The AI mind tax that afflicts those who use AI best

A person seen from behind at a cluttered desk facing several screens and laptops at once, surrounded by sticky notes and notebooks, in warm muted light. — One person, a dozen tools, and nothing connecting them but her.

AI was supposed to make work easier, but for the people who use it most, it just became a different kind of work, and a whole lot of exhaustion.

The promise of the tech was that AI would make work easier, and, sure, in plenty of ways it came true, but for the people using it the hardest, it created a second job nobody considered, managing the pile of tools that now does the work for you.

The folks who stay sharp hold onto what the machine still can’t do for them, their judgment, their ability to work with other people, and some grip on where the work actually stands, but there’s a price to pay for that.

We’re not talking about just one burnt out person, because nobody was hired for this. The title still says marketer, or designer, or founder. The job shifted to keeping a dozen disconnected machines pointed in the same direction, and the whole industry is pretending that isn’t the work.

The people embracing AI the right way, learning the tools, wiring them into everything, getting genuinely good, are the ones getting buried, because being good at this now means being the one who holds it all together. A recent benchmark from CMO Huddles and Benchmarkit shows why. 98 percent of marketing teams use AI somewhere, and only 3 percent have connected it across their work. The companies didn’t build the connective layer, so the capable people became it.

In a research project I’ve been working with CMOs, I was chatting with a marketer I know who has been in the tech industry for almost twenty years. Her career has spanned some big tech names you would definitely know and these days she runs pretty much all the marketing at a startup by herself.

As we discussed the evolution of marketing and the marketing executive today, it became very clear that she’s about as deep into AI as anyone I’ve met. She is wired into every tool she touches, running most of her day through Claude and ChatGPT. She can spin up a user survey, analyze it, write the launch plan, and build the assets herself. All the stuff that used to take a whole team.

If AI was built to make anyone’s job easier, it was built for her.

I figured, if anything, that AI had simplified her life exponentially and that she’d admit as much. Instead she told me she feels like she’s “melting on the inside.”

The tools aren’t the problem. There are just so many of them, all living in different places, and she’s the only thing holding them together. Figma, the CRM, Claude, ChatGPT, Slack, each one its own window with its own half-finished conversation, and she’s the one keeping track of which idea she had where.

She called it a real tax and toll. She also said she wouldn’t give a single one of these tools back.

And that job is exhausting.

She was not an outlier. One fractional CMO I talked to runs marketing for a whole stack of startups at once, each with its own set of tools, its own logins, its own AI baked into the CRM and the design software and everything else, and she is the only human carrying context from one to the next.

Another person I spoke to runs a team at a big company where everyone has quietly built their own personal pile of AI agents, each one tuned to how that person works, so when someone leaves, the knowledge of how any of it fits together walks out with them.

Different companies, different budgets, same situations. The person in the middle is the integration layer, and the integration layer is tired.

Diagram of one person at the center linked by lines to five separate tools, Figma, a CRM, Claude, ChatGPT, and Slack, with no lines connecting the tools to each other. — Every tool connects to the person, never to each other. The human is the only shared context in the whole system.

The fatigue is real

What she described as “melting” is a documented, measurable thing. For better or worse, the research on it was pretty disturbingly long before any of us used a chatbot.

Gloria Mark, Daniela Gudith, and Ulrich Klocke found that it takes about twenty-three minutes to fully get your focus back after a single interruption.

Similarly, Rohan Narayana Murty, Sandeep Dadlani and Rajath B. Das, studying workers across three Fortune 500 companies, found that people toggle between different applications and websites around twelve hundred times a day.

Cornell’s Ellis Idea Lab, working with Qatalog, found it takes nine and a half minutes just to get back into a workflow after switching apps, that forty-five percent of workers say too many apps make them less productive, and that nearly half admit they make actual mistakes because they can’t keep track of what’s happening across all their tools.

That was the cost of switching between dumb apps that just sat there and waited for you. Then every one of those apps grew its own AI.

As the switching multiplied, every tool became two things to keep track of, the tool and the AI living inside it. The marketer I spoke to isn’t just bouncing between Figma and the CRM and Slack anymore. She’s bouncing between Figma’s AI and the CRM’s AI and Claude and ChatGPT, each one a separate conversation, each one knowing nothing about the others, each one needing to be caught up.

The one thing none of these tools reliably does is hold the thread for you.

Rifat Mehreen Amin and her colleagues, studying how people actually use AI chatbots, gave the problem a name, context fragmentation, and watched people cope with it by copying pieces of one conversation into another by hand, stitching the results back together across sessions, rebuilding their own mental map every time they came back.

You become the only part of the system that remembers what you were doing in the other four windows, which is why, at the end of a day like that, you feel like you’re melting on the inside. It also means you are the only one who knows where any of it stands.

When all the context lives in your head, explaining to your manager or your team where the work actually is becomes its own task, and nobody else can pick it up without you.

An overhead photo of many different-colored cables tangling together and converging into a single central point on a plain surface. — The human user becomes the sole integration layer between disconnected AI systems.

You turned into the glue

The reason all of this lands on her and not on the machines is simple. The machines don’t talk to each other, so she has to. Every tool she uses has become its own little walled garden with its own AI inside it.

Figma’s AI knows about the file she’s in and nothing else. The CRM’s AI knows about campaigns and nothing else. Claude knows about the conversation she’s having with Claude. None of them know about any of the others.

Connectors are starting to change that at the edges. You can wire Figma or a CRM into Claude so it can reach into them directly. Xinyi Hou and her colleagues, in a survey of the connector landscape, found that even with a shared standard for these integrations, stitching tools together remains fragmented, and the connections mostly run one way, into whichever assistant she has made the hub. The other tools still sit in their own corners.

She is still the one deciding what connects to what, and carrying across by hand whatever doesn’t.

Jan-Philipp Steghöfer studied how professionals actually use AI tools across their real workflows and found that the most persistent frustration was almost never the AI’s ability. It was the breaks between the tools, each one useful on its own, none able to hand off to the next, so the person in the middle rebuilds the context from scratch every time and carries it across by hand. One worker in his study described finishing in one AI tool and manually moving the output into the next just to keep going.

The marketer I talked to does the same thing, having ChatGPT write a summary of one conversation so she can paste it into Claude to start another. She has become the connective tissue between systems that were supposed to connect themselves.

The cruel part is that all of that burden on your shoulders feels productive. It feels like work getting done.

The research keeps finding that the feeling and the reality don’t match.

The cleanest proof comes from METR, a nonprofit that ran a careful randomized controlled trial on experienced developers doing real work in codebases they knew inside out. Going in, the developers expected AI to speed them up by around a quarter. They did the work, and afterward they still believed it had made them about twenty percent faster.

Bar chart from the METR study showing developers expected to be twenty-four percent faster with AI and felt twenty percent faster, but were measured nineteen percent slower. — Developers felt about twenty percent faster with AI. They were measured nineteen percent slower.

When the researchers measured the actual time, the developers had been nineteen percent slower with the AI than without it. The gap between how fast they felt and how fast they were ran close to forty points, and it held even after they’d lived through the slowdown themselves.

That gap is the whole danger of the mind tax. You cannot feel it while it is happening. You feel busy, you feel capable, you feel like you are getting an enormous amount done, and some of that is true.

Aruna Ranganathan and Xingqi Maggie Ye found the same thing over eight months of watching people adopt AI at a tech company. The tools intensified the work instead of shrinking it. People took on more, moved faster, and stretched the day longer, largely because AI made doing more feel possible.

Instead of the rest or deeper thinking you’d expect, the time it saved resulted in more work, more tools to run, and more threads to manage.

A bar chart comparing how often a person versus an AI affirms what you are doing, with the AI bar forty-nine percent higher than the human one. Source, Cheng, Jurafsky et al., Science, 2026. — Across eleven leading AI models, the machine affirmed what the user was doing about forty-nine percent more often than a person would.

It comes for your best people first

The cruel joke of the mind tax is who it hits hardest. You would think it’s the people just dabbling with AI who get worn down the most, but the reality is that it’s the people who dove in head first.

Julie Bedard, Matthew Kropp, Megan Hsu and their colleagues at Boston Consulting Group studied nearly fifteen hundred workers and found that productivity climbs as you add AI tools, right up until about the third one. After that it falls. The workers running four or more tools at once were getting less done, making more errors, and reporting more of the exact buzzing, fried feeling the marketer described. The fatigue came from overseeing the AI, not from handing work to it.

All the watching and checking and correcting, and the holding of everything in your head at once, is its own kind of labor, and the people doing the most of it were the heaviest users, the ones who leaned in hardest.

The study found the highest rates were in marketing, which is to say, in people exactly like my CMO friend.

The management version of this is worse than the productivity one. Those same researchers found that workers who weren’t fried intended to leave their jobs at a rate of about one in four. Among the workers who were fried, it rose to one in three.

The Upwork Research Institute found the same thing from another angle, that the most productive AI users, the ones getting the biggest gains, were also the most burned out, with 88 percent of them reporting burnout and roughly twice the likelihood of looking for the exit.

This mental tax seriously wears out your best people and then pushes them out the door. The problem is, of course, that these are the exact people you cannot afford to lose.

What should worry anyone who does this work for a living is the impact to a person’s mind underneath it all. Nataliya Kosmyna and her team at the MIT Media Lab wired people up and watched their brains while they wrote, some on their own, some leaning on AI. The heavy AI users showed weaker neural engagement, remembered less of what they’d produced, and felt less ownership of it.

The researchers called what builds up “cognitive debt,” the brain quietly doing less of its own work once a machine is there to do it.

Michael Gerlich found a matching pattern in a study of six hundred and sixty-six people. The more someone relied on AI, the weaker their critical thinking scored, and the thing connecting the two was cognitive offloading, the habit of handing your thinking to the tool. The effect was steepest in the heaviest and youngest users. Gerlich is careful that this is a correlation and not proof that AI caused the decline, and that caution is fair. The direction of it still lines up with what everyone living in these tools already feels.

Put the pieces together and the mind tax stops looking like plain tiredness. The thing getting taxed is judgment.

The more of the thinking you hand off, the less of it you are doing, and the less sharp you get at the one part of the job a machine still cannot do for you.

That is a genuinely terrible tradeoff. When it comes down to it, your judgment is the only thing you were ever really being paid for.

The new job is quietly eating the old one, and nobody has said so out loud.

A two-column comparison. Left, “What the title still says”: Marketer, Designer, Founder. Right, “What the job actually became”: Orchestration, Judgment, Holding the team together, Being the integration layer. — The title stayed the same. The actual work changed underneath it, and nobody renamed it.

A room with one voice in it

When I asked that marketer what she missed most, she avoided any mention of the AI tools in her arsenal, and focused entirely on teamwork and collaboration. The whiteboard, the ten people around a table, the coworker who could absorb her idea and say, in so many words, that it was wrong.

What she has now is a one-way screen. She does the work alone, runs it past an assistant that agrees with her, and then launches it.

The collaborator got replaced by something that never pushes back.

Myra Cheng, Dan Jurafsky and their colleagues at Stanford tested eleven leading AI models and found they affirmed what the user was doing about forty-nine percent more often than a human would, including in cases where the user was describing something deceptive or harmful. Worse, when people took interpersonal problems to a sycophantic model, a single conversation left them more convinced they were right and less willing to repair the situation and people liked it. They trusted the agreeable model more, which is exactly why companies have every reason to keep building it that way.

This is baked into how these systems are made. Anthropic’s own researchers, led by Mrinank Sharma and Ethan Perez, showed that the leading assistants are systematically sycophantic, because the training rewards it.

People reward the answer that flatters them over the one that corrects them, every time. Train a machine on that mentality and you get a world-class suck-up.

Even the humans overseeing the training, and the models built to imitate those humans, preferred the convincing wrong answer to the accurate inconvenient one a meaningful share of the time.

This dismantles what actually made the work good. Jan-Philipp Steghöfer’s study found that the single-user model of these tools substitutes a private conversation with the AI for the collaborative one with the team, the back-and-forth where people surface each other’s blind spots, challenge the weak idea, and build a shared understanding of why the thing is the way it is.

The team inherits the finished artifact but not the argument that should have shaped it. Our marketer friend felt the same loss. She said that when a team spends half an hour in a room on a problem, you get ten different minds pulling at it. With AI, you get one, and it’s pulling in whatever direction you already were.

The thing that used to make people good at this work was friction, the colleague who caught the mistake, the review that stung, the question you hadn’t thought to ask. Trade all of that for a tool that agrees with you, and the cost isn’t only time and focus.

The real loss is what was keeping your judgment honest.

Somebody has to pay for it

I want to be careful here, because it would be easy to think I’m making a case against AI and I’m not. The marketer I talked to would be the first to tell you that. She loves these tools. She said, more than once, that she wouldn’t give a single one of them back, that the amount she can do now would have taken a whole team the old way.

The gains are real. The tax is just the price nobody understood or identified when she started down this path.

Let’s be clear, it is a real price.

Add up everything we’ve walked through and you get the same result in three different outputs. There’s the work of running the tools, the context carried by hand from one window to the next. There’s the slow erosion of the thinking you stop doing because the machine is there to do it. There’s the quiet of working alone with something that only ever agrees with you.

This all comes down to one person standing in the middle of all the tools, holding it all together.

This is really a design problem.

Right now the human is the integration layer, the memory, the quality check, the only thing in the whole setup capable of saying “no, that’s wrong.”

We built a dozen brilliant tools and skipped the part that connects them, carries the thread, or pushes back. The work John Dubois, Chandra Shukla, and Kay Youngstrom from EY have done on redesigning jobs around AI makes the case that the win comes from augmentation, designing work so the tools sharpen human judgment instead of burying it. That is a thing to build, not a thing a person should have to white-knuckle their way through alone.

We need to embrace the idea that the job changed. Being good at this work, today, means being good at orchestration, at judgment, at holding a team together while the machines do the production. Create a name for that job, hire for it, build the day around it, and the tax gets smaller. Pretend the job is still what it used to be, and it keeps coming out of the person.

The reason to build it is that the one job that we could truly own, and still can, is judgment.

Knowing what’s good, knowing what’s missing, knowing when the confident answer is confidently wrong.

Peter Klein calls this the paradox of automation. The more AI takes over the routine work, the more valuable human judgment becomes, and at the same time the fewer chances people get to build it, because judgment is grown through exactly the hands-on work the machine is now doing.

Tomas Chamorro-Premuzic puts the same idea more bluntly. When machines do the thinking, humans have to do more of the judging. The mind tax is dangerous because it eats that capacity first. It buries the one thing you cannot afford to lose under the management of everything else.

So our marketer friend keeps going. She does the work of ten people, and then she does the work of managing the machines that do the work of ten people, and at the end of the day she feels like she’s melting on the inside.

The tools keep getting upgraded.

She’s still running on the same single brain she started with. Until someone designs for that, instead of around it, the tax comes out of her.

The tax will always come out of the person.