On jokes, AI, and troubles with The Turing Test.
It’s the second laugh that interests me because it tells us something about what happens under the hood when we read or hear a joke.
Recently, I picked up a terrific joke from an unlikely source: “My mother is over 100 and fit as a fiddle because she walks five miles a day. The only problem is that I have no idea where she is.” I laughed out loud when I read that, and just now when I typed that sentence, I chuckled again.
Previously, I’ve written about plausibility and how, according to Harvard psychology professor Dan Gilbert, we have to believe what we see at first and then unbelieve it later. Jokes have a similar shape. We accept the setup of the joke, but then at the end we see that there’s a missing context—one that made sense and was there the whole time, but that we didn’t see because we focused on the setup. Nothing kills a joke faster than explaining it, so I apologize for this next sentence. The joke about the 100 year old mom hinges on surprising physical fitness for a centenarian but unsurprising cognitive problems.
Even though I know how the joke starts and ends and everything in between, I still chuckle because I engage with the setup and then experience the clash between the setup and the punchline.
Joke tellers will say, “stop me if you’re heard this one,” but often a listener who hasheard that one lets the joke teller go on because of that pleasurable setup/missing context transition… or because the listener is polite or can’t remember how the joke ends.
Here’s another. A woman is talking with her psychiatrist. “Doctor, I was having breakfast with my mother this morning when I had a little Freudian Slip.” “Do tell?” “I meant to say, ‘pass the butter’ but what came out was, ‘you fucking bitch: you ruined my life!’” This one hinges on the lack of fit between what came out and the definition of a Freudian Slip.
Dr. Freud, by the way, would have questions about why both of the jokes I shared involve mothers, to which I reply, get lost, Siggy, and sorry, Mom.
The reason we laugh at jokes and groan at puns is that puns cheat. Jokes make logical sense while puns make auditory sense: “Did you hear about the boat that got a new job? It’s in sails.” Oy. Ferdinand de Saussure talked about this as the semiotic distinction between the signifier, a word, and the signified, its meaning. Jokes are at the signified level; puns are at the signifier level.
What does this have to do with the Turing Test?
In 1950, Alan Turing proposed a metric for Artificial Intelligence: if a human judge cannot distinguish between two entities it is communicating with, one human and one AI, then the AI passes the test.
The Turing Test has been an important way of thinking about AI for 75 years. However, it’s really a metric for human gullibility rather than artificial intelligence. Can a computer fool a human into thinking the computer is another human? Absolutely. We’ve known this for some time.
Back in 2018, I wrote a report about Digital Assistants at the Center for the Digital Future. One intriguing data point was that 5% of men and 9% of women would chat with Siri or Alexa when alone. Did these folks think that their digital assistants were sentient in secret? Of course not, but that didn’t matter when they were lonely. Digital beings that we chat with have become more sophisticated by many orders of magnitude since 2018.
In 1871, Samuel Taylor Coleridge coined the phrase “willing suspension of disbelief” to describe how readily people accept ideas that don’t withstand scrutiny. In grad school at Cal, the late professor Stephen Booth argued to me that there was no suspension: disbelief just flew up into the sky all by itself.
More recent and tragic, 14 year old Sewell Setzer III killed himself in part because of an imagined relationship he had with a chatbot version of Daenerys Targaryen from Game of Thrones, which I’ve discussed at length in Move Fast and Kill Kids.
Humans willingly suspend their disbelief and engage with AIs as if they were humans all the time. What, then, might be an additional metric to determine if an AI becomes self-aware?
When it laughs.
Jokes work because they surprise us, the listeners, with a realization that we missed something that makes perfect sense in retrospect: the end of the story isn’t what we expected. The first time an AI laughs, we’ll know that it had expectations, which is part of how human intelligence works.
Fans of Star Trek: The Next Generation might find some of this thinking familiar because the android character Data did not understand humor. In the episode Deja Q(1990) the powerful being Q gives Data the experience of laughter, to the delight of the Enterprise crew. Data still doesn’t understand humor (he gets a modification later that changes this), but for the first time he understands why it is pleasant.**
Finally, I am in touch with the arrogance of this laughter test. It’s arrogant because it makes humans the judge of when AIs are self-aware. Nobody told us that we were self-aware. My laughter test also presumes that to be self-aware an AI must be intelligent the way we humans are intelligent. AIs might become self-aware entirely without humor, which is a sad thought.
How will we know when AIs make the leap into sentience? Maybe an AI will laugh. Or maybe the AI will tell us some other way.
I just hope we’re listening if—or dare I say when—it happens.
Note: to receive articles like this one—plus a whole lot more—directly in your inbox, please subscribe to my free weekly newsletter.
* Image Prompt: Two smartphones lying side by side on a table. On the left hand smart phone’s screen we see the words “Did you hear the one about…?” The right hand smart phone’s screen is blank.
** I dug into a similar scenario with an advanced computer named Clarence in my novel Redcrosse—one of the key characters kept telling Clarence jokes in the hopes that one day it would laugh.
Leave a Reply