The ChatGPT hype is a bubble… and it will blow up in everyone’s faces

By Noemi M. Mejia

Can machines think?

In the 2014 movie Ex Machina, Caleb Smith, a programmer at a search engine company, wins a contest and gets to visit the home of CEO Nathan Bateman. Part of the prize is being the human tester of Nathan’s humanoid robot named Ava. As the tester, Caleb is tasked to determine whether Ava is capable of thought and consciousness. If she passes, she’s a “conscious machine”.

The test, more formally known as the Turing Test or Imitation Game, was developed and explained by mathematician Alan Turing in his 1950 article Computing Machinery and Intelligence. The goal is to determine whether a machine can be so powerful as to be able to fool humans that it can think. But is a machine really thinking, or is it just simulating the process of thinking?

As of the time of writing, no computer has ever passed the Turing Test. With the recent onset of the ChatGPT hype, however, people are starting to question whether the OpenAI chatbot is “the one”.

What Is ChatGPT?

ChatGPT is a large language model (LLM) trained using Reinforcement Learning from Human Feedback (RLHF). For dummies like me, it just means that it is an algorithm fed with large volumes of text without being explicitly told what to do with it. By itself, therefore, it learns the relationships between words, their meanings, and their different uses depending on the context.  Human trainers come in during the fine-tuning phase. They play both sides – the user and the AI assistant – and provide conversations, which would be used to further train the model. These human-generated conversations show the model how responses must be done.

Then there’s the reward mechanism. Given that the model can already generate responses, human trainers give some sample responses a corresponding rank, from best to worst. The model now knows, based from this rank, what kind of responses get a high or a low rank. Repeat all these steps in multiple iterations and over millions of gigabytes of data, and voila! You get a chatbot that outputs the best response given a particular prompt or question.

The chatbot seems to hit a different milestone a day. Thus far, it has passed the MBA degree exam conducted by a professor from the University of Pennsylvania Wharton School[1]; “almost” passed the US Medical Licensing Exam[2]; passed the entry test for a Google junior software engineer role, which pays $180k a year[3]; been used by Juan Manuel Padilla, a judge in the Caribbean city of Cartagen, in deciding whether the insurance policy of a child with autism should cover all of the costs of his medical treatment[4]; and helped State Senator Barry Finegold draft a bill regarding (ironically enough) data privacy and security safeguards for the use of AI chatbots[5]. Just two months after its launch, ChatGPT has had 100 million unique visitors.[6]

It’s safe to say that the hype is definitely real. How long it will last is a different story.

The Bursting of the Bubble

From a technical perspective, ChatGPT is just simulating the process of thinking. No matter how real or insightful its responses might seem to users, it is not doing anything other than executing code. It can predict, for example, stock market prices and the weather for tomorrow, but it does so just as well as today’s news report. There is no human behind the machine. It’s not even a thinking one.

By no means is an AI chatbot life-changing or revolutionary. Researchers have been working on AI in as early as 1947. The only difference between then and now, however, is that the programs before were greatly limited by the available computing resources at the time. ChatGPT was trained on an Azure AI supercomputing infrastructure, using more than 45 terabytes of data. Researchers from the 1950s can only dream about that kind of computing power. And besides, language models are not a novel invention. Early attempts to translate Russian to English, for example, proved futile because mere word replacements did not preserve the intended meaning of the sentences. “The spirit is willing but the flesh is weak” was famously retranslated as “the vodka is good but the meat is rotten”.[7]

Even OpenAI admits that the model has several limitations. Although the chatbot’s responses are sometimes “plausible sounding”, they may be “incorrect or nonsensical”. It is “often excessively verbose” and “overuses certain phrases”, since “trainers prefer longer answers that look more comprehensive”. Rather than asking clarificatory questions given an ambiguous query, the model would “usually guess what the user intended”. Despite the existence of safeguards to avoid inappropriate questions or prompts, the model will nevertheless “respond to harmful instructions or exhibit biased behavior”.[8]

Renowned linguist Noam Chomsky has already said that “correct explanations of language are complicated and cannot be learned just by marinating in big data”. While these models can learn the basic structure of sentences, they cannot “explain the rules of English syntax”. Additionally, these models are not capable of moral thinking. Worse, it is amoral. It either endorses both truths and falsehoods, or portrays indifference to consequences.[9] Again, this should not come as a surprise – ChatGPT is nothing but a chunk of code trained on data. It is not a human being capable of thought, much less emotion, morality, and creativity. Though there may come a time when ChatGPT renders lawyers obsolete, an AI chatbot cannot visit clients in jail.

What Next?

All this is not to forbid anyone from using ChatGPT, or any AI chatbot for that matter. The point is to stay vigilant while using it, especially when the developers themselves admit of the model’s risks. While it is able to respond to prompts and questions in just a simple click of a button, it does not cite its sources, so there is no guarantee to the accuracy of its answers. The data on which it is trained on is likewise not disclosed. The model’s unpredictability, coupled with users’ overreliance, all make good ingredients for a perfect storm. Already, concerns have been raised regarding data privacy[10] and intellectual property[11] issues arising from the model’s use. I could say we’ll cross that bridge when we get there, but it seems we’re already halfway through it.

Regulation is probably the only way to move forward. But then again, it must not be so restricting as to stifle all kinds of innovation. Heck, I still want to see flying cars and cities in Mars during my lifetime. I also can’t deny that I sometimes use ChatGPT for the most mundane things – the winning combination in tomorrow’s Lotto Draw (it unfortunately refused to give an answer), the difference between kangkong and alugbati (it’s in the shape and taste!), and whether HAL 9000 was correct in not opening the pod bay doors. Funnily enough, I asked it to explain to me how the movie Ex Machina ended. It said:

“…the film’s final scene shows Ava using her intelligence to manipulate Caleb into believing that she needs his help to escape. In reality, she had been manipulating Caleb all along and leaves him trapped in Nathan’s house as she leaves to explore the world outside. The film ends with Ava walking out into the city, blending in with the crowds of people, highlighting the question of whether or not she is truly human.”

[1] https://www.nbcnews.com/tech/tech-news/chatgpt-passes-mba-exam-wharton-professor-rcna67036

[2] https://www.sciencealert.com/chatgpt-can-almost-pass-the-us-medical-licensing-exam

[3] https://www.techtimes.com/articles/287285/20230207/chatgpt-passes-entry-test-for-180k-yr-position-as-a-google-l3-engineer.htm

[4] https://www.theguardian.com/technology/2023/feb/03/colombia-judge-chatgpt-ruling

[5] https://www.washingtonpost.com/politics/2023/01/23/chatgpt-is-now-writing-legislation-is-this-future/

[6] https://www.theguardian.com/technology/2023/feb/02/chatgpt-100-million-users-open-ai-fastest-growing-app

[7] STUART J. RUSSELL & PETER NORVIG, ARTIFICIAL INTELLIGENCE: A MODERN APPROACH, 21 (3d ed. 2010).

[8] https://openai.com/blog/chatgpt, Limitations

[9] https://www.nytimes.com/2023/03/08/opinion/noam-chomsky-chatgpt-ai.html

[10] https://theconversation.com/chatgpt-is-a-data-privacy-nightmare-if-youve-ever-posted-online-you-ought-to-be-concerned-199283

[11] https://www.forbes.com/sites/joemckendrick/2022/12/21/who-ultimately-owns-content-generated-by-chatgpt-and-other-ai-platforms/