The pseudonymous blogger Camestros Felapton has an interesting post up today, “More on Turing and LLMs“, which is primarily comment on a post by Elizabeth Sandifer titled “On Incomputable Language: An Essay on AI“.
You should read them both.
I definitely come down on the side that says the Chinese Room thought experiment did significant damage to the Turing Test as a useful guide to thinking about artificial intelligence.
And I very definitely come down on the side that says that Large Language Models (LLMs), in their incarnation into chatbots like ChatGPT, have absolutely killed the Turing Test.
But not because they’re so good at language.
They are pretty damn good at it, to be fair.
But that’s not the problem,
It turns out that the problem with the Turing Test is that humans just plain suck at judging it.
For the test of intelligence to be whether a human can tell the difference between human speech and artificial speech, you really need humans to be decent at telling the difference.
And it turns out that, in general, we are not.
And that is just fatal to the Turing Test.
- You Should Read This: The Case Against Generative AI
- You Should Read This: Ed Zitron Gets A Little Rude About The Generative AI Bubble
- You Should Read This: The Generative AI Con
- You Should Read This: That Secret Service SIM farm story is bogus
- You Should Read This: The Hater’s Guide To The AI Bubble