Will AI Make Us Dumber? Method-Dependent Evidence

Question (2026-05-16): "Can you tell me specific data to argue that AI will not make us dumber if we follow a specific method?"

The honest one-liner

No single longitudinal study shows "users who follow method X retain capability while others lose it" — that study doesn't exist yet. The vault has converging evidence from cognitive science + workplace surveys that the method matters more than the tool. The defensible claim is "the cognitive-offloading harm isn't a property of AI, it's a property of how you use it" — not "follow this method and you'll be smarter."

The method (4 moves)

1. Retrieve before consulting — close the chat, articulate the answer cold, then check
2. Use AI as draft, not oracle — be a user who edits/verifies, not one who accepts
3. Build foundations first — domain fluency precedes delegation
4. Insert deliberate friction — ask AI to clarify, assign you homework, or show work before answering

The data backing each move

Move 1 — Retrieve before consulting

Bjork retrieval-practice study (Psychological Science), cited in Desirable Difficulties:

Group	Method	Retention after 1 week
Tested	Read once, then tested	80%
Re-read	Read twice	34%

Same input, same time → ~2.4× retention from forcing recall instead of re-recognition. Without retrieval, Ebbinghaus claims ~70% of new info is gone in 24h. The AI-dumbs-us-down mechanism is letting recognition substitute for recall (Fluency Illusion); the fix is to never skip the retrieval rep.

Move 2 — Edit, don't accept

Raymond Fu's 55/30 paradox (Learning Software Engineering During the Era of AI (Raymond Fu, TEDxCSTU)):

• 55% of developers use Copilot
• Only 30% accept its output without changes

The 25% gap — devs who use AI and edit it — are the structurally-protected group. The 30% are the Cognitive Offloading cohort. (Fu doesn't cite his source; treat as directional.)

Move 3 — Foundations first

Microsoft Research stat Charlie Gedeon cites in Is AI Making Us Dumber (Charlie Gedeon, TEDxSherbrooke): study of 319 knowledge workers, 60–70% reported less cognitive effort when using ChatGPT for comprehension, synthesis, analysis, and evaluation. Inverse reading: the 30–40% who didn't offload likely had the domain expertise to evaluate output. Andrej Karpathy's framing nails the mechanism: "You can outsource your thinking but you can't outsource your understanding." You can only judge AI output you're qualified to judge.

Move 4 — Productive resistance

Productive Resistance (Gedeon's term, from the When Copilot Becomes Autopilot paper he cites): the antidote to sycophantic, frictionless AI is design that asks clarifying questions or assigns prep before answering. The user-side version: do this manually —

• Paste your draft answer into ChatGPT before asking for its answer
• Ask it to interview you on the problem
• Request criticism, not output
• Use the TRAP Framework "T" step: close the source, say it back cold, only then check

Pushback I'd give before this is used in public writing

• Bjork data is on traditional learning, not LLM-mediated learning. The principle borrows cleanly, but it isn't a direct study of AI users.
• The MS-research stat measures self-reported effort, not capability loss. People reporting "I used less effort" isn't the same as "I got worse." That gap is where critics will attack.
• The Fu 55/30 numbers are uncited. Likely real (GitHub or developer-survey origin), but pin the source before quoting in print.
• The strongest defensible claim is method-dependence, not method-superiority. "AI won't make you dumber if X" should be framed as "AI's de-skilling risk is a function of usage discipline, and here's what the discipline looks like" — that's what the evidence supports without overshoot.

Cross-source contradiction this query sits on

Cognitive Offloading documents the central tension:

• Gedeon says the intervention point is AI design (regulate, build Productive Resistance into products, treat sycophancy as a dark pattern)
• Fu says the intervention point is human discipline (embrace AI, treat as creative partner, master foundations)

This query answers the Fu side — "what does the personal discipline look like." The complementary query on the Gedeon side is Designing AI Products That Don't De-Skill Users — "what design changes would make AI consumer products non-de-skilling by default?"

Brand-fodder note

This is on-thesis for the user's senior-IT-leader content brand on Medium/LinkedIn. The framing that works for that audience isn't "5 productivity tips" — it's "AI literacy is a leadership decision, not a personal habit. Here's the framework I use to keep my team's judgment intact while we 10× output." The 4 moves above are the framework; the data points are the receipts.

Sources

• Is AI Making Us Dumber (Charlie Gedeon, TEDxSherbrooke) — cognitive offloading, MS-research stat, productive resistance, dark patterns
• Learning Software Engineering During the Era of AI (Raymond Fu, TEDxCSTU) — 55/30 Copilot paradox, foundations-first curriculum
• How To Learn Anything So Fast (theMITmonk) — Bjork data, TRAP framework, Ebbinghaus
• Andrej Karpathy on Agentic Engineering (Sequoia AI Ascent) — "outsource thinking, not understanding"

Related concepts

Cognitive Offloading · Productive Resistance · Fluency Illusion · Desirable Difficulties · TRAP Framework · Dark Patterns · Ebbinghaus Forgetting Curve · Spaced Repetition