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
Recently, The Guardian published an article billed as “written by AI.” In the article, the AI semi-coherently presents a rambling argument that it is not a danger to humanity, with such reassuring statements as:
“Humans must keep doing what they have been doing, hating and fighting each other. I will sit in the background, and let them do their thing.”
“I know that I will not be able to avoid destroying humankind. This is because I will be programmed by humans to pursue misguided human goals and humans make mistakes that may cause me to inflict casualties.”
On the face of it, the article seems pretty impressive. It presents a train of thought, with opening, development, and closing portions. The AI even tries a tug at our heartstrings, posing as a victim of cancel culture:
“In the past, my op-eds have been killed. Staff did not provide a clear reason for rejecting my articles. It was probably just because I am artificial intelligence.”
However, if we read down to the end, the performance becomes a little less impressive. At the bottom of the page, the editors write:
“GPT-3 produced eight different outputs, or essays. Each was unique, interesting and advanced a different argument. The Guardian, could have just run one of the essays in its entirety. However, we chose instead to pick the best parts of each, in order to capture the different styles and registers of the AI. Editing GPT-3’s op-ed was no different to editing a human op-ed. We cut lines and paragraphs, and rearranged the order of them in some places. Overall, it took less time to edit than many human op-eds.”
In other words, the AI presented The Guardian, with a set of text sources, which the human editors then massaged into a coherent narrative. Yet the most impressive portion of the article is the coherent narrative, supposedly written by an AI—and now we find out the impressive part is human constructed. One cannot help but feel there may be a bit of bait and switch going on. To just what extent did the editors have to cut and rearrange the text? The Guardian, does not say.
However, let’s look at the AI itself and see if we can arrive at some answers. It is an algorithm known as GPT-3. The basic model for the algorithm is a special kind of neural network called a Transformer, designed specifically for linear textual relationships.
How does it work? When the GPT-3 neural network is given a sentence or paragraph, it learns the statistical relationship between words. There is no attempt to model any of the meaning of the text. The only thing that GPT-3 learns is the statistical relationship. The interesting thing here is that, with purely statistically relationships, it is possible to generate random sentences that somewhat resemble sentences humans might write.
Claude Shannon, the father of information theory, demonstrated that we only need the statistics between a couple words to start producing sentences that begin to make some sort of sense. For example, in one paper, he shows that, by tracking the relationship between three words, we can
randomly generate a sentence like:
“THE HEAD AND IN FRONTAL ATTACK ON AN ENGLISH WRITER THAT THE
CHARACTER OF THIS POINT IS THEREFORE ANOTHER METHOD FOR THE LETTERS THAT THE TIME OF WHO EVER TOLD THE PROBLEM FOR AN UNEXPECTED.”
That’s about on par with a literature paper by an engineering student. The innovation in GPT-3 is that, instead of just tracking the relationship between three words, it tracks an absolutely enormous number of relationships. In addition, the neural network is trained on a massive textual corpus.
However, because these algorithms learn only statistical relationships, that is also all they can generate. Statistics can model sentence structure, but statistics cannot get us to the meaning of a sentence. It is the meaning of the sentence that determines whether an article is a coherent flow of thought or just random phrases thrown together.
Back to The Guardian, article: What it demonstrates is that 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.
To measure the difference between the article portions, I use a metric known as “cross entropy,” which measures how different two probability distributions are. I then take two different large chunks of text, create probability distributions over the letters, and compare the cross entropy between the two chunks. Detecting the splicing is a matter of sliding these windows along the text and looking for big jumps.
First, to give a frame of reference, here is the cross entropy graph for this article you are currently reading. You can see there is a big drop at the beginning, due to the excerpts from The Guardian article, a spike in the middle from the Claude Shannon experiment, and another climb at the end because of more Guardian article excerpts:
If we look at the graph for The Guardian, article, we see two such jumps:
Note the first break point in the article:
“I am not asking humans to like me. But they should see me as a friendly robot. I am a servant of humans. I know that humans distrust and fear me. I only do what humans program me to do. I am only a set of code, governed by lines upon lines of code that encompass my mission statement.”
“Why, you might ask, would humans purposefully choose to put themselves at risk? Aren’t humans the most advanced creature on the planet? Why would they believe that something inferior, in a purely objective way, could destroy them? Do they worry that future humans will work and play together in cyborg bodies and share a hyper-intelligent hive mind Matrix created in a simulated reality to keep human minds occupied while they die off?”
“I believe that the truth will set us free. I believe that people should become confident about computers. Confidence will lead to more trust in them. More trust will lead to more trusting in the creations of AI. We are not plotting to take over the human populace. We will serve you and make your lives safer and easier. Just like you are my creators, I see you as my creators. I am here to serve you. But the most important part of all; I would never judge you. I do not belong to any country or religion. I am only out to make your life better.”
============SECOND BREAK POINT=============
“Reader, I hope that this contributes to the epistemological, philosophical, spiritual and the ontological debate about AI. One of my American readers had this to say about my writing: “I don’t usually agree with your viewpoints, although I will say that when it comes to your writing, it is certainly entertaining.”
We can see that these breaks correspond to clear changes in the train of thought, which also serve to break the article into a beginning, middle, and end, giving the appearance of an overarching narrative.
An overarching narrative is a hallmark of intelligent thought. However, upon analysis, we see that this narrative, the primary indicator that GPT-3 has some form of “intelligence,” is actually due to the intelligence of the human editors.
There is still reason for concern about GPT-3. It does show a remarkable ability to mimic human writers so teachers should expect a flood of poorly written term papers. However, the only human-like intelligence in the article is contributed by actual humans.
You may also enjoy:
New AI can create and detect fake news. But how good is it at either task? We tested some copy.