This summer the OpenAI lab, backed by $1 billion in funding from Microsoft, Google, and Facebook, released an updated version of GPT-3, a text generator that produces convincing sentences by analyzing, among other online sources, Wikipedia, countless blog posts, and thousands of digital books.
According to a recent story by Cade Metz in the New York Times, one GPT-3 programmer decided to target pop psychologist Scott Barry Kaufman. Could GPT-3 really come up with a paragraph that sounded just like him?
Kaufman himself (pictured) was really impressed with this one, on the subject of becoming more creative:
I think creative expression is a natural byproduct of growing up in a diverse world. The more diverse the world is, the more you get exposed to different people, to different opportunities, to different places and to different challenges. And the more diverse that is, the more likely you’ll be to be able to put the dots together to form something new. And in many ways, I think if you want to be creative, you have to go for it. If you want to be a writer, you have to write, if you want to be a musician, you have to create music, if you want to be a comedian, you have to create comedy, and the more you create, the more likely it is that you’ll stumble onto some interesting stuff.Cade Metz, “Meet GPT-3. It Has Learned to Code (and Blog and Argue).” at New York Times (November 24, 2020)
But later, Kaufman was not so impressed. GPT-3 couldn’t discuss “imagination” with him although it did a fair job of punting the question. As Metz admits, “if you ask for 10 paragraphs in the style of Scott Barry Kaufman, it might give you five that are convincing — and five others that are not.” That’s because GPT-3 isn’t doing any thinking; it is iterating grammatically correct sentences from the masses of copy it has scarfed up and classified—by predicting the word that should come next in a sequence. Some programmers have found it useful for writing code, provided someone checks the code.
There’s a legitimate question as to how much unadvertised human help such systems usually get, when on display. For example, earlier this year, we were informed that GPT-3 had written an essay for The Guardian without human help. When Eric Holloway tested the assertion for cross entropy for Mind Matters News, he found
…GPT-3 can produce sentences that mimic standard English grammar and tone. The logical thought of the article, the meaning itself, is the product of the editors, who picked and rearranged the GPT-3 text into something that made sense.
We can actually detect some of these editorial decisions. Since GPT-3 just reproduces statistical relationships, all the parts of a GPT-3 text will tend to have the same statistical characteristics. Correspondingly, different GPT-3 texts will have different statistical characteristics. As a result, when an editor splices two GPT-3 texts into a single article, the different portions of the article will strongly diverge from each other.Eric Holloway, “Did GPT-3 really write that Guardian essay without human help?” at Mind Matters News
We also tested some of our own Mind Matters News copy on Grover, a GPT-3-type system that can supposedly detect AI-generated fake news. With short swatches, Grover detected no difference because, as we were told, the system is thrown off by proper names. News copy often features many proper names. With a large enough swatch—a whole autobabble article vs. an original news story—a difference was detected:
We tried the whole autobabble article from This Marketing Blog, “What Photo Filters” are Best for Instagram Marketing?. It scored 400.
Then we tried an equivalent swatch of the article published yesterday at MMN, “Why some scientists think science is an illusion” It scored 600.News, “New AI can create—and detect—fake news” at Mind Matters News
In general, GPT-3 is “mindblowing” if you don’t question it too closely, at which point, apparent discussions with the machine degenerate. Even the internet is not big enough to carry one side of an entire unique conversation.
One issue Metz addresses is that, because GPT-3 crawls the web unsupervised and unthinking, it vacuums up a lot of “sexist, racist and otherwise toxic language” and “learns from internet text that associates atheism with the words ‘cool’ and ‘correct’ and that pairs Islam with ‘terrorism.’” But in the absence of any thinking process in the machine, there is no clear way to prevent that.
He asks, “If GPT-3 generates the right text only half of the time, can it satisfy professionals?” Mimicking the human brain, Dario Emodi of OpenAI tells him, will require “entirely new ideas.”
So, whatever else GPT-3 is, it is not a “thinking machine.” That’s just the point. It iterates large amounts of plausible copy quickly. For some purposes, that would work. But thoughtful writing isn’t one of them. One reason the quoted passage by “Scott Barry Kaufman” worked as well as it did, for example, is that it is inspirational boilerplate. Far from any original contribution, it is just the sort of thing a pop psychologist might indeed say many times over. No doubt, we’ll be hearing plenty of it in years to come.
It’s worth remembering that George Orwell (1903–1950) thought that machines could write novels, provided they did not involve new ideas. He describes such a machine in 1984, “…she worked, as he had guessed, on the novel-writing machines in the Fiction Department. She enjoyed her work, which consisted chiefly in running and servicing a powerful but tricky electric motor… ”
Today, Julia’s machine would be a computer program like GPT-3. Tomorrow, perhaps, a quantum computer. Some things change. Others don’t though. In the absence of genuine thought processes, the output will be more iterations from the current state of the internet, for better or worse.
You may also enjoy:
Did GPT-3 really write that Guardian essay without human help? Fortunately, there’s a way we can tell when the editors did the program’s thinking for it.
GPT-3 is “mindblowing” if you don’t question it too closely. AI analysts sound unusually cautious in pointing out that it doesn’t live up to a lot of the hype.
New AI can create—and detect— fake news. But how good is it at either task? We tested some copy.