Striving To Stay Human In an AI World

“In a place where there are no humans, strive to be human.”

– Pirkei Avot 2:5

We are now two years out from the public debut of AI large language models (LLMs) – ChatGPT being the most well-known of these. These tools have become seemingly ubiquitous in workplaces and schools. It is often tempting to rely on these new technologies to outsource an email reply to a coworker or produce a catchy title for a paper. These LLMs are, admittedly, good at fixing computer programming code and summarizing text. It’s no surprise, then, that a Pew Research Center study from earlier this year found that a quarter of Americans have used this kind of online chatbot.

As an environmental sociologist, I often use surveys to see what people think about climate change and how we can solve it. A year ago, I fielded a survey to see what people thought about hydrogen energy, a new way of powering industrial manufacturing that minimizes its climate impact. I asked my respondents to give me any questions they had. Since hydrogen energy is still relatively unknown by the general public (I myself hadn’t heard much of it before starting this project), I expected questions about how it worked, its applications, and its safety. And these definitely poured in.

But I was also flooded by responses that looked perfect but, upon closer inspection, were probably fakes. How can this be? A normal person would probably give a short response, no more than 10 or so words. Many humans wouldn’t capitalize or use punctuation, and they’d probably give only a few questions. But an LLM would. In fact, when I tested it myself, one AI gave fifteen questions, and another gave twenty. All the LLM-written responses were perfectly spelled, capitalized, and punctuated. (The research paper for my study is available online for free. Both these results and the sample AI responses are included there.) 

Though these responses may have been distinguished by being overly “perfect,”  AI is bad at many, many other things. To know why, we need a quick understanding of how these LLMs work. Like all artificial intelligence models, LLMs “learn” based on what is given to it. From there, it takes some request and produces an output that it thinks best answers this request based on its previous training. Like a child who learns 2 + 2 = 4 and can then figure out that 2 + 3 = 5, an LLM can make reasonable guesses at what answer to give. But, just like how the jump from addition to multiplication is a big hurdle in elementary school, an LLM cannot magically come up with a response it hasn’t learned, let alone predict the future. Ask it who will win the next season of The Bachelor, and it won’t even know where to start. These AI chatbots have already received criticism for parroting climate change denial, election denial, and vaccine skepticism it got from blogs and other websites.

Because these chatbots can only repeat information they have received, when there is not enough information, they don’t simply admit that they don’t know, and  start to make inaccurate and often bigoted assumptions based on what has been fed to them. Recent research has found that LLMs tend to flatten the human experience to a singular point of view, seeing all members of a certain race or sex as having the same opinion. Each person has a unique and valuable perspective, but these bots reduce whole groups of people to monoliths at best and stereotypes at worst. For a people that prizes its diversity of opinions (“two Jews, three opinions”), we cannot allow our experiences to be flattened or simplified.

LLMs have other problems, too. The high amounts of computer processing power needed to calculate their responses require a lot of water to cool the processors, and the additional electricity required has caused fossil fuel plants to reopen. LLMs have also been accused of stealing work from news sites, such as The New York Times. Interestingly,  Jewish authors’  work has been used for LLM training data more than other groups of people, making us more vulnerable to LLMs’ plagiarism.

When I looked through this data with two research assistants, we suspected about 7.5% of my responses were copied from an LLM. I tried using a few tricks that are standard in online surveys for weeding out respondents who aren’t paying attention, and the LLM responses jumped to over 9%. 

Computerized methods of detecting AI usage aren’t perfect, either. There are valid ways that people can produce a response that looks like AI. Somebody who is using text-to-speech (like in the dictation feature on your phone) will usually produce text with perfect spelling, capitalization, and punctuation, all in a shorter amount of time than a person typing would.

The reason LLMs are successful is that they work. Until I looked more closely, I thought that some respondents were just writing eloquently and formatting everything well. But this was a mirage. The automatism of LLMs threatens the creativity and individualism of human experience that is so often celebrated in Judaism. Rabbi Lord Jonathan Sacks writes that the “use of language, not to describe something already in existence but to create something that didn’t exist before, that links us to God.” LLMs copy what already exists; only humans can create something new. Only through our humanity can we link ourselves to God. 

(This post is part of Sinai and Synapses’ project Scientists in Synagogues, a grass-roots program to offer Jews opportunities to explore the most interesting and pressing questions surrounding Judaism and science. Frederic Traylor is a PhD candidate in Sociology at Rutgers University and led science programs for kids at Anshe Emeth Memorial Temple in New Brunswick, NJ as part of the grant.)