There is a kind of progress that erodes what it claims to improve. Roads built for cars hollowed out the pedestrian city. Ultra-processed food engineered for convenience rewired our appetite and metabolism. Now, artificial intelligence—the most capable cognitive tool our species has ever constructed—is doing something similar to the very faculties it was designed to augment.
The irony sits in plain sight, yet we keep averting our gaze.
Worldwide spending on AI reached nearly $1.5 trillion in 2025, with projections crossing $2 trillion in 2026. Corporate AI investment hit $252.3 billion in 2024 alone, growing more than thirteenfold in a single decade. Governments race to build AI gigafactories. Boardrooms declare AI integration a competitive imperative. The pace is dizzying, the commitment structural. Meanwhile, budgets for human development, critical thinking education, and cognitive resilience remain, at best, an afterthought—at worst, a line item being cut to fund the infrastructure that replaces the need for them.
Ask yourself: When was the last time your organisation invested comparably in developing the minds that operate the machines, and the hearts that propel those minds to care?
10 Minutes to Wake Up
That is the duration that stopped researchers in their tracks. A new study from Carnegie Mellon, Oxford, MIT, and UCLA—1,222 participants, three randomised controlled trials—offered the first large-scale causal evidence that brief AI use can measurably impair both cognitive performance and the willingness to persist through difficulty. Participants solved fraction problems and SAT-style reading comprehension questions. Half worked alone. Half had access to a GPT-5 sidebar chatbot. Then, without warning, the AI was removed. The three problems to be addressed were identical for both groups.
The AI cohort did worse in terms of outcomes once they were deprived of their artificial assistants. They also skipped problems at roughly twice the rate of those who had worked unaided throughout.
Previous assumptions about deskilling placed the danger in the slow lane—a gradual drift across months of dependency, a gentle erosion barely perceptible until the skill was already gone. This study relocates the concern. The shift happened within the first session; 10 to 15 minutes of AI assistance was sufficient to alter both performance and motivation. Agency decay is an acute risk, and it strikes fast.
The skip rate deserves special attention. Skipping is a motivational measure. It signals something about willingness, about tolerance for uncertainty, about the felt ratio of effort to reward. The authors propose a mechanism worth sitting with: Once AI collapses time-to-answer to near zero, unaided work starts to feel harder than it originally felt (and actually is). The reference point for reasonable effort resets downward. The mind recalibrates its baseline—and finds its own unassisted capacity suddenly disagreeable by comparison.
A boiling frog effect strikes—dependency accumulating without notice, like temperature changes too gradual to trigger an alarm. Based on the study, it appears that those individuals who use the AI to get direct answers show the steepest decline. Those who use it to request hints or clarifications perform roughly on par with the control group. A caveat applies—the experiment worked with self-reporting, and participants chose their own usage style. Still, the implication threads through the data clearly enough: The way we engage with AI shapes what it does to us.
The Mechanics of Agency Decay
The cognitive machinery of the human species evolved over thousands of years to do hard things. Solving problems under uncertainty. Holding ambiguity. Persisting through frustration. Pursuing purpose. These capacities are, in the deepest sense, the substrate of agency—the raw materials of autonomous thought and meaningful action. They are also, like all biological systems, subject to the logic of use-dependency. Faculties exercised regularly stay sharp. Faculties delegated atrophy.
Agency decay is the process by which this delegation, repeated and normalised, gradually reduces a person’s actual capacity—and felt willingness—to act, think, and decide independently. It operates through comfort. Through convenience. Through the perfectly rational-seeming choice to let the faster, more accurate system handle it.
The hybrid world we now inhabit runs on this logic. Artificial intelligence handles the research, the first draft, the calculation, the routing, the summary, and the translation. Human beings increasingly occupy the role of approvers—scanning outputs, clicking «Accept,» moving on. This is the operational picture in countless organisations, classrooms, and households. At each handover, a little more cognitive territory quietly shifts.
Artificial Intelligence Essential Reads
What makes this genuinely dangerous—rather than merely inconvenient—is the asymmetry of awareness. The capabilities lost in small increments rarely announce their departure. You simply find, one day, that wrestling with a hard problem feels unreasonably taxing, that the instinct to skip arrives faster than it used to, that your reference point for what counts as acceptable difficulty has moved somewhere you did not consciously choose to put it.
The Intelligence Gap
The machines, for their part, keep improving. The trajectory of artificial intelligence is steep, sustained, and generously funded. Natural intelligence—the original, embodied, evolved kind—receives no comparable investment. The systems for cultivating it are underfunded, the practices for maintaining it are undervalued, and the cultural norms that might encourage deliberate cognitive effort are, at this particular historical moment, swimming hard against the current.
We are, in effect, building an extraordinarily sophisticated prosthesis while letting the limb weaken underneath it.
This is the central irony of the hybrid era. Capability and dependency grow in tandem. The smarter the tool, the less friction stands between question and answer—and the more our own wrestling capacity falls into disuse. The dividend promised by AI—time freed for higher-order thinking—slips away precisely because the higher-order thinking muscles have been quietly relieved of their workload.
The gap between what AI can do and what we are developing ourselves to do has the makings of a civilisational blind spot.
A Practical Takeaway: AGENCY
A – Audit
Map the cognitive tasks you have stopped attempting alone. Awareness is where reclamation begins.
G – Grapple
Choose at least one hard problem per day to wrestle with unaided. Struggle is not inefficiency — it is the mechanism of cognitive maintenance.
E – Engage
Use AI as a hint machine, not an answer machine. Ask for a direction, a clue. Keep the solving in your hands.
N – Notice
When the urge to give up arrives faster than it used to, treat that as information, not instruction. The discomfort is usually a calibration error, not an actual incapacity.
C – Calibrate
Periodically work without AI on tasks you would normally delegate. Track what happens. Your baseline for reasonable effort needs tending, or it will quietly migrate.
Y – Your sovereignty
Agency does not hold itself. In a world engineered for frictionless assistance, the deliberate choice to think hard—and keep thinking—may be the most consequential habit you can build.
The intelligence gap is real. The investment gap is real. The window to close both—while we still notice what we are losing—is open. But the window of opportunity is closing quickly.
www.psychologytoday.com
