


ChatGPT was released two and a half years ago, and we have been in a public panic ever since. Artificial intelligence can write in a way that passes for human, creating a fear that relying too heavily on machine-generated text will diminish our ability to read and write at a high level. We’ve heard that the college essay is dead, and that alarming number of students use A.I. tools to cheat their way through college. This has the potential to undermine the future of jobs, education and art all at once.
The Titanic is indeed headed toward the iceberg, but the largest problem — at least at the moment — is not the college essay, the novel or the office memo. It’s computer code. I realized this last year when I was teaching a course on A.I., language and philosophy. When I asked my students how they use chatbots, one told me that whenever he has a spreadsheet full of data (such as results from a lab experiment or information collected from a survey), he was trained in high school to write a quick bit of code to parse and analyze that data. But now, he told me, he just throws the spreadsheet into ChatGPT, which analyzes it more quickly and requires him to do almost nothing.
That’s when it hit me: A.I. is just as much a challenge to numeracy — our knowledge and ability to use mathematics and reason quantitatively — as it is to literacy.
In February, the A.I. engineer Andrej Karpathy reported on X that he was engaged in a new form of software development he called “vibecoding.” Using nothing more than a series of spoken prompts to a chatbot, he was conducting ad hoc experiments on data and said he would “barely even touch the keyboard.” He said this allowed him to “forget that the code even exists,” leaving the grunt work to the A.I. and simply directing from above. Mr. Karpathy’s post went viral, and many others acknowledged they were doing the same.
By some accounts, though, vibecoding isn’t going well. The code that Mr. Karpathy’s prompts create has been reported to be inefficient and riddled with irreversible errors. Worse, programmers using the method say they’ve found themselves not merely forgetting that code exists but forgetting how to code. As is the case with reading and writing a language, code is one of those things where if you don’t use it, you lose it. Early studies indicate that humans who use A.I. could become less creative over time.
Something not unlike vibecoding has already entered the marketplace. Google claimed in 2024 that A.I. wrote over 25 percent of all of the company’s code, and Microsoft recently reported similar numbers as it fired thousands of employees, including many software engineers. Amazon has also adopted streamlined A.I. coding practices, which workers say changes software engineering fundamentally, making a job defined by intellectual effort into something more like industrial drudgery. A.I. companies themselves see the writing on the wall: OpenAI, for example, is in talks to spend a cool $3 billion to acquire Windsurf, a company that offers an A.I.-driven coding assistant.