Cognitive load in the age of AI: the paradox that is exhausting your teams
Generative AI promised to free us from tedious work. Two years after its widespread deployment in businesses, studies converge on a disturbing finding: when poorly integrated, it intensifies work instead of simplifying it. An analysis of a phenomenon still largely invisible in organizations—and concrete proposals for regaining control.
The paradox: more AI, more mental fatigue
On paper, the equation was simple: AI automates repetitive tasks, freeing up employees for high-value thinking. The reality observed on the ground tells a different story.
An eight-month ethnographic study conducted by Aruna Ranganathan (Haas School of Business, Berkeley) and Xingqi Maggie Ye at a technology company with approximately 200 employees reaches an unambiguous conclusion: AI tools have not reduced work, they have increased it. constantly intensified. Employees launch AI productions "in the background" while they process other files, multiply the back-and-forth between heterogeneous tasks, and pay the price of "attentional residue" with each switch: part of the cognition remains attached to the previous task, degrading the processing quality of the next.
In early 2026, the Boston Consulting Group put a name to the symptom: the "AI brain fry"In a survey of nearly 1,500 full-time employees across all sectors, researchers found that intensively monitoring AI tools was the most mentally taxing form of engagement. Workers tasked with closely monitoring AI agents report +14% mental effort, +12% mental fatigue and +19% information overloadThe consequences are both business and human: people in a state of "brain fry" commit 39% more major errors and are 39% more likely to actively seek to leave their job.
Benchmarking: what major studies say
Four research corpora, published between mid-2025 and mid-2026, paint the same picture from different angles:
1. The “workslop” — Stanford Social Media Lab & BetterUp Labs (2025-2026). This portmanteau word ("work" + "slop") refers to AI-generated content that appears polished but lacks substance. A survey of 1,150 American employees revealed that approximately 40% received such content in the past month, with each incident requiring an average of [amount missing]. 1 hour 56 minutes of corrective work at the recipient's address. Estimated cost: $186 per employee per monthThat's more than $9 million per year for an organization of 10,000 people. The relationship cost is just as concerning: about half of the recipients judge the sender to be less reliable, less creative, and less competent than before.
2. The phantom return on investment — MIT (2025). The MIT Media Lab report from July 2025 estimated that 95% of organizations saw no measurable return on their investments in generative AI, despite billions spent. The workload and verification overhead explain a good part of this evaporation of gains.
3. “Cognitive debt” — MIT Media Lab (2025). An electroencephalography study on essay writers shows that passive use of AI (delegating without prior effort) reduces cognitive engagement and can ultimately weaken the skills that are no longer being used. An important distinction: introducing AI after An initial independent effort produces much better results. The problem is not the tool, it's the sequence of use.
4. Critical thinking under pressure — Microsoft Research (2025). Among knowledge workers, the higher the trust in AI, the less critical thinking effort is reported—while confidence in one's own abilities is associated with greater vigilance. The training challenge is therefore twofold: mastering the tool AND strengthening human judgment.
The five mechanisms of overload
Behind these figures lie well-identified cognitive mechanisms:
- The cost of production decreases, the cost of processing remains human. Generating a report costs nothing; reading it, checking it, arbitrating it, does. Every individual time saved often translates into a collective overload: the time "saved" is immediately reabsorbed by reading the work of "augmented" colleagues.
- The invisible verification work. Reviewing, correcting, contextualizing, and validating AI outputs is not listed in any job description. Yet this "phantom supervision" consumes the most demanding cognitive functions: judgment, selective attention, and ultimate responsibility.
- Multitasking multiplied. AI makes it technically possible to manage several projects in parallel. The human brain, however, hasn't changed: each interruption leaves an attentional residue that accumulates throughout the day.
- The expanded sphere of responsibility. BCG researchers observe that AI has not reduced the workload: it has expanded what everyone feels responsible for — more production, more control, more information, in the same amount of working time.
- The erasure of borders. The ease of access to tools promotes micro-tasks in the evening, on weekends, during breaks — all of which are moments of cognitive recovery that are cut short.
Six proposals to regain control
The good news: none of these mechanisms are inevitable. They depend less on the technology itself than on how organizations integrate it. Here's what we recommend—and implement—for our clients.
1. Establish a clear charter of use and governance
The additional cognitive load skyrockets when the rules of the game are unclear: who checks what? What level of quality is expected? When should the use of AI be disclosed? A short usage charter, co-created with the teams, drastically reduces constant arbitration. It must, in particular, establish an anti-workslop standard: "AI helps you produce, but you remain the author." — you should never send a colleague content that you have not personally reviewed, understood and taken responsibility for.
2. To train in discernment, not just in the use of tools
Knowing how to prompt isn't enough. Studies from MIT and Microsoft converge: what protects against cognitive debt is the ability to choose when use AI, in what order (human effort first, AI then to amplify), and how to to retain control over the final judgment. That's precisely the philosophy of our format. Prompt Party®, in 1 hour or half a day: real-life situations where everyone learns to manage cognitive delegation rather than being subjected to it.
3. Protect attention as a strategic asset
Implementing periods of deep work without notifications, limiting the number of AI tools active simultaneously per employee, and grouping verification tasks rather than scattering them: these are all simple organizational measures that directly address attentional clutter. The teams that will succeed will not be those that pile up the most tools, but those that build the clearest workflows.
4. Make the verification work visible
As long as the review and validation of AI outputs remain invisible, they are neither planned, nor acknowledged, nor supported by tools. Integrate this time into workload estimates, team rituals, and even job descriptions. What is designated can be regulated.
5. Measure quality, not volume
If performance indicators focus on the quantity of deliverables, AI will inevitably produce workloads. Shift the metrics towards impact: deliverable rework rate, downstream correction time, internal recipient satisfaction. An AI deliverable that generates two hours of rework for a colleague is a net loss, not a gain.
6. Oversee the deployment with pilots and ambassadors.
Rather than hasty generalizations, proceed with targeted, measured, and then scaled-up use cases. A network of trained AI ambassadors plays a key role: disseminating best practices, providing peer feedback on usage, and reporting pain points. This is also the explicit recommendation of researchers from Stanford and BetterUp: mutual support and feedback among colleagues are among the most effective ways to combat workload stagnation.
In conclusion: AI should help us think better, not less.
Cognitive workload in the age of AI is not a technological problem, it's a problem of work design. Organizations that treat their teams' attention as a scarce resource—on par with budget or time—will transform the promise of productivity into a sustainable reality. Others will accumulate cognitive debt, the interest on which is paid in errors, disengagement, and departures.
Sources and references
Rock, D. & Weller, C., AI is frying our brains — here's what leaders need to do about itFortune, April 2026.
Bedard, J. et al. (BCG), When Using AI Leads to “Brain Fry”, Harvard Business Review, March 2026
Ranganathan, A. & Ye, X. (UC Berkeley, Haas School of Business), 8-month ethnographic study on AI-induced work intensification, 2026
BetterUp Labs & Stanford Social Media Lab, AI-Generated “Workslop” Is Destroying Productivity, Harvard Business Review, September 2025
MIT Media Lab, Report on the ROI of Investments in Generative AI, July 2025
MIT Media Lab, EEG study on the "cognitive debt" associated with ChatGPT use, 2025
Microsoft Research, The Impact of Generative AI on Critical ThinkingCHI Conference, 2025
To learn more: discover how to adopt AI without sacrificing team well-being, our analysis of productivity paradox and our article on the workslop.