Yesterday, Meta announced it is cutting 10% of its global workforce — roughly 8,000 people — and redirecting $115 billion toward AI infrastructure. Teams are being reorganised into AI-focused pods. The future, per the memo, is agentic.
It is a significant bet; one that Microsoft, Amazon, and Google have taken on in some form. The logic is consistent: AI does more, so we need fewer people to do it. What is not being addressed though, is what happens to the judgment of the people who remain — the ones now tasked with directing, governing, and course-correcting AI systems they increasingly let do the thinking.
There is a growing body of research on what is being seen as a developing "cognitive debt."
The brain fry risk
The term "cognitive debt" comes from a 2025 MIT Media Lab study: 54 participants asked to write essays using ChatGPT, Google, or unaided. The brains using ChatGPT showed the weakest neural connectivity — and when those participants later wrote without AI assistance, the weakness persisted.
Cognitive debt accumulates through passive AI use — accepting the output, skipping the friction, letting the tool do the reasoning you could have done yourself. Across subsequent research, the consistent finding is not that AI damages thinking, but how it is used does.
In a 2026 study by Sarah Baldeo, published in the APA journal Technology, Mind, and Behaviour, participants who accepted AI's first answer reported lower confidence in their own reasoning afterwards. Those who pushed back — challenged the output, edited it, rejected it — reported the opposite. The variable wasn't the tool. It was the stance taken towards it.
A BCG Henderson Institute study published earlier this year in Harvard Business Review found that the drain extends beyond active use entirely. Monitoring AI — overseeing outputs without engaging with them — produces what workers themselves are now calling "brain fry." Productivity peaks at three AI tools simultaneously; above four, cognitive strain rises even as output appears to increase.
Productivity is a false idol
The metrics that make AI adoption look successful can be measuring the wrong thing entirely — output volume rising whilst the judgment behind it is degrading.
65% of employees in AI-adopting organisations report that AI has improved their productivity and efficiency — but at what cost and for how long?
The MIT research found that 80% of participants accepted the AI answer even when it was wrong — not because they were careless, but because the workflow was designed so that checking felt unnecessary. The friction of verification had been optimised away.
Meta's restructuring makes the stakes concrete. The bet is that the remaining workforce will have the judgment to run the pods. That may be true, but it assumes the judgment is being maintained: that the years of doing the work, making the calls, being wrong and learning why, have produced something durable enough to survive a transition that now also includes the normalisation of passive AI use.
Not in my name
There is an element that doesn't appear in the research but belongs in any honest conversation about AI at work: ownership.
When a piece of work leaves your organisation with your name on it — a strategy, an analysis, a client recommendation — you are accountable for the judgment in it. Not for the speed at which it was produced, not for how efficiently the tool was used, but for the work itself. AI-assisted work doesn't change that accountability. It raises the bar for it, because the ease of production makes it easier to submit something that looks right without having tested whether it is.
The organisations building good AI practice are building this into how their teams work: the understanding that using AI well isn't just about efficiency, but about being able to stand behind the output — to explain it, defend it, and improve it; which requires having thought it through.
AI as a whetstone
Most enterprise AI training programmes are designed to increase speed and confidence with the tool: how to prompt, how to integrate AI into existing workflows, how to get outputs faster. The research suggests the cognitive risk isn't that people can't use AI, it's that they use it in ways that gradually hollow out the judgment they're meant to be applying to the output.
AI adoption training must teach people to use the tool in ways that make them sharper, just like a knife on a whetstone. Three things that make a practical difference:
Protect the first draft. Workflows that use AI to generate the first draft and human review to polish it invert the cognitive load in a way that atrophies original thinking over time. The draft is where the thinking happens. Build its protection into how teams are trained to work.
Train disagreement as a habit. The research is clear: the users who maintain their cognitive confidence are the ones who override, challenge, and reject AI outputs. That habit doesn't develop naturally in a workflow optimised for speed. It has to be designed in — not as a compliance step, but as a practice. The question isn't "is this good enough?" It's "do I agree with this, and why?"
Measure what the tool is doing to thinking, not just to output. Adoption rates, time savings, and content volume tell you nothing about whether the judgment that makes the work valuable is being maintained or depleted. There is no dashboard for this yet, but consider monitoring revision rate of AI output, decision quality over time, and qualitative judgment audits where people work on a thinking or strategy task without AI input first.
The ones who remain
The question behind Meta's announcement — and behind every similar restructuring — is not how many roles AI can absorb, but what kind of workforce emerges on the other side.
The research points toward a bifurcation. On one side: people who have used AI in ways that enhance their thinking — who are sharper, more precise, better at the high-order judgment that AI cannot replace. On the other: people who have used it passively, whose critical faculties are less confident, who are increasingly good at supervising outputs they may no longer interrogate over time.
This is not an argument against AI transformation. It is an argument that the organisations investing billions in AI infrastructure need to invest proportionately in the cognitive health of the people operating it - not as a welfare measure, but as a business imperative.
There is a postscript to Meta's announcement that deserves to be read alongside the cognitive debt research. Days before the layoffs were confirmed, Reuters reported that Meta had begun installing tracking software on employee computers — capturing mouse movements, keystrokes, and periodic screenshots under a programme called the Model Capability Initiative.
The stated purpose: to train AI agents to perform white-collar tasks by learning exactly how humans do them. The programme sits within a broader internal initiative Meta has rebranded the Agent Transformation Accelerator. Workers have not been asked whether they consented, they are being informed and logged.
The end game of Meta's surveillance programme seems clear: the thinking of the ones who remain is being captured as training data for the systems that will make them soon unnecessary. To Zuckerberg, the humans are the inefficiency — and he is executing on it in the race for AI supremacy. But at some point, optimisation cannibalises itself. You cannot harvest human judgment indefinitely whilst removing the humans who provide it. In May, 8,000 people will find out their work is done. The rest will still be doing it — they just won't know for how long. And when they are gone, and the agents talk to each other, what thinking will be left to borrow?