AI is Overrated

Almost 40 percent of Americans have used generative AI, and 24 percent of Americans use it weekly. With such a high percentage, it’s worth asking: can we be sure that AI is trustworthy? I don’t think so. AI (specifically ChatGPT), as of the moment, isn’t as good as everyone says it is. It is inaccurate, biased, and has trouble reliably conducting research.

First, AI is not accurate. According to a study conducted by Purdue University, ChatGPT is wrong 52 percent of the time when asked questions about computer science. This is an astounding error rate for an AI model that claims to be good at coding. According to another test by Orhani, Senad, and Xhevdet Spahiu, ChatGPT has less than a 60 percent accuracy rate when it comes to math questions. 

These errors are called “hallucinations." These hallucinations can be as common as 27 percent of the time (all topics combined) with some AIs even hallucinating in response to a simple question. For example, when asked, ChatGPT said that the Golden Gate Bridge, which sits in California, was transported through Egypt. So, yes, AI is terrible at pretty much every subject except for creative writing.

Next, certain types of generative AI could be biased. Since AI is trained by humans who have biases using data that has bias, those biases get passed down to the model. Although both ChatGPT and Gemini, Google’s AI, have instructions, also called “guardrails”, not to give opinions on controversial topics (politics, for example), that doesn’t mean bias doesn’t show up. According to a study at UCLA, ChatGPT does have biases. Researchers asked ChatGPT questions regarding immigration and other topics, and ChatGPT’s answers tended to be left-leaning. They attributed this to the human input used in training the AI and humans’ implicit biases. 

Finally, AI isn’t good at research. ChatGPT can provide sources, but these sources are sometimes outdated, untrustworthy, or completely nonexistent. This makes ChatGPT even more unreliable, as it may say something that is not backed up by any evidence (again, such as the Golden Gate Bridge being transported through Egypt). To test this out myself, I tried asking it to give me recent events near where I live, and it ended up outputting events that happened in 2013. 

Although ChatGPT may have issues, it still is a major advancement. It can code much faster than any human and write poems that rhyme without missing a beat. If we continue to develop AI, we may see some of these issues lessen. However, as of right now, ChatGPT is very inaccurate in CS, math and many other subjects. It has biases built into it, even if these biases are unintentional; furthermore, ChatGPT won’t always give relevant results for research. For now, don’t believe everything ChatGPT, or any other generative AI, says, and make sure to check your sources. Oh, and maybe don’t use it to do your homework, because it’s really not good enough to get you an A…