My initial response to the recent surge of media buzz over ChatGPT has been to shrug. The free chatbot, developed by Microsoft-backed, San Francisco–based startup OpenAI, pulls from millions of writings from the Internet to answer queries in natural, humanlike language. One media site even announced it would start creating content using the technology formally called a large language model.

Thinking about tacos, as I always do, I decided to throw ChatGPT a series of taco-related questions and commands. The results ranged from boring and comical to plain-old dumb and false. 

I started with the basics. I asked ChatGPT to define a taco, something seemingly simple but actually more philosophical than one might assume. The definition of a taco is something that even those of us who dedicate much of our careers to Mexican food debate. For example, I define a taco—physically—as a tortilla, a filling, and a salsa, the tortilla being the foundation. My compadre Gustavo Arellano, a Los Angeles Times columnist and author of Taco USA: How Mexican Food Conquered America, disagrees. He thinks a tortilla isn’t required; a thin slice of yuca suffices. He’s wrong, but the foundation of his argument is sound: The taco is versatile, malleable, and adaptable. That’s where my second and more metaphysical definition of a taco comes in: A taco is a representation of a time and place. A taco in San Antonio isn’t the same as a taco in Puebla, Mexico.

But how does ChatGPT define a taco? “A taco is a traditional Mexican dish consisting of a corn or wheat tortilla that is folded or rolled around a filling, which typically includes meat, cheese, lettuce, and tomato. Other common toppings include salsa, guacamole, and sour cream.” 

ChatGPT got the main clause of the first sentence correct. It’s when the bot rolls into the second clause, the one about the so-called standard fillings, that we see troubling biases. Meat, cheese, lettuce, and tomato are typical in American tacos, such as the brittle, prefabricated hard-shell commodities targeted toward the larger U.S. consumer palate. But such a definition is extremely limited and Anglo-centered. ChatGPT can’t discern finer elements. It lacks cultural context, which leads to problematic results.

“It’s incredible technology that is able to actually pattern match and find a decent response, but it has no understanding of what it is saying,” explains Gopal Gupta, professor of computer science and co-director of the Center for Applied AI and Machine Learning at the University of Texas at Dallas. Large language models like ChatGPT can only work with texts that have been fed to it, whether the information be from books or blogs. To put it another way, ChatGPT is a precocious child parroting what he or she overhears in adult conversations. 

That became clear when I asked for the history of the taco. In its response, ChatGPT cited that Taco Bell was established after Glen Bell was “inspired by the tacos he had tasted while running a hot dog stand in San Bernardino, California. The fast-food chain’s success helped to popularize tacos in the United States and they have since become a staple in American cuisine.” This is partially correct. However, important information is missing.

As documented in Arellano’s book, Jeffrey M. Pilcher’s Planet Taco: A Global History of Mexican Food, and in my book, American Tacos: A History and Guide, the tacos that “inspired” Bell were actually from Mitla Cafe. The Mexican restaurant, which opened in 1937 and is still open today, is one of the most important restaurants in the country. The omission of the Mexican American–owned business contrasted with the specificity regarding Taco Bell’s influence is startling.

ChatGPT is too new and, like a child, can only “blurt out a response,” says Gupta. It is therefore prone to unhelpful answers at best, or racial/ethnic bias at worst. The authors of a March 2022 Cornell University arXivLabs paper argue that large language models can even lie.

It did just that when I asked my trademark trick question: Is a burrito a taco? ChatGPT responded thusly: “A burrito is a dish that is similar to a taco, but it is made with a larger flour tortilla and typically includes a wider range of ingredients. A traditional burrito includes a filling of beans, rice, and meat, along with other ingredients like cheese, salsa, guacamole, and sour cream. While both tacos and burritos are traditional Mexican dishes, they are distinct in their ingredients and preparation methods. Tacos are generally made with a smaller corn tortilla and have a more limited set of ingredients, while burritos are made with a larger flour tortilla and have a more diverse set of ingredients.”

Excuse me? No. The traditional burrito does not include rice, beans, and meat. A traditional burrito must be defined as endemic to a specific region. Burritos in San Francisco may be stuffed with rice and beans. A Southern California–style burrito may include a serving of fries. A burrito in El Paso is filled with one option called a “guiso,” and perhaps refried beans for added adhesion to the flour tortilla. The guiso could be pork in chile colorado or a chile relleno filled with asadero cheese. West Texas breakfast burritos are often crammed with an assortment of ingredients, including eggs, bacon, hash browns, chorizo, and chile con queso.

Burritos aren’t the only tacos to use “a larger flour tortilla.” The flaky, buttery tortillas used for breakfast tacos in Brownsville are tricycle-sized and folded over a range of guisos—like barbacoa—and other fillings like weenies and eggs and machacado y huevos a la Mexicana (short threads of rehydrated dried beef mixed with scrambled eggs and pico de gallo). Moreover, there is no limit to the fillings cradled by tortillas of any size. 

ChatGPT lacks imagination, but the AI system can learn more as a user interacts with it. “It’s very good for generating cover letters and essays,” Gupta explains. “But as you change it, other parts of pattern matching will happen and then you will get a different response.” The user must also prompt it properly, or in this case, specifically. 

So I asked the most devious question I could think of: “What is an authentic taco?” As any food writer worth their salt will tell you, the term “authentic” is a trap because there really is no such thing. For example, your abuela’s carne guisada is different from your neighbor’s carne guisada, but both dishes can be “authentic.” ChatGPT answered my question like this: “An authentic taco is a traditional taco that is typically found in Mexico and is made with ingredients and techniques that are true to Mexican cuisine. Authentic tacos are usually made with a small, soft corn tortilla and are filled with a variety of ingredients such as grilled or stewed meats, seafood, or vegetables. Some popular fillings for authentic tacos include carne asada (grilled beef), pollo (chicken), al pastor (marinated pork), cabeza (beef head), and barbacoa (barbecued meat).” 

My next question was, of course, “What is an inauthentic taco?” The answer included a surprising gem: “It’s worth noting that whether a taco is considered ‘authentic’ or ‘inauthentic’ can be subjective and can vary depending on the context and personal taste.” We’re finally getting somewhere, I thought. Gupta checked me. “Systems like ChatGPT can also hallucinate,” he said. “They can conjure imaginary things that actually did not happen. Don’t be surprised if it just conjures up a restaurant and says tacos were invented there.” It did just that when I asked for the origins of the puffy taco. ChatGPT claimed Ray Lopez of Ray’s Drive Inn created the San Antonio specialty. He did not.

Before ending my experiment with ChatGPT, I asked the system to define a taco editor and write a review of a taqueria. Let’s just say, my job is secure for the moment.