ChatGPT is the future

On Thursday, June 8, 2023, a lawyer by the name of Steven A. Schwartz stood with his head hung low—humiliated—before a federal judge in Manhattan. He was there to represent a man named Roberto Mata in his lawsuit against the airline Avianca. Schwartz had asked ChatGPT for past legal cases that could support his client's case against Avianca. ChatGPT cheerfully replied with case names like "Martinez v. Delta Air Lines, Zicherman v. Korean Air Lines and Varghese v. China Southern Airlines," and—perhaps feeling a similar sense of cheerfulness at the sheer convenience of this newfangled online tool—Schwartz went ahead and cited these cases in the final document he submitted to the court. However, any such cheerfulness quickly dissipated, for if Schwartz had done any cursory examination of the cases provided by ChatGPT, he would have soon discovered that none were real. ChatGPT had made them all up.

ChatGPT is a form of artificial intelligence. The term "artificial intelligence" conjures science fiction images like HAL 9000 from the film 2001: A Space Odyssey or Agent Smith from The Matrix or Skynet from The Terminator. In reality, modern artificial intelligence is just a bunch of math (albeit extraordinarily fascinating math!). ChatGPT takes in an immense amount of data and generates an internal model to predict the most likely next word in a sentence, given the context of the many words that preceded it. At a high level, that is what modern AI is: a system that constructs a mathematical model for a particular task given a set of data and uses that model to make predictions, and that learns from the outcome of those predictions to improve the model. In the case of ChatGPT, those predictions were limited only to what words should come next in a particular context. As Steven A. Schwartz learned, ChatGPT's responses are not guaranteed to be factual.

ChatGPT is not the first of its kind. Chatbots—computer programs that try to engage in natural language conversations with human users—have existed for decades. As a kid, I remember trying out Cleverbot, a chatbot that launched in 2008 and was the successor to Jabberwacky, a chatbot launched in 1997. I only tried Cleverbot once, and I remember quickly becoming bored. As I recall, Cleverbot's responses tended to be short and uninspiring, often just one or two sentences (or even just "yes" or "no").

Because of my experience with Cleverbot, I was initially skeptical of ChatGPT when OpenAI first released it to the public in November 2022. One of my first questions to ChatGPT was, "What is the airspeed velocity of an unladen swallow?" The chatbot first tried to give a numerical answer to the question, and I was disappointed because it did not mention that the question was a reference to the movie Monty Python and the Holy Grail. (My disappointment slightly abated when ChatGPT correctly identified the reference on my follow-up question: "What was my previous question a reference to?")

Nevertheless, the more prompts my friends and I threw at ChatGPT, the more astounded I became. Within seconds, ChatGPT could write a multi-paragraph, remarkably factual essay where Cleverbot would have returned a boring, single-sentence response. Within seconds, the program could compose a parody of a particular pop song with lyrics in support of Georgia Tech football. Within seconds, the program could write code capable of solving almost any simple programming question in almost any programming language we threw at it, including some obscure ones like OCaml, Scheme, Typed Racket, and even one called Brainfuck. Late last year, when I was preparing for a software engineering job interview, I read online that a common interview question at a particular company was to design a program that could simulate a game of checkers. Thus, I asked ChatGPT to generate a Python program that could simulate a game of checkers. Within seconds, ChatGPT cheerfully responded with a working solution. Lo and behold, when the interview arrived, simulating checkers was indeed the question.

Today, I am thoroughly convinced. ChatGPT is the future. It is revolutionary technology. When I talked with my friends, some of them waved off my excitement, pointing to cases where they were able to get ChatGPT to respond with factual inaccuracies or uncreative stories (indeed, it's not hard to do—my mathematical friend was able to get ChatGPT to claim that Fermat's last theorem was false, although after asking follow-up questions, ChatGPT was able to realize it was wrong and even apologized for its mistake). My excitement, however, is not about today. What we are seeing today is the worst iteration of the new technology. In other words, it is only going to get better from here. We not only have to start discussing the current problems with the technology but also the future problems that we will need to deal with as this technology continues to improve and proliferate.

Students at all levels have started cheating by using ChatGPT to write their essays and solve their problem sets. It has given rise to new programs that attempt to detect whether a student used ChatGPT to write their paper, but these ChatGPT detectors are not infallible. We have already seen cases where instructors have accused their students incorrectly, jeopardizing their abilities to graduate. Rather than rely on ChatGPT detectors, the new technology has forced educators to change how they teach. According to an article published in the Times, "Gone are prompts like 'write five pages about this or that.' Some professors are instead crafting questions that they hope will be too clever for chatbots and asking students to write about their own lives and current events."

The story does not end at text generation. 2022 also saw a rise in the prominence of AI art generation, notably when an AI-generated digital painting won an award at the Colorado State Fair's annual art competition. The award sparked intense debate over the ethical implications of AI-generated artwork. Whereas ChatGPT produces text in response to short prompts from the user, artwork-generating AI programs produce artwork in response to short prompts from the user. Does the user who wrote the prompt deserve to be awarded for the quality of the AI-generated artwork? It doesn't seem fair to compare AI-generated art to art that humans crafted by hand—if any competition in art is even desirable, human-created artwork and AI-created artwork should be judged separately. As you recall, AI works by trying to make predictions based on observed data and learning from the outcomes of those predictions. Given the right prompt, an AI might produce an artwork similar to the style of a particular human artist whose work was part of the dataset on which the AI trained. In the future, courts could have the option to construe such works as "derivative works" under copyright law, which are creative works "based on one or more preexisting works". By law, in order to publish a derivative work, the creator must first get the permission of the copyright holder of the original work (although copyright holders often choose not to enforce this—for example, most fan fiction stories on the Internet are derivative works published without permission, but the copyright holders rarely go after fan fiction authors because fan fiction often benefits the copyright holder more than it harms, as it generates more interest in the franchises).

The last decade has also seen the rise of "deepfake" technology, in which AI programs can digitally recreate a person's likeness. In the 2016 film Rogue One: A Star Wars Story, Disney used computers to digitally recreate characters whose original actors are now deceased, such as Grand Moff Tarkin (Peter Cushing, d. 1994) and Princess Leia (Carrie Fisher, d. 2016). Disney has continued this practice with recent releases, such as digitally recreating a young Luke Skywalker in The Book of Boba Fett (2021–2022) and The Mandalorian (2019–). Luke Skywalker's original actor, Mark Hamill, is not dead but is now much older than what the digital recreations portray. Where should the line be drawn here? It is not difficult to imagine a case in the future where a particular actor could be too expensive to hire directly, so a film production company could decide to have that actor's likeness computer-animated without the actor's permission as a cheaper alternative. That does not seem fair to the actor.

In the past few years, we have also seen dramatic improvements in the ability of artificial intelligence to imitate someone's speaking voice based on just a few seconds of audio. There is now a sea of content on the Internet featuring AI-generated voice imitations. For example, countless YouTube videos feature AI-generated voices of Joe Biden and Donald Trump arguing over topics you would never expect the actual figures to have even heard about, especially video games, ranging from Splatoon lore to which Legend of Zelda game is the best to which Kingdom Hearts game is the best. Although many of these audio clips are supposed to be humorous, they have sparked controversy among voice actors, who see their jobs in jeopardy from this new technology. In an interview, Mike Pollock, the voice actor of Doctor Eggman from the Sonic the Hedgehog franchise, said that he routinely stumbles across AI-generated audio clips based on his characters' voices. He has tried to contact the apps that generate these voices, as well as the YouTube users who upload videos featuring them, to remove him from their platforms. It is unclear whether these platforms have any legal obligation to honor his requests.

New regulations for the use of artificial intelligence are necessary. As a starting point, there needs to be a clearly defined principle that everyone owns their own voices and likenesses by default. In order to publish AI-generated content simulating someone's voice or likeness, the law should require you to get that person's advance permission. This would be an equitable doctrine that not only protects the interests of actors but also prevents harassment of ordinary individuals. It is not hard to imagine a world in which someone with a grudge against you might try to defame you by publishing a fake version of your voice saying something you would never say. These restrictions do not have to be indefinite. Currently, in the United States, creative works are typically protected by copyright until 70 years after the author's death. Perhaps it doesn't need to be quite that long, but something similar could apply to human voices and likenesses as well.

We will also need to educate the public carefully about the capabilities of artificial intelligence. Stephen A. Schwartz's humiliation before the U.S. district court resulted primarily from his personal lack of diligence (even if ChatGPT were the magical search engine he thought it was, he should have still double-checked to make sure the cases he was citing actually supported the arguments he was making), but perhaps Schwartz would have had the instinct to be more diligent if we had encouraged him and other members of the public to participate more in the debates surrounding artificial intelligence.

These are the concerns swirling in my head as I continue to interact with these new technologies. On the whole, however, my primary emotion remains excitement, not worry. One day, not too long from now, there will be a kid in the bedroom of her parents' house who will have an amazing idea for a movie. To see her idea come to life right before her eyes, she will need only to write it down.