In 1950, Alan Turing famously created what’s now known as the Turing Test, a way of deciding whether a computer is intelligent. If the computer could converse so fluently that it passed as a human? Presto: That’s artificial intelligence.

Turing’s test became the north star for generations of AI pioneers. For decades, they’ve labored mightily to mimic basic human skills, with wild success: We’ve now got AI that can hold conversations, draw pictures, or play expert rounds of chess, Go, and fast-paced video games.

But now some AI thinkers wonder whether we’ve succeeded a little too well—at the wrong task. Mimicking human abilities, they believe, has led to direct economic competition between people and machines. Maybe Turing led us astray and copying humans was the wrong goal.

That’s the argument of Erik Brynjolfsson, who is the director of Stanford’s Digital Economy Lab and has long written about AI’s effect on labor. In a recent paper, he argued that Turing-chasing scientists have driven up wage inequalities.

As Brynjolfsson sees it, AI creators have automated humans out of many forms of work—from obvious jobs such as bank tellers to less-visible roles, like inventory managers and legal researchers. Automation has made companies more productive, but the productivity gains go to the owners of firms, not to workers. This dynamic, Brynjolfsson argues, is “the single biggest explanation” for why wages have mostly stagnated over the past few decades while the ranks of millionaires and billionaires have bulged.

He calls it “the Turing Trap.” It’s certainly true that human-like AI is on a roll: Behold the rise of uncannily deft visual-art generators such as Dall-E and Midjourney. This summer, a game designer entered a Midjourney creation in the Colorado State Fair art contest and won first prize, apparently without any of the judges suspecting the work had been done by a computer—nailing, as it were, an aesthetic Turing test.

Brynjolfsson gets why AI creators have been so enchanted with mimicking human abilities. It caters to a desire to play god, creating life forms in our own image. “Every culture has a myth about this,” Brynjolfsson says. Ancient Greeks told stories of the inventor Daedalus producing mechanisms that walked like men, Jewish folklore had the golem, and real-life inventors have been crafting humanlike automata from early Islam to Renaissance Europe. Modern sci-fi is simply littered with AI that walks and talks like humans.

But mythology may not be the best framework for software development.

Brynjolfsson thinks real economic growth lies in building AI that augments humans: It should do things people can’t.

“We need to change the target,” he says. Consider AlphaFold, DeepMind’s AI for predicting protein structures. Predicting the structure of proteins requires manipulating chains of amino acids in millions of possible combinations—which isn’t something humans can readily do. But by using AlphaFold, scientists could potentially become super-scientists, able to explore far more possibilities for drugs and medical treatments than they could on their own. When I spoke to DeepMind CEO Demis Hassabis last winter, he argued, much like Brynjolfsson, that augmentation was the promising way forward. “What I’m hoping for is AI as this sort of ultimate tool that’s helping science experts,” he said. He anticipates “a huge flourishing in the next decade,” and says that “we will start seeing Nobel-Prize-winning-level challenges in science being knocked down one after the other.”

Instead of merely saving costs by replacing humans with a bot, Brynjolfsson notes, augmentation increases people’s productivity. Better yet, some of the economic value of that productivity would accrue to workers because their augmented labor would become more valuable. It wouldn’t all be hoovered up by the billionaire owners of the tech.

The catch is that augmentation is hard. When you’re simply mimicking human behavior, you know (more or less) whether you’ve nailed it. (The computer can play checkers: success!) But inventing a form of AI that’s usefully different from the way humans operate requires more imagination. You have to think about how to create silicon superpowers that fit hand-in-glove with the abilities unique to people—such as our fuzzy, “aha” intuition; our common-sense reasoning; and our ability to deal creatively with rare, edge cases.

“It’s 100 times easier to look at something existing and think, ‘OK, can we substitute a machine or a human there?’ The really hard thing is, ‘let’s imagine something that never existed before,’” Brynjolfsson says. “But ultimately that second way is where most of the value comes from.”

At the Stanford Institute for Human-Centered AI, director Fei-Fei Li wanted to know what people actually wish to have automated. Her group went to the US government’s “American Time Use Survey,” which chronicles people’s daily tasks. Li’s team picked 2,000 everyday activities that could viably be done by AI and robots, then asked people to rate how much they wanted that task automated, “with zero being hell no, I don’t want robots to do this, and the maximum being please, I’m dying to have a robot do this,” Li says.

“Open a Christmas present for me” was zero; “cleaning the toilet” was high. Obvious enough, but there was more complex stuff in the middle, such as “recommending a book.” The only way to find out what people want, Li notes, is by asking them—not by barging ahead and designing AI based on sci-fi fantasies.

Here’s another wrinkle: It’s not always obvious how the two kinds of AI are different.

One could argue that DALL-E and other image generators are a pure Turing play because they replicate the human ability to create art. The internet currently groans under the weight of essays claiming human artists are about to be serially unemployed by AI. But creators can also use the apps to punch above their weight, such as when a video game designer used Midjourney to generate art for a space shooter. That looks a lot like augmentation.

What’s more, many jobs are harder to entirely automate than you might think. In 2016, deep-learning pioneer Geoff Hinton argued that we should stop training radiologists because “it’s just completely obvious that within five years, deep learning is going to do better than radiologists.” (He added that it might take 10 years.) But there are still tons of radiologists employed, and there probably will be in the future because the job of a radiologist is more complicated than Hinton suggests, as noted by Andrew McAfee, a colleague and coauthor of Brynjolfsson’s who codirects the MIT Initiative on the Digital Economy. AI might be better at noticing potential tumors on scans, but that’s only one small part of a radiologist’s job. The rest of it includes preparing treatment plans and interacting with scared patients. Tumor-spotting AIs, then, might be better seen as augmenting those doctors.

To nudge companies away from Turingism, Brynjolfsson suggests some changes to government policy. One area ripe for reform is the US tax code. Right now, it taxes labor more harshly than capital, as recent work by the Brookings Institute found. Companies get better tax treatment when they buy robots or software to replace humans because of write-offs such as capital depreciation. So the tax code essentially encourages firms to automate workers off the payroll, rather than keeping them and augmenting them.

“We subsidize capital and we tax labor,” Brynjolfsson says. “So right now we’re pushing entrepreneurs—whether they want to or not—to try to figure out ways to replace human labor. If we flip that around, or even just level the playing field, then entrepreneurs would figure out a better way.” That might be one way out of the trap.