how to tackle AI brain fry

A few weeks ago I hit a wall. I had been using AI tools every single day, across every part of my work. I was producing more than I ever had. And I was more tired than I had been in years.

The instinct was to blame myself. Maybe I was not disciplined enough. Maybe I was using the wrong tools. Maybe I just needed to figure out the right prompts.

Then I went digging into the research. And what I found completely reframed everything I thought I understood about AI and productivity.

This is not an article against AI. It is an article about something the AI productivity conversation has been systematically avoiding: the human cost of doing more, faster, without a plan for what that actually means.

Part 01

What the Research Actually Found

UC Berkeley’s Haas School of Business researchers Aruna Ranganathan and Xingqi Maggie Ye spent eight months embedded inside a 200-person American tech company. They observed in person twice a week, tracked internal communications, and conducted over 40 in-depth interviews across engineering, product, design, and operations.

Their conclusion, published in Harvard Business Review in February 2026, was direct: AI tools did not reduce work. They consistently intensified it.

The researchers identified three specific mechanisms:

1. Task Expansion

AI lowered the barrier to entering specialized work. Product managers started writing code. Researchers handled engineering tickets. This felt like empowerment. The hidden cost: engineers then spent significant untracked time reviewing and correcting their colleagues’ AI-assisted output. Nobody’s calendar showed this. Nobody’s metrics counted it.

2. Blurred Boundaries

Prompting an AI feels like chatting, not working. So workers filled lunch breaks, meeting gaps, and mornings with “one last quick prompt.” The natural pauses that allow the brain to recover simply stopped happening.

3. Relentless Multitasking

Multiple AI threads running simultaneously created a feeling of momentum. It also created constant context-switching and cognitive fragmentation that the brain is not designed to sustain at scale.

The cycle this created is worth understanding in full:

Brain Fry: A New Clinical Category

A separate study from Boston Consulting Group and UC Riverside surveyed 1,488 full-time US workers and introduced a term that I think is going to define the next decade of workplace conversation: “brain fry.” They define it as mental fatigue from excessive use or oversight of AI tools beyond one’s cognitive capacity.

This is not traditional burnout, which builds slowly over months. Brain fry is acute. It can hit within a single intense workday. And the numbers around it are alarming.

The one positive data point in all of this: when AI replaced genuinely routine and repetitive tasks, burnout decreased by 15%. The problem is not AI.

The problem is AI applied without intention to cognitive work that was never designed to be accelerated.

The CEO-Employee Perception Gap

The Upwork Global Study revealed something that I think is the most structurally important finding in this entire body of research.

Same technology. Completely opposite realities. CEOs are measuring output metrics: volume, speed, headcount ratios. Employees are living the input metrics: the cognitive effort, the expanded task scope, the invisible supervisory labor of reviewing and validating AI outputs. Both are accurate. They are measuring different things entirely.

ActivTrak’s 2026 State of the Workplace data, drawn from 96,000 workers, makes this concrete. Since widespread AI adoption, email volume increased 104% and messaging volume increased 145%. Yet deep focus time - the sustained, uninterrupted concentration that produces actual thinking - dropped 9%. More communication. More coordination. Less actual work.

The Cognitive Debt You Are Quietly Accumulating

MIT Media Lab’s research introduced a concept that reframes the entire productivity conversation: cognitive debt. Their study found that higher confidence in AI output was directly associated with less critical thinking. Workers who trusted AI more thought less independently.

But here is the finding that is almost never discussed: evaluating, verifying, and integrating AI outputs actually requires more cognitive effort than doing the task independently in the first place.

AI offloads mechanical execution while concentrating cognitive demand at the oversight layer. For knowledge workers, whose value is judgment and critical thinking, this is not relief. It is a redistribution of effort toward the most expensive cognitive work, with the added stress of accountability for outputs you did not fully generate.

Microsoft Research’s 2025 paper on AI and critical thinking confirmed this directly. Workers described the shift not as lighter work but as different, and more mentally taxing, work.

Part 02

Why We Cannot See This Clearly

The California Management Review published what is probably the most methodologically rigorous review of AI productivity assumptions to date, covering 37 software development studies, meta-analyses of 106 collaboration experiments, 28 creativity studies involving 8,214 participants, and 371 estimates of AI’s labor-market outcomes.

It dismantled several assumptions most of us have been operating on.

The paper’s conclusion is precise: AI’s productivity dividend is real in specific contexts, for specific users, under specific workflow designs. But it is far from automatic or universal.

Deloitte’s 2025 Global Human Capital Trends report reinforced this from a different angle: AI is creating silent, unintended impacts. Bburnout, loneliness, and increased workloads. As workers increasingly collaborate with AI systems. The mechanism is precise. AI handles repetitive tasks but simultaneously raises the cognitive ceiling of every remaining task. Workers no longer toggle between easy and hard. They operate in perpetual high complexity mode.

We Have Seen This Pattern Before

NBER identified that the current AI productivity paradox directly mirrors the IT productivity paradox of the 1980s, when personal computers arrived in every office and productivity statistics refused to move for nearly a decade. Massive investment, widespread adoption, stagnant macro-level outcomes.

Their argument: implementation lags are the dominant explanation. The technology has arrived. The organizations, workflows, norms, and metrics required to extract genuine value from it have not.

We Are Measuring the Wrong Things

UC Berkeley Executive Education made an important argument: MIT found that 95% of generative AI pilots delivered zero measurable ROI despite $30-40 billion in investment. But this is not an AI failure. It is a measurement failure. Organizations are applying industrial-era, quarterly ROI metrics to a transformational technology that does not produce immediate balance-sheet returns.

Gartner reinforced this with a finding that stopped me:

Part 03

The Psychology Behind It

The term “cognitive overload” has become a buzzword. The peer-reviewed research behind it is more specific and more alarming than the buzzword implies. Frontiers in Psychology (2025) validated a six-dimension clinical structure for what they call digital burnout in the AI era.

This is not a list of inconveniences. It is a clinical taxonomy of a psychological condition that did not exist at scale five years ago.

The Distinction Nobody Is Teaching You

PMC research introduced a concept I have found essential for thinking about AI honestly: the difference between scaffolding and substitution.

The uncomfortable truth: the short-term experience of both feels identical. You cannot tell in the moment whether AI is building your capability or eroding it. The divergence only becomes visible over months and years, when you try to do the work without the tool and find that you cannot.

Research by Macnamara et al., 2024 - Cognitive Research: Principles and Implications makes this even more unsettling: expert-level skill decay from AI scaffolding is invisible to the performer. People feel equally confident even as their unassisted performance degrades. You do not know what you have lost until the tool is removed.

Rinta-Kahila, Penttinen et al. (2023) at Aalto University documented what this looks like at organizational scale. They studied an accounting firm after automation. When the automation system was removed for maintenance, employees could no longer perform the underlying accounting tasks manually. The skill had not partially atrophied. It had disappeared entirely. The organization had optimized itself into fragility.

Part 04

What Actually Helps: An Honest Framework

Everything above leads to the question I kept asking as I worked through this research: so what do you actually do about it? Here is what the evidence points to.

Before You Adopt Any AI Tool, Ask Three Questions

The Two-Week Signal

After two weeks of using any AI tool, ask one honest question: has the volume of that category of work increased since I started? If yes, the tool expanded your baseline. It did not give you time back.

Build the Practice, Not Just the Stack

The UC Berkeley researchers are explicit on this: the problem is not the technology. It is the absence of norms around how it is used. Here are the five organizational interventions the research collectively points to: