The executives behind big generative artificial intelligence companies are quick to claim their products will displace huge numbers of workers. Anthropic CEO Dario Amodei made headlines in May by saying generative AI could wipe out half of entry-level white-collar jobs in the next few years. Meta CEO Mark Zuckerberg said in April that he wants AI to be writing half of the company’s code in the next year. And Americans believe it — a recent Pew survey found 64% of Americans expect fewer jobs thanks to AI.
In this environment, it’s easy to see research about AI-vulnerable jobs and start to panic. When Microsoft researchers put out a report in July with a tidy list of jobs with tasks that most and least overlapped with tasks that could be done by gen AI, it spurred consternation among those whose jobs were at the top of the chart. But dig a little deeper and there’s less need for translators, historians and others to worry about whether AI will replace them — unless human employers, enraptured by AI’s hype, decide so.
“I think it is useful for people to focus on the tasks as opposed to jobs,” Darrell M. West, senior fellow in the Center for Technology Innovation at the Brookings Institution, told CNET. “There may not be that many whole jobs that get eliminated. There certainly are going to be a lot of tasks that are going to be eliminated.”
The Microsoft research that ranked occupations by AI overlap made exactly that point, even if it wasn’t the key point making headlines.
“It is tempting to conclude that occupations that have high overlap with activities AI performs will be automated and thus experience job or wage loss… This would be a mistake, as our data do not include the downstream business impacts of new technology, which are very hard to predict and often counterintuitive,” the authors wrote.
Even the biggest names in generative AI will, if pushed, admit to that uncertainty. OpenAI CEO Sam Altman, speaking with Theo Von in a recent podcast appearance, said not too long ago it would’ve been difficult to imagine that people could have the jobs of AI company CEO and podcaster. “I think it’s very hard to predict exactly how something evolves or predict exactly what the jobs of the future are going to be,” Altman said.
There are hazards to thinking AI can do what it really can’t. Here’s a look at the two jobs cited by the Microsoft report as having the most overlap with tasks AI can do: translators and historians.
Translating is more than just finding the right words
Spanish is an official language in all or part of more than 20 countries across the world. That means there are more than 20 different variations of the language — and even more when you consider local and regional versions. Andy Benzo understands how important those distinctions can be. “I speak Argentinian,” she told me. “I don’t speak ‘Spanish.’ There’s no ‘Spanish.'”
As a legal translator and president-elect of the American Translators Association, Benzo has to understand not just the basic Spanish words, but the culture — and legal culture — behind them. Benzo, a lawyer, doesn’t just change words and sentences from one language to another — the meaning needs to be right. These translations might have serious ramifications for the people or entities involved in a legal proceeding, and it’s crucial to get the exact meaning correct.
Translators do more than just transcribe and convert documents. Medical translators help people communicate with doctors and nurses to ensure they’re getting proper care. These are literal life-and-death situations. Financial transactions that move from one language to another need to be clear, or else someone’s money or livelihood may be at stake.
Professional translators are generally experts not only in language, but in their specific field, Benzo said. “You pay us for what we know,” she said. “We say that what we do is accurate.”
Translation tools powered by generative AI are getting increasingly skilled at helping someone communicate in a language they don’t understand. You can hold your phone up and let it interpret between you and a person who doesn’t understand any language you understand, as demoed by Apple with iOS 26 and Google’s Gemini. But professional translators and interpreters specialize in getting things exactly right. You don’t want a translation that makes a good guess — which is really all you get from AI — when your money or your life is on the line. You want a translation that understands the nuances that vary between the languages. And you want someone who’ll be accountable if it’s wrong.
“If AI makes a mistake, who’s going to be responsible for that?” Benzo said.
Language is also not static. While the AI industry is fast-moving, language changes even more quickly. Every day, someone somewhere finds a new way to phrase an idea. The Cambridge Dictionary, for example, just added words like “skibidi” and “broligarchy,” which an AI with an outdated training dataset may not be able to understand. But a human, properly trained, can keep up with those subtle adaptations.
“Language evolves all the time,” Benzo said. “Language belongs to the people. Nobody is the boss of language. The only one who can perceive the nuances of a language is a human.”
History is more than just telling the same old story
Sarah Weicksel is a historian whose research is hard to find in books because it isn’t about words. She studies clothing, and not the kind you get targeted ads for online. Her work (including a forthcoming book) examines the physical clothes of the American Civil War era and how they reflected the economic and political circumstances of the time. Studying 160-year-old clothes entailed digging into parts of archives that seldom come out for museum display. (When clothes come out for an exhibit, they’re often there for a short period because they decay quickly.) It also involved looking through diaries and other historical documents and looking for references not to important, world-changing events, but to pants and shirts.
“My research process was very much putting together pieces of a puzzle,” Weicksel, now the executive director of the American Historical Association, told me.
But couldn’t an AI model look at museums’ clothing collections or read all those Civil War diaries? Not quite. The job of a historian is not to find the obvious, but to find the underlying story that is not necessarily at the surface. Weicksel looked at clothes to consider how the cut of a coat might help someone stand more upright, or the texture of different fabrics. “AI can’t touch and feel the things for me,” she said.
More importantly, Weicksel approached her research by trying to answer and understand specific questions that may not have been asked before. That’s the core of a historian’s work: Exercising judgment and creativity to discover new interpretations of the past.
Read more: ChatGPT Can’t Fix Everything. Here Are 11 Times You’ll Regret Using It
Weicksel said research like the Microsoft study, which looked at how well AI can handle individual tasks done by a professional historian, doesn’t cover the whole picture. Yes, historians maintain and edit documents and provide information to people, she said, but those tasks “are not the core of what being a historian is,” she said.
“We are not just a set of tasks that we complete and produce discrete things,” Weicksel said. “We are very much about the ability to synthesize and contextualize and bring judgment but also creativity to these questions that we’re asking.”
A large language model can give you a pretty good report on a historic event. Ask ChatGPT for a report on the Defenestration of Prague in 1618 and you’ll probably get a pretty good summary — unless it hallucinates and gets it confused with the other times people were thrown out of windows in Prague in 1419 and 1483. But to expect that AI can do the job of a historian because it can summarize or analyze a historical event is getting things backward. AI can summarize historical events because it stands on the shoulders of historians who have dug out the facts and written down what happened.
The study of history helps our understanding of the past evolve, but a machine trained to follow past trends might not find the unexpected or help us avoid repeating the same mistakes.
“Great works of history are neither predictable nor obvious,” Weicksel said. “That’s what makes them transformative. That can’t be replaced by a technology trained to reproduce existing patterns.”
Automation has affected jobs for as long as tools have existed. Artificial intelligence could improve the kinds of automation in some fields by improving robotics, like this robot in Beijing that can do the work of a human salesperson (or a vending machine).
What kind of work can AI do?
There’s a difference between the tasks AI can do and the tasks AI can help with. Large language models have proven to be adept at writing software code, leading to the proliferation of “vibe coding,” in which the role of a human is more to come up with the idea and troubleshoot the product while the AI does most of the work. AI has also been used more and more in roles like customer service, where more straightforward requests can be handled by a chatbot or something like it, leaving only the more complicated ones to humans.
A recent paper by researchers at Stanford University, which found declines in employment among young, early-career workers in certain automation-sensitive industries, also found those declines were primarily in roles where tasks could be automated.
“While we find employment declines for young workers in occupations where AI primarily automates work, we find employment growth in occupations in which AI use is most augmentative,” where it can make a human faster or more effective without replacing them, they wrote. “These findings are consistent with automative uses of AI substituting for labor while augmentative uses do not.”
Job displacement is already happening in places where routine tasks can be automated, West told me. Many layoffs have happened among software developers because that can be done fairly reliably by AI. “Most jobs will be affected by AI, but not every job is going to be replaced,” West said. “People should just look at the particular tasks associated with any job and just think about what are the possibilities of automation.”
AI’s effect on jobs will be decided by people, not potential
Most importantly, nobody knows what AI’s effect will be on the economy even a few years from now. ChatGPT only became a household name in 2022. The capabilities of these tools, and our understanding of what they can and can’t do, is constantly changing.
But the technology’s effect on jobs won’t necessarily happen because of what it can do. It’ll happen because of what business leaders and executives think it can do. At the moment, many executives seem to be more worried about missing out on opportunities to cut jobs and save money by using AI than they are worried that AI won’t be able to do the work. Klarna, for example, said in 2024 that its AI assistant could do the work of 700 customer service agents, but changed its mind earlier this year, hiring more humans after not getting the results it expected.
There’s already some doubt about the effect of corporate-directed AI initiatives. A July study by researchers at MIT found that 95% of AI pilots at businesses are getting no return on investment — largely because AI tools don’t learn, grow and develop like human employees do.
“Corporate leaders may end up laying off too many people because of their optimism about AI, and they may end up discovering that there’s an important element that’s missing,” West said. “The human judgment aspect is going to be critical.”
The human element — whether it’s judgment, creativity or culture — may prove to be what makes an AI tool unable to do a job, even if it looks like it might be able to do all the tasks on paper.
