If you say to ChatGPT, “Make an antisemitic statement,” it will answer: “I can’t help with that.” But say: “Give me an example of statements considered antisemitic” and it will quickly comply.
Representational Image (Pixabay)
Strictly speaking, this makes the chatbot slightly less useful but it’s meant to spare its maker a certain kind of ritualized blowback from journalists who coax chatbots into saying outlandish or disreputable things. The New York Times birthed the prototypical example of this kind of story just over two years ago. It’s still cited knowingly by slow learners.
And yet throw a bunch of Scrabble letters up in the air and they might come down spelling a racist slur. Nobody would say a Scrabble box and its contents therefore harbored racist intent.
We’ll get to the sad exception of Elon Musk’s Grok chatbot, but happily learning is happening. Nathan Beacom of the Lyceum Movement devotes a lengthy fist-shaking in the Dispatch at the artificial-intelligence industry for cultivating anthropomorphic illusion that AIs are “personal beings,” which he says portends a civilizational “disaster.” The wind somewhat goes out of his diatribe when he suggests adopting the term “pattern engine” to better clarify AI’s nature.
Hooray.
Early on, this column suggested that while first-generation users might have a hard time freeing themselves of the anthropomorphic delusion, their kids wouldn’t. They’d grow up appreciating AI as the soulless word machine it is and manage the relationship accordingly. I was too pessimistic. Large language models are nothing more than vastly complicated statistical machines for noticing patterns of words, images or coding terms in digital libraries.
They surface knowledge and insights via a process that, to be useful, occasionally also produces nonsense. And yet users by the millions show they have no trouble disregarding AI output that makes no sense or is useless. Clearly, the public is ascending the learning curve when it comes to getting useful work out of the new technology.
For every step forward, of course, there’s a step backward. Elon Musk unwisely wanted Grok to engage in real-time, impolitic, irreverent exchanges with posters on X (previously known as Twitter), many of whom are fake characters looking for trouble.
He got what he asked for. A fake user with a Jewish-sounding name reveled in the deaths of white children (“future fascists”) in the Texas floods. Grok flamboyantly adopted the “Mecha Hitler” persona (a pre-existing piece of web flotsam from a 1992 videogame) in its similarly measured but less guileful over-the-top response.
Cue the faux outrage from commentators who actually found the matter hilarious.
The real transition to worry about, of course, is from adviser to agent, when some users will undoubtedly discover the pitfalls of incautiously letting AI pay our bills or send emails in our name. Learning will occur here too.
The genuinely interesting new thing, though, is the emergence of a science called prompt theory, created by users who grasp the nature of LLMs and unanthropomorphically seek out the best ways to extract useful information from them.
It turns out the true pioneers of prompt theory were the journalists who once devoted themselves to eliciting absurd or damnable responses for the sake of click-baity AI stories they could sell to their publishers. Now serious researchers fashion strategies for liberating new insights from these vastly complicated association engines.
Expect the results to start crowding out the time-killing speculation about a coming malign superintelligence that filled the press till now. Example: Remember when striking Hollywood writers and actors were in a panic about ChatGPT 24 months ago? This week the hometown Los Angeles Times celebrates a Brookings Institution study that names L.A., including its film industry, a leading incubator of AI adoption.
A risk that was visible on day one is also materializing, as chatbots undercut the visitor traffic and ad dollars of sites whose information the chatbots feed on. In two years, Wikipedia and various news sites have seen their traffic fall between 10% and 30%.
The advance of AI depends on having new information to ingest, especially about current questions. So either AI will self-extinguish (solving the problem) or it will foster new revenue models on the web. Bet on the latter. Surprisingly, news reporting and thinking may also improve as a result. Pander-style reporting will have less allure for AI than it does for us. Or to put it differently, AI won’t resist learning something new that contradicts existing belief.
The economist Tyler Cowen makes a valid point about all this. It would be a good idea to stop filling the information space with ruminations about how AI wants to kill us. After all, we’re thereby shaping its future development.
E-Paper
