which by their awfulness accidentally corroborate the theory he is proposing. We expect the joke to be funny; we get tumbleweed; we laugh at the thwarted expectation.

It’s enough to notice, for our purposes, that Kant is suggesting as a theory of the joke a set of necessary and sufficient conditions for laughter. He is, if you like, giving us an algorithm.

But the algorithm leaves out or does not capture the most important thing about jokes: what it feels like to find them funny. Algorithmic accounts of consciousness are reductionistic, and it is the way of scientific reductionism to ignore or eliminate the subjective character of human experience.

I’ll make the point in a more contentious way. The AI ascendancy is technologically and scientifically bewildering and yet metaphysically insignificant. The arguments against the possibility of machine consciousness are a priori: an algorithmic system, no matter how complex, is syntax, processing and speed. A simulated consciousness does not become conscious just because it simulates in more impressive ways.

And if machines are not conscious then, a fortiori, they cannot be self-conscious.

Claude’s cleverer descendants will no more get a joke than an abacus could a century ago, and Berry’s early AI lacks the perspective to realise that “she” is the object of whatever humour there is in his staple quip.

Which brings us to Screwtape.

Lewis uses a disputable demonology to make the point that there are jokes and there are other jokes and that there is an ethics (and perhaps even a theology) of humour. If laughter is the universal language of empathy – a truth familiar to anybody who attends AA meetings or Catholic wakes – then jokes are more culturally and socially contextual.

What they have in common, that said, is that they compel us to relate to each other not as centres of consciousness but as moral agents. The joke assumes capacities for empathy, sympathy and the suspension of belief. It reminds us that we are composed not just of intellect and appetite but also of will.

Here’s the question, then. Are the defenders of strong AI – the theory that machine intelligence or consciousness is philosophically plausible and practically inevitable – focusing on the wrong set of competencies when trying to convert the rest of us?

Instead of trying to persuade us that the systems are conscious, should they be aiming to persuade us that they understand the ‘in joke’, genuinely find it funny, and know when to repeat it and when not?

And if Claude learns in the end to turn the joke on himself – the most ingratiating thing a human being can do, after all – wouldn’t even the most incorrigible Platonist begrudgingly concede that this hunk of metal really is ‘one of the lads’?

A proposal, then. The Turing Test should be rewritten to consider that laughter and jokes aren’t add-ons to consciousness but rather constitutive of it.

Human beings are moral agents. Amusement and humour are conduits of that agency. For a machine to be conscious he or she must be in on the joke, which requires further that he or she is ‘in on’ himself or herself. Which I doubt can ever happen.

And I haven’t even mentioned the specific complexities of slapstick.