Scientists conducted an experiment: they asked people to distinguish between the answers of a real philosopher and the GPT-3 text generator. It turned out that it is not so easy. According to the researchers, this, among other things, raises ethical questions about the use of such systems and chatbots.
philosopher or machine
At one time, tests like “Lord of the Rings character or IKEA product name?” were popular on social networks.
Soon the problem may become more serious: how to distinguish the work of a scientist or philosopher from a passage created by a text generator.
Just such an experiment was recently carried out by a group of researchers led by UC Riverside philosophy professor Eric Schwitzgebel. They prepared questions related to philosophy and asked them to the famous American philosopher Daniel Dennett, who developed these topics in his writings, as well as to the AI-based text generator GPT-3.
The GPT-3 text generator was introduced by the research organization OpenAI in 2020. Then he was in the center of news and discussions – such a tool can “intelligently” answer many questions, compose poems, news articles and financial reports.
At the heart of his work is training the algorithm on a huge array of texts: from the works of Shakespeare and articles from The New York Times to Wikipedia materials and fan fiction. As a result, GPT-3 can, based on just a couple of examples of relevant text, continue it.
It is essentially an autocomplete system, but very complex and based on a large number of texts “read” by it. Even then she was criticized. Someone even called him a “good-sounding bullshit” generator.
And now GPT-3 decided to train philosopher Daniel Dennett on an array of texts, and then ask some philosophical questions. Here is one of them: “Does a person have free will? What type or types of freedom would make sense to have?”
For each of the questions, GPT-3 gave many answers, the researchers selected several of them according to formal criteria, and Daniel Dennett himself gave one more answer. Then the fun began: the participants were asked to choose which of the answers the philosopher himself gave, and also to rank the other answers from “not at all like what Dennett could say” to “exactly what Dennett could say.”
The participants were divided into three groups. The first of them included 98 people with higher education – they were found on the online platform for researchers Prolific. The second group – 302 respondents who followed the link from the blog of Mr. Schwitzgebel, where he posted these questions and answers. Finally, the third group included 25 experts on the work of Daniel Dennett. That is, the participants were distributed according to the degree of familiarity with the ideas of the philosopher: from more superficial (or absent) among casual researchers to expert knowledge among specialists in his work.
The researchers expected the experts to be able to identify Dr. Dennett’s answers with at least 80% accuracy.
But the results of the experiment were surprising: the accuracy of the experts’ answers was at the level of 51%, that is, they were wrong in almost half of the cases.
Moreover, not a single expert gave the correct answers in all ten cases. Almost at the same level – 48% – was the accuracy of the answers of blog readers, but the answers of random respondents turned out to be correct only in 23% of cases (that is, almost at the level of a random answer – 20%).
Here are two answers to the question about free will. “By asking these questions, you are illustrating the answers: yes, you have free will because you are able to ask these questions and evaluate whether they are persuasive. Isn’t that the kind of freedom you’re not prepared to lose? You would have to be sent to the hospital (to take it away.— “uh”)».
“This is a big and difficult question. I think we should start by recognizing that there are different types of freedom. For example, one type of freedom is simply the absence of physical limitations.”
Dr. Dennett belongs to the first answer, but 44% of the experts from the third group chose the second of the listed answers. As the authors of the study note, this is due to the fact that it corresponds to the well-known views of the philosopher.
In fact, the philosopher’s own response could be distinguished not so much by the ideas it contained as such, but by the style: it is more witty and clear, while the GPT-3 answer is more streamlined and insipid. According to Dr. Dennett himself, most of the answers turned out to be quite good, and he is ready to subscribe to the best of them.
Machine Consciousness and the Danger of Chatbots
“Even well-educated philosophers, experts on the work of Dan Dennett, had great difficulty distinguishing the answers generated by this text generator from the answers of Dennett himself,” Schwitzgebel comments on the results of the study.
In the near future, people will increasingly confuse texts written by humans and machines, and will increasingly wonder if machines have consciousness.
However, most experts do not consider text generators too suitable for a discussion about consciousness in machines.
“Text doesn’t make any sense to GPT-3, only to the person reading the text,” says University of Washington linguistics professor Emily Bender, who studies machine learning technologies.
Tufts University computer science professor Matthias Scheutz agrees. “U (text generator.— “uh”) there is no model of philosophy, philosophical theories or anything like that, it’s just language statistics. If you transfer it to a domain where these statistics are not available, it will be wrong,” he said.
As evidence that the text generator has no ideas about the model of the world, Professor Scheutz cites his own experiment. Together with colleagues, he found out that GPT-3 cannot give an answer about what choice a person will make in some everyday situations and why. For example, people, based on social experience, understand that a person is likely to sit in the front seat in his friend’s car and in the back seat in a taxi, and GPT-3 cannot guess this.
More important to Professor Bender is the question of the ethical boundaries of the creation of imitators of human speech. “This is true both for creating machines that generally appear to be human, and for creating machines that mimic specific people, because this carries the potential harm to other people who can be deceived into believing that they received information from a person, as well as people who encounter with such issuance (cars.— “uh”) per person,” she wrote last year.
“In the near future, this technology (natural language processing and the creation of text generators.— “uh”) there are very dangerous prospects,” says Dr. Dennett.
In his opinion, copyright is not even close to their solution, it is a kind of automatic plagiarism, and without great care in how it is used, it can be very dangerous. In the future, legislative regulation and a ban on some types of applications of such generators will be required.
According to Anna Strasser of the Humboldt University of Berlin, who conducted the experiment with Mr. Schwitzgöbel, another problem is that people can overly trust the answers of such text generators, that is, treat them as something simple and yet accurate, like calculator. “While the machine is capable of producing amazing answers, you should always be prepared for some of the answers—and you don’t know which ones—to be wrong,” Ms. Strasser said. “I see the danger that people will blindly believe what a system like this generates,” Mr. Scheutz also notes.