There is a disturbing new trend among students. Computer science students, particularly. No, it isn't a trend you will find on the clock app. It's the ubiquitisation of Large Language Models (LLMs) like ChatGPT, Bard, Copilot, and LLaMA. I believe we are going down an extremely dangerous path.
I present my arguments against LLMs in three parts. These are informed by my own experiences, research, and observations. The first part talks about the overall harms and risks, the second part talks specifically about the software engineering industry and the third part is, well, a sort of silver lining. As I complete my Master's from the University of Southern California, I reflect on the past six years as a student. I have lived through this change. Right from the publication of the ground-breaking "Attention is all you need" in 2017 to the public release of ChatGPT in 2022. Any conversation I have with my peers now is marred by the mention of one of these LLMs but mostly ChatGPT. Artificial intelligence has already taken over our conversations. PS: If you are a current student, skip to part two, the section titled "Just 'ChatGPT' it!"
Injurious to Health
LLMs are still experimental. Granted, they have been in the wild for a couple of years now but recall that OpenAI themselves are a young institution with an even younger product. They have to constantly update their content filtering and downstream harmful text detection algorithms as they acquire more users. They will continue to do so for a long time due to the generative and unpredictable nature of these models. It isn't as easy as it was before to disable the guardrails on ChatGPT but still very possible. Some early examples display the extreme harms of such models if content is not filtered. Research shows that ChatGPT and LLaMA are both heavily biased in the nationality and gender dimensions. How can you still, in good faith, use these tools to develop applications? No amount of content filtering, guard-railing, or fine-tuning will eliminate the bias it currently demonstrates. This is because of the amount of biased data these monster-sized models ingest during training. Generalized models come with a penalty.
People don't think about the energy impact of training and then serving such complex models. If a cloud service is offered for free, individual environmental conscience takes a hike. It is estimated that ChatGPT alone uses 1 GWh each day, which is the equivalent of the daily energy consumption for about 33,000 U.S. households. Models are getting larger, and are being trained with more data. The number of companies that are building products based on or using LLMs is also ever-growing.
And we are barely concerned with Intellectual Property (IP) issues this beast generates. The law will forever play catch-up with technology but certain issues will be too huge to ignore in the very near future. Independent artists, contributors to open-source software are all rightfully concerned that their work has been used by a for-profit organization to create a product that is so widely used. There are several gray areas with respect to the ownership of the content that it spits out as well. Does OpenAI own the content that comes out of ChatGPT? Does the prompt engineer? Who is then responsible when things go awry?
"Just 'ChatGPT' it!"
I hate when services become verbs. You may argue that these models are still safe to use in low-impact tasks such as writing emails, and captions or writing up simple scripts in Python. You forget about the far-reaching consequences of using these types of services on your personal development! I break down all the reasons I think you should abandon ChatGPT if you are a student in the following enumerated list.
-
While building your latest web-application, you come across an error. One of your React components is unable to render on your localhost. You quickly open ChatGPT, directly copy the entire error from your console to the chat window and blindly hit enter, hoping to find the solution quickly. As I see it, there are two major issues here. Firstly, debugging errors from warnings/logs/error messages is a very important engineering skill. Even better if you can cross-reference the original documentation and change log. Many times, in the past, I have been able to solve an error I encounter by carefully reading all related messages and performing a quick Google search to find associated documentation. Secondly, the act of googling the error, browsing multiple answers on Stackoverflow, poring through documentation itself has massive benefits! You have learnt 15 new instances where the same error could surface and are now adept in dealing with all of them. ChatGPT is only going to give you a single, point-fix. Often, it hallucinates this fix but that is an issue that requires its own point.
-
Imagine you rushing to your Teaching Assistant with a very important question or clarification regarding the logistics of your exam which is scheduled to begin within the next hour. Now, imagine her or him “making up” an answer instead of being honest and admitting their ignorance on the matter. You would be furious in the future, when you learn that you have been lied to or misinformed. That is exactly what ChatGPT does! Not only does it hallucinate information but it is eerily confident in it. Confident enough to be convincing, even. The gaps in ChatGPT’s knowledge are very evident when used in niche areas where there is little information available to begin with. Seek this information instead of a muddled version of it.
-
If you are constantly asking ChatGPT to write your emails for you or asking for help on your essays, you may risk under-developing your communication skills. I have watched (in awe) my senior colleagues draft up succinct and clear emails effortlessly, like a skilled weaver, putting together a beautiful Persian rug. That right there is decades of experience in communication. Yes, actual, penned communication and not auto-generated. Moreover, the more you write, the more of a style you develop. If you read enough text generated by a model, you will see that it lacks a style. It lacks color and it lacks taste. Personalization of communication is paramount. It helps you develop relationships, be it personal or professional. Imagine getting an apology, generated by ChatGPT.
-
Say you are asking ChatGPT to ideate for you. You are hosting a brunch this Sunday and wanted ideas for the menu. You run to your computer, key in a simple prompt and read the response. A fig and strawberry salad, lamb chops and a hearty leek soup sounds perfect. However, what about all the things that ChatGPT did not offer in its response? What about all the other, thousands of things, you may have enjoyed better than the suggestions but never saw the light of day because you based your menu of a (supposedly) artificially intelligent agent? We are taking away the creative elements embedded in everyday tasks. We are firing our organic neurons less often. I fear that we are losing something that is innately human by offloading work onto computers; thinking. This low-impact example only demonstrates the harms of LLMs. Extension to a high-impact example is left as an exercise for the reader.
Silver lining?
Some tasks that require little supervision such as formatting, greatly benefit from the use of such services. Mundane, repetitive or low-impact tasks should be automated using ChatGPT as it can save certain woman-hours. However, there are studies that show that these models are not well-behaved even with templates or greatly vary their responses on certain semantic alterations of the prompt. The greatest benefit the world of computer science is seeing right now, due to the widespread penetration of these services is that of awareness. Once again, issues of fairness, tractability and transparency are in the limelight. Many technologies, before ChatGPT suffered from the same issues but now we see more conversations surrounding it. Hopefully, it leads to more solutions than problems. The 2024 year will be one for the history books as the peak of generative AI adoption meets multiple, large elections. Will misinformation reign supreme? We will never know.
I remember wondering how my parents got through college without a Google search. When I asked them, they curtly replied saying, "libraries and textbooks exist." I can see the next generation of students wondering how we got through college without ChatGPT. "Google search and Stackoverflow exist." What a time to be alive.