Typos as a symbol of prestige: How to write so it doesn’t sound like AI

“I wrote to five CEOs and four replied,” says Ben Horwitz, a student at Harvard Business School. CEOs don’t usually respond to emails from strangers. He also asked them to get coffee or attend a meeting with students — nothing too important. But Horwitz had a trick: he had created an app that mimicked the writing style of these executives, with typos, no greetings, just a single line of six or eight words. And it worked.

Horwitz has turned that trick into a business idea: the app Sinceerly, which converts formal and meticulously well-written texts — like those produced by AI — into something more disjointed. Sinceerly offers three levels of conversion: “subtle,” “human,” or “CEO,” which is the ultimate mode for sloppy writing.

“I make tons of typos myself, and my inbox is overflowing with AI junk,” Horwitz explains to EL PAÍS via message. “I realize that I’m already unconsciously writing in Sinceerly’s ‘subtle’ mode. I’m in ‘human’ mode for my good mornings (brief, but with typos), and I aspire to the CEO style: clear, to the point, and very, very brief,” he explains. Beyond the idea, which is half-joking and unlikely to become a big business, Horwitz wants us to notice something that’s already happening: we see so many texts written by AI that we dedicate some of our effort to figuring out if the author has cheated. One way to make it clear that something is human is to write with typos, a new pinnacle of human prestige in the age of AI. “Subconsciously, I’ve become a worse writer,” says Horwitz. “I realize that I unintentionally write with grammatical errors on purpose. So that people know I’m human. It makes me a little sad.”

He’s not the only one becoming depressed by this decline, due to AI’s relentless pursuit of perfection. “Now we doubt everything; whether a video or a text was written with AI. I’m very saddened by the direction this is taking, and the overuse of AI makes us less reflective,” says Nerea Satrústegui, a writer and professional editor.

Spelling mistakes as a reward

Typos and a certain carelessness in grammar are something people have accepted more quickly than it seems. Julio Alonso, professor of AI at UNED (Spain’s open university), detected something like this back in 2023, but he still feels compelled to assign it some value: “A few months ago, I confessed to another professor that now, when I find spelling mistakes in a student’s work, I have to control an impulse to give it a higher rating, because my subconscious interprets those kinds of errors as a sign of effort, that they haven’t delegated that work to AI. He replied laconically: ‘And you think the students don’t know that?’” he explains.

The vicious cycle that could be created if we all continue down this path will be ironic: “This poses a circular problem: people add errors to signal that it’s human, and then we need AI to do the same. I hope we don’t get to that point, although I have no idea,” says Jenna Russell, a researcher at the University of Maryland.

All of this is happening because there are humans who want to make it clear that they don’t use AI to write. It’s a valuable signal today, at least for now. A journalist from The New York Times revealed on X that a publicist had told him that email open rates increased when they included a typo in the subject line because it was a human sign. The same applies to written text: “We place a lot of value on stress testing. So it’s logical that someone who writes by hand is starting to worry about whether their style might lead people to mistake them for an AI,” says Alonso.

Typos are the easiest way to avoid appearing AI-generated. But for those who want to write correctly, it’s not easy to stand out. There are resources that AI uses more frequently, and which humans, upon seeing them, intuitively recognize as machine writing: hyphens, the phrase “it’s not this, nor that, but this over there,” or periods and a moral closing. But it’s a dubious method. Hyphens no longer appear as often because OpenAI made its ChatGPT use them less. Furthermore, knowing too well how AI writes can have unpleasant implications: “Linguistic markers can change, but as things stand right now, the more AI-generated text you read, the more obvious these signals become, and more and more people are talking about them and keeping track of what the current ones are,” says Russell, who has found that those who best detect AI-generated text are those who use it the most.

According to Russell, there are some tools like Pangram and GPTZero that are acceptable for detecting text, at least in English: “They are generally reliable,” he says.

There is, however, no reliable solution beyond finding your own voice and having something to say, which doesn’t seem like a panacea either. “To avoid sounding like AI, you need to know who you are when you write. To know your voice, your tone, your filler words, your way of constructing an idea. No tool can give you that. It can help you in the process, yes, but you have to make the effort. Your voice, your tone, is the first thing you lose when you delegate writing without that prior work,” says Satrústegui. It’s a path, but it doesn’t truly solve the problem.

The best option is to get to the root of the problem: not the writing itself, but the ideas. Russell has just published a paper that attempts to determine whether the specific ideas in a text were conceived by a human or an AI. They discovered that by extracting features from fictional stories based on more than 10,000 human-written narratives and 50,000 written by AI, they were able to train a simple model to detect whether a story is AI or human-written: “And without using any stylistic features!” says Russell. “But there’s still a long way to go in understanding how to detect structural differences between human and AI thinking. And as AI improves, who knows how far apart we’ll be in 10 or 20 years,” he adds.

Sign up for our weekly newsletter to get more English-language news coverage from EL PAÍS USA Edition