By Christopher Mims
"USA250: The Story of the World's Greatest Economy" is a yearlong WSJ series examining America's first 250 years. Read more about it from Editor in Chief Emma Tucker.
On July 6, 1962, the Energy Department set off a hydrogen bomb 75 miles southwest of Las Vegas, in the hopes of proving that nuclear weapons could be used for peaceful purposes, such as earthmoving. The blast sent fallout as far away as South Dakota, and exposed more than 13 million Americans to radiation.
Project Plowshares, the program that begot this test, bears more than a passing resemblance to today's efforts to apply ever more powerful artificial-intelligence models to every area of human endeavor. Both are examples of the misbegotten belief that just because humans can do something, they ought to at least try.
There is even the uncomfortable parallel between the misgivings of those who built the bomb and those who are developing today's AI models, who worry that the monster they are unleashing is so dangerous that it risks the future of humanity.
Overuse of today's AI models isn't going to lead to a massive plume of radioactive fallout, but it could have a number of far-reaching consequences that will seem obvious in retrospect. These include mass layoffs, angry and alienated customers, and the destruction of a huge amount of value at companies that are misapplying it.
I haven't arrived at these conclusions lightly. For the past decade, I've been collecting insights and case studies from those who build AI, and those on the front lines figuring out how to use it to do their jobs. I've talked with the CEOs and lead engineers at every leading AI lab, the Nobel Prize- and Turing Award-winning computer scientists on whose shoulders they stand, and dozens, if not hundreds of everyday folks creatively applying the resulting tools in their day-to-day lives.
I distilled all I learned into my most recent book, "How to AI," focusing on how to use AI to your advantage. But I left out something important: When not to use AI.
As I give talks and take questions from readers, I've come to realize that those who are the most skilled at using AI are equally knowledgeable about where they should keep hands off. Were they civil engineers, they'd be the sort to say that using an overwhelmingly powerful tool in the wrong circumstances, such as a hydrogen bomb in the course of digging a canal, is a bad idea.
So when should you not use AI? Based on my conversations, it's when one of the following is core to what you're looking for:
Empathy
Dan Leiva is a former executive at Apple and eBay, where he pioneered the use of AI to deliver customer service, and to help call-center workers do their jobs. He says the world is full of things that AI can do -- but shouldn't. One of the most obvious is handling any customer interaction that requires empathy.
When someone calls in seeking to cancel an account, for example, it's a cinch for AI to handle that. "But let's say it's a widow calling up to cancel her husband's account," says Leiva. "Do you really want AI handling that interaction?"
Natalie Desseyn is a double-board-certified nurse practitioner in family and psychiatric medicine. She uses a half dozen different AI tools, for everything from taking notes during sessions to filing claims with insurance companies. Taken together, she says they save her 15 to 20 hours a week, allowing her to handle 300 patients as a solo practitioner. And yet the most important thing these tools do for her, she insists, is make it possible for her to be fully present -- eyes up, reading the faces of her patients as they pour out their distress -- during psychiatric counseling sessions.
"In psychiatry, if you're not engaging with the patient, you're going to miss a lot of symptoms," says Desseyn. "I have a couple of patients that I'll ask, 'Are you thinking about hurting yourself?' And if I'm not looking at them, I'm not going to get the right answer." That level of presence is possible now -- and wasn't in the past -- because she knows her AI transcription and note-taking tool is capturing every nuance of the interaction, and her insurance-filing tool will digest it into claims.
Whether it's customer service or counseling, it's where the AI is absent -- and the ways in which it makes it possible for humans to be present in that empty space -- that counts the most.
Authenticity
When I give talks to people in small and midsize businesses, one of the easiest ways to get them to sit up and pay attention is to tell them they shouldn't be using AI to write their marketing copy.
This flies in the face of what they've been told about the easiest wins in applying AI to their business. The problem isn't that AI isn't a facile writer -- it's that in a world in which everyone can use it to churn out copy, what matters isn't just showing up on social media and on the corporate blog, but authenticity.
Big companies know this. Ironically, it's the firms most likely to be building AI that are now desperately seeking human "storytellers" to break through the endless AI-generated slop on LinkedIn and elsewhere, with real stories written by actual people. This also helps explain the rise of a new class of influencers, who reach audiences with direct-address, short-form video, which is much harder to fabricate with AI.
Transparency
When he was head of payment operations at eBay, one of Leiva's biggest concerns about incorporating generative AI into the company's systems was whether it would be possible, after the fact, to explain the decisions it had made.
This should be a concern to anyone using generative AI in a highly regulated field, whether that's payments, medicine or the law, he adds. While engineers have spent decades trying to create "explainable AI," a defining characteristic of today's large language models is that they remain biased black boxes that are capricious in their decision-making.
When transparency about how they reached a conclusion isn't possible, AI systems need to be able to at least show their work, so that humans can verify it after the fact. Mike Walsh is CEO of LexisNexis, the legal research company that has been offering computerized legal research services since 1973. Recently, he told me that generative AI means his customers are now spending more time using his service to draft legal documents than to do case-law research. The reason they can do so with confidence is that, unlike generic chatbots which are infamous for hallucinating nonexistent cases, his company's system is only allowed to cite cases that it can verify actually exist in the company's databases.
Lawyers who use such systems have emphasized to me that they still have to review what AI generates -- every time. In a world in which liability falls on the shoulders of whoever used an AI, rather than its creators, there is simply no other way.
More isn't always better
There is a vast gulf between how much productivity CEOs believe AI will yield and how much employees report it actually helps. There are as many reasons for this as there are areas where AI is useful. One challenge is the huge variability in the applicability of AI: While some tasks can be handed over to AI almost entirely, most cannot.
Companies that are too quick to lay off workers on the assumption that AI can do their jobs risk wrecking their future competitiveness in two ways. The first is that they can lose critical institutional knowledge. The second is that they risk hurting their own talent pipelines. While it may be tempting to replace junior engineers with AI, doing so means that when senior engineers move on, a company will no longer have the humans required to review the work of those AIs.
This is already playing out at a handful of companies that proclaimed a pivot to AI and later had to either rehire employees or increase head count. Skeptics believe that recent layoffs at tech companies are cover for labor trends that have nothing to do with AI, such as efforts to reduce head count after a period of over-hiring, or a push to switch from salaried to contract labor. Another reason such layoffs may prove to be temporary: In many cases AI seems to shift the burden of work to those who are most skilled at using it, rather than lightening the load overall.
While it's already apparent which companies are falling behind because they've failed to embrace AI, in just a few years it will become apparent which companies are falling behind because they leaned too far into AI, says Leiva. "We will see who over-automated, and which ones lost the knowledge that lets them differentiate themselves from other companies," he adds.
In the end, Project Plowshares had the opposite of its intended effect. Instead of demonstrating new uses for nuclear bombs, it showed when they shouldn't be used at all. In a future in which AI is ubiquitous and taken for granted, today's efforts to throw it at every problem on earth may lead to the same outcome.
(END) Dow Jones Newswires
June 21, 2026 11:00 ET (15:00 GMT)
Copyright (c) 2026 Dow Jones & Company, Inc.
Comments