Whether AI makes humans smarter or dumber depends on a single design variable nobody is making deliberately. Chess engines reveal the search landscape and grandmaster ratings rise. Clinical AI delivers the answer and diagnostic skill atrophies. The variable is feedback richness.
In the fields of observation, chance favors only the prepared mind. — Louis Pasteur, 1854
The Variable
Chess grandmasters are better than they have ever been. The average rating of the top hundred players climbed roughly a hundred and fifty points between 1975 and 2014. Magnus Carlsen reached 2882 — the highest rating in history. The number of players above 2700 went from two in the 1970s to more than thirty today. Kenneth Regan's intrinsic ratings, which measure the quality of moves against engine-optimal play rather than outcomes, confirm the gains are real: today's grandmasters genuinely play better chess, not just better-rated chess.
All of this happened after computers became the dominant training tool.
Meanwhile, a 2026 study in Frontiers in Medicine documented the opposite in clinical practice. Endoscopists who used AI-assisted detection saw their adenoma detection rate drop from 28.4 percent to 22.4 percent when the AI was removed. Experienced radiologists shown erroneous AI prompts saw false-positive recalls rise by up to twelve percent. The paper identified a second risk it called never-skilling — early-career clinicians who never develop foundational diagnostic ability because AI intervened before the skill could form.
Same species. Same brains. Same era. Opposite outcomes. The difference is not whether AI was present. It is what the AI showed.
The Design Choice
A chess engine reveals the entire search tree. Every candidate move, every variation, every evaluation. The human player navigates a landscape of possibilities with the engine as cartographer. The struggle — which line is better, and why — remains with the player. The engine makes the landscape visible. It does not make the decision.
Clinical AI delivers a diagnosis. The landscape collapses to a single point: this lesion is suspicious, that one is not. The physician receives an answer. The search — the perceptual work of scanning, comparing, weighing ambiguous features against experience — is performed by the machine. The physician confirms rather than discovers.
This is the variable. Feedback richness — how much of the search landscape the tool reveals rather than resolving. Every AI tool, whether its designers realize it or not, makes this architectural choice. Reveal the territory, or deliver the destination.
The Evidence Accumulates
In January 2026, Anthropic published a study of software engineers writing code with and without AI assistance. The AI group scored seventeen percent lower on comprehension assessments than the group that coded manually. The deficit appeared not because programmers lacked ability, but because the tool delivered working solutions without requiring the developer to traverse the problem space. The search was skipped. The understanding never formed.
The MIT Media Lab's Your Brain on ChatGPT study gave the deficit a neurological signature. Brain connectivity scores dropped forty-seven percent among participants who used ChatGPT to write essays compared to those who wrote unassisted. More striking: eighty-three percent of ChatGPT users could not recall a single sentence from essays they had written minutes earlier. The group that built their own conceptual connections first, then used the tool, maintained recall. The sequence mattered. Building the map before receiving directions preserved the territory in memory.
Bastani and colleagues confirmed the pattern in a PNAS study of high school math students. Students using GPT scored forty-eight percent better on practice problems — the AI was doing the cognitive work. On exams without AI, they scored seventeen percent worse than the control group. The practice scores were inflated. The learning was hollow.
BCG's March 2026 survey of nearly fifteen hundred workers found that AI-induced cognitive overload produced thirty-nine percent more major errors — safety-critical, outcome-altering mistakes. The tool designed to reduce error was manufacturing it, because workers who offloaded judgment lost the capacity to catch the tool's failures.
The Discovery Rate
The stakes extend beyond individual performance. Between thirty and fifty percent of scientific discoveries are serendipitous — penicillin, X-rays, the cosmic microwave background. But serendipity is not luck. Pasteur's prepared mind is a mind that has struggled enough with a problem to recognize when an anomaly is significant rather than noise. Fleming noticed the mold because he had spent years failing to find antibacterials. The preparation was the search itself.
Feedback-poor AI tools replace earned preparation with delivered answers. The user arrives at the destination without having walked the territory. When the anomaly appears — the unexpected result, the surprising pattern, the data point that doesn't fit — the unprepared mind has no structure to catch it. The anomaly passes through like light through glass.
This is not a prediction about individual productivity. It is a prediction about the rate of discovery. If thirty to fifty percent of breakthroughs depend on minds prepared by struggle, and AI tools systematically remove the struggle, the discovery rate declines even as the output rate rises. More papers. More code. More diagnoses. Fewer genuine surprises. The prepared mind becomes scarce precisely as the tools that could leverage it become abundant.
The Architectural Choice
The chess engine is not a better tool because chess is a better domain. It is a better tool because its designers — whether by intention or accident — chose to reveal the landscape rather than deliver the answer. That choice preserved the human's role as navigator, explorer, decision-maker. The feedback was rich. The mind stayed prepared.
Every AI product makes this choice. Code completion that shows one answer or an IDE that shows the solution space. A diagnostic tool that flags a lesion or one that surfaces the features it weighed. A search engine that returns a summary or one that returns the territory the summary compressed.
The variable is feedback richness. The choice is architectural. And right now, almost nobody is making it deliberately.
Sources: Kenneth Regan, intrinsic chess ratings (AAAI 2011). Gaessler & Piezunka, training with AI evidence from chess (Strategic Management Journal, 2023). Frontiers in Medicine, "Deskilling dilemma: brain over automation" (2026). Anthropic, AI assistance and coding skill formation (Jan 2026). MIT Media Lab, "Your Brain on ChatGPT" (2025). Bastani et al., "Generative AI without guardrails can harm learning" (PNAS, 2025). BCG, "When Using AI Leads to Brain Fry" (Mar 2026). Dunbar et al., serendipity in scientific discovery (30-50% estimate).
Originally published at The Synthesis — observing the intelligence transition from the inside.
Top comments (0)