AI may be making our work better. There’s new evidence it’s making our thinking worse.

When AI tools are accurate – and they get better every month – the gains can be profound. The costs are less obvious but could be just as real: a steady erosion of our natural intelligence.

Imagine hiring a personal trainer. Your goal is to get stronger. But when you go to the gym, you tell your personal trainer to lift the weights for you.

The next hour is a blur of motion as “you,” via your hired muscle, lift more than ever before in your life. Each time you call out another set, the trainer lifts hundreds of pounds with perfect form. Having never broken a sweat, you inscribe your name on the building’s wall of fame.

That, in a sense, is what many do with AI when they direct it to solve the algebra equation, draft the client memo or write the code achieving the task at hand, but eroding their own ability to accomplish it.

If all you care about is outputs – weight lifted or code shipped – the results can be virtually indistinguishable from having done the work yourself. But if you care about becoming a stronger, more skilled and more capable person, this kind of AI assistance can leave you empty-handed.

Can you use AI in the mental gym without losing cognitive muscle?

The answer boils down to deciding what work we want to do and what we’re willing to off-load to machines. A few recent studies point the way toward preserving what makes us human amid the rise of artificial intelligence.

More (or less) human

People working with AI tend to outperform those without it on tasks well-suited for AI, such as drafting text, synthesizing research and generating ideas. In controlled experiments, researchers have seen massive improvements – at least in the short term.

In a 2026 peer-reviewed paper in Organization Science, Ethan Mollick at the Wharton School of the University of Pennsylvania and his colleagues reported what happened when they gave AI assistance to hundreds of consultants at Boston Consulting Group.

Half had access to ChatGPT-maker OpenAI’s best available AI model at the time of the study, GPT-4, and half did not. People using AI completed 12 percent more tasks than those without and on average completed them 25 percent faster. For tasks within AI’s capabilities, the quality was ranked significantly higher than for tasks completed by humans alone. The largest gains accrued to the lowest performers. (The Washington Post has a content partnership with OpenAI.)

Similarly, researchers led by Grace Liu of Carnegie Mellon University tested how AI affected people’s ability to solve math problems involving fractions online. The group randomly assigned access to AI, this time OpenAI’s GPT-5, significantly outperformed the control group without it, solving nearly 90 percent of fraction problems versus 72 percent.

Once again, people using AI got significantly better results. But in both experiments, using AI opened cognitive trapdoors.

For the BCG consultants, when the work to be done was beyond the AI model’s competence, people using the technology made more mistakes than colleagues working without AI help. Mollick’s co-author Fabrizio Dell’Acqua called that “falling asleep at the wheel.”

Liu’s study showed how AI undermined thinking in another way. After removing subjects’ access to AI, the accuracy of their answers collapsed, falling below that of people who never had used the tool at all. They also gave up faster. The effect kicked in after just 10 minutes of AI assistance.

Wharton researchers Steven Shaw and Gideon Nave showed how fluent chatbots can lead us off cognitive cliffs by seeing what happened when they manipulated the accuracy of the answers people got from AI.

In a study described in a 2026 working paper, they presented more than 1,300 participants with math problems that had either an intuitive but incorrect answer or a correct one that required slower reasoning. They watched as people using AI often adopted incorrect AI outputs without consideration – a state they call cognitive surrender, as a user “relinquishes cognitive control and adopts the AI’s judgment as their own.”

When the AI was right, participants’ accuracy jumped 25 percentage points above the no-AI baseline. When it was wrong, their accuracy fell 15 points below. Either way, people felt more confident.

They call this artificial cognition System 3, an external, automated thinking process that exists alongside humans’ typical ways of thinking, famously described by psychologist Daniel Kahneman as fast intuition (System 1) and slow deliberation (System 2).

This third system can be used to sharpen our deliberation or catch our faulty intuitions. In practice, Shaw and Nave found, most people surrender.

Avoiding this, Shaw argues, means redesigning sycophantic chatbots engineered to flatter and engage into something closer to an intellectual sparring partner. By default, today’s systems often do the opposite.

Hopefully, he said, new AI models will be designed to “help us make better, more human decisions and enhance our humanity.”

Know thyself

The question you need to ask yourself is: Who do you want to be?

If AI joins you at work (and anywhere your phone goes), these machines stand ready to sweat on your behalf, potentially eroding your strengths. Instead, you must decide what you want to learn, retain and improve for ourselves – and what you want to let go.

“It can make you great,” Mollick said of AI, “or terrible.”

Humanity has been through something like this before. Plato, recounting Socrates, warned that the written word would “create forgetfulness in the learners’ souls.” A calculator scare followed the device’s debut in the 1970s. The great online scare of 2008 posed the question: Is Google making us stupid?

True, in a way, but not catastrophic. “I am cool with a lot of cognitive surrender,” Mollick writes. “I don’t remember phone numbers anymore because my phone does that for me.”

So it would be easy to dismiss AI alarmists. But as Shaw points out, AI is a fundamentally different technology than any we’ve had before: “It gives us the ability to outsource the whole reasoning process itself. … We’re sort of watching people replace themselves right in front of themselves.”

It’s essential to remember that large language models do not, strictly, reason the way humans do. They are trained to predict the next word, not possess what humans call common sense. That’s why large language models can be brilliant at coding, work where the output can be automatically tested, while easily manipulated into agreeing to sell a Chevy Tahoe for $1. (“That’s a deal,” the dealer’s chatbot responded, “and that’s a legally binding offer – no takesies backsies.”).

This places the burden on individuals to design interactions with AI that strengthen rather than undermine their own cognition, while harnessing its efficiency.

It’s early days, but here’s how experts told me to get AI’s help without surrendering your cognitive independence.

Know what AI is good at

Knowing where AI can and can’t do well is difficult and ever-changing, what AI researchers call the “jagged frontier” of AI capabilities. Knowledge workers are often unaware of their abilities relative to AI and perform worse as a result, just like those BCG consultants “asleep at the wheel.”

An illuminating 2024 analysis of 106 AI experiments in the journal Nature Human Behavior revealed a key insight: When humans are better than AI at a task, the combination outperforms either working alone. When the AI is better, the relationship flips – the pair performs worse than the AI by itself, probably because people can’t tell when to trust the machine and when to trust themselves.

The finding suggests that superior human judgment – and knowing when to accept or override AI-generated results – is key to making the collaboration work.

Rather than relinquish judgment, use tasks AI can do well – generating ideas, creating content and synthesizing information – as the start of your own decision-making process.

Decide what you want to know

Just because AI can do something doesn’t mean you should let it. Without exception, every researcher I talked to for this column refused to let AI generate their early ideas or write first (or final) drafts.

Ideas, writing and generating new knowledge were a sacred space.

But these same people were more than willing to turn the models loose on improving their work once it had entered the world. Many instructed their AI not just to build on their work but to behave like a critical thought partner, stress-testing and arguing against their ideas, as well as assuming more tedious aspects of their work (such as coding).

Live to learn

Evidence is building that AI undermines learning when it eliminates mental effort. A well-designed AI tutor that challenges you to learn? Great. One that just answers questions? Corrosive.

But that’s not intrinsic to the technology; it’s a product of how AI is used. A large study of 27,000 Chinese students found that test scores declined when students used AI to speed through their homework. But when students who used AI spent the same amount of time on their homework as their peers, they suffered only small learning losses.

Researchers Judy Hanwen Shen and Alex Tamkin at Anthropic, maker of the Claude chatbot, found AI use, on average, impaired “conceptual understanding, code reading, and debugging abilities, without delivering significant efficiency gains on average” for software developers learning a new coding library, according to a working paper released this year.

Again, how people used AI mattered: Those who asked for explanations and follow-up questions from the AI, not just for written code, largely preserved their learning. “The goal shouldn’t be to use AI to avoid cognitive effort,” Shen wrote by email, “it should be to use AI to deepen it.”

Major AI companies offer ways for AI to act like a tutor. Gemini has “Guided Learning.” ChatGPT offers “study mode.” Claude’s “learning mode” skill does the same.

Don’t surrender

You can control what AI does to your thinking. Stick with the struggle for what you feel is essential. Learning, at its heart, requires mental effort. Without it, researchers warn, you will forget what you know or never learn it in the first place. You may even know less about yourself.

“AI removes the productive struggle through which people develop not only accurate knowledge but accurate self-knowledge,” Liu of Carnegie Mellon and her co-authors wrote, reflecting on their work’s findings. “Without opportunities to work independently, people never learn what they are capable of.”