Recap: Use AI to Think Better, Not to Think Less — The Memory Paradox Talk at Upper Bound 2026

By: on May 27, 2026
A backlit hand with water slipping through the fingers, evoking memory running out

This is the first in a short series recapping Upper Bound 2026, Amii's AI conference in Edmonton — 11,000 people, 22 countries, 200-plus sessions. I'm writing up the talks worth carrying home. First one: the session that shaped my recent post on quizzing yourself.

The Talk: "Memory Paradox in the Age of AI"

One session at Upper Bound 2026 was Dr. Nidhi Sachdeva's "Memory Paradox in the Age of AI" — TELUS Stage 3, May 21, in the AI Literacy & Education track. The frameworks she laid out are the actual seed of the cognitive-debt post I published a few days later. So this is a recap: a walk through the six frameworks she presented and the research behind each, with the facts and references straight from the source.

The paradox she names, in one sentence: the more we let AI think for us, the less we engage the very mental processes that make learning durable. AI makes the work feel easy, and ease is deceptive. You can look like you're learning while nothing actually lands.

Her core message — the line the whole talk hangs on, as I carried it out of the room:

"Use AI to think better. Not to think less."

Who She Is

Dr. Nidhi Sachdeva is an evidence-informed learning designer at the University of Toronto's OISE (Department of Curriculum, Teaching and Learning), with fifteen-plus years of post-secondary teaching behind her. She co-writes The Science of Learning newsletter (tens of thousands of subscribers) and chairs researchED Toronto. Her 2023 doctoral dissertation is on designing evidence-informed microlearning for graduate courses.

Worth stating plainly: she's a learning scientist first, AI second. The frameworks below aren't hype-cycle takes — they're decades of cognitive-science research pointed at a new question.

Framework 1: The 85% Sweet Spot

Optimal learning happens at roughly an 85% success rate. Not 100%, not 50%.

  • Too easy → no adaptation. The brain coasts.
  • Too hard → overload and frustration. You bail.
  • ~85% right → the band where neural growth and pattern recognition happen.

From my notes on the slides: "Difficulty is not the enemy. It's the signal that learning is happening."

Reference: the 85% figure comes from Wilson, Shenhav, Straccia & Cohen, "The Eighty-Five Percent Rule for optimal learning" (Nature Communications, 2019). It rests on Robert and Elizabeth Bjork's desirable difficulty research — conditions that make learning feel harder in the moment produce better long-term retention.

Framework 2: The Flip — Provoke Thinking, Don't Replace It

The move: before the AI hands you the answer, you go first.

  • Instead of AI writing the answer → predict it first, then compare. (Retrieval practice.)
  • Instead of AI summarizing a text → critique and refine the AI's summary. (Elaboration.)
  • Instead of AI explaining a concept → visualize it and teach it back. (Schema building.)
  • Instead of taking AI's feedback at face value → justify your reasoning out loud first. (Self-explanation.)

This is the framework that became my quiz-yourself routine. "Predict first, then compare" is retrieval before resolution — the retrieval has to happen in your head before the model resolves it, or the wiring never gets built.

Reference: retrieval practice (Agarwal and colleagues — see retrievalpractice.org), elaboration and self-explanation strategies surveyed in Dunlosky et al. (2013).

Framework 3: Amplifier vs. Substitute

Same AI tool, two very different outcomes — and the difference isn't the tool, it's what the learner brings to it.

  • The Amplifier: already has real knowledge and mental models. Uses AI to evaluate outputs, sharpen prompts, push further than they could alone. AI multiplies what's already there.
  • The Substitute: still building the knowledge. Mistakes the AI's fluency for their own understanding, skips the retrieval step, and can't separate what they know from what the model knows.

Her framing, as I jotted it: "Same AI. Different outcomes. The difference is what learners bring with them."

The danger is the substitute looks competent right until the tool is taken away — AI fluency standing in for genuine expertise, with the gap invisible until they have to perform without it.

Framework 4: Cognitive Offloading vs. Cognitive Outsourcing

  • Cognitive offloading (acceptable): using AI to extend your reach. You still do the core thinking; the AI handles routine or amplifies it. Draft with it, then critique, refine, and own the result.
  • Cognitive outsourcing (dangerous): skipping the thinking entirely. Ask for the answer, paste it in, call it learning. The understanding never lands.

This is the distinction underneath what I called "cognitive debt" — outsourcing is the act, the debt is what it leaves behind. The substitute trap is just cognitive outsourcing by another name.

Framework 5: The Retention Revolution

The principle: "Retrieve to retain."

  • Space it out. Learning distributed over time beats one long cramming session. (Spaced practice.)
  • Don't re-read. Recall. Testing yourself beats passive review for long-term retention. (Active recall.)
  • "Every repetition builds the pattern." Each retrieval lays down a stronger neural trace.

For anyone learning through AI, the takeaway is blunt: don't just consume AI's explanations. Use AI to quiz you, not to answer for you, and space the sessions out instead of one passive marathon.

Reference: spaced and retrieval practice both rank at the top of Dunlosky et al. (2013), "Improving Students' Learning With Effective Learning Techniques" (Psychological Science in the Public Interest); the test-enhanced learning effect comes from Roediger & Karpicke (2006).

Framework 6: Four Principles to Design With

The talk's practical checklist:

  1. Practice unaided. Mental math, handwriting, reasoning without a copilot. Build the base skills independently before delegating them.
  2. Embrace difficulty. Aim for that ~85% success rate. Productive struggle is the mechanism, not a bug.
  3. Internalize foundations. Some things have to live in your head before you can think with them. Don't skip the foundational work.
  4. Use AI as a supplement. Let it prompt and critique — but you do the core thinking. Amplifier, not substitute.

The Research Base

None of this was invented for the AI moment — it's established cognitive science aimed at a new target. The foundation she draws on:

  • Cognitive Load Theory — John Sweller. Working memory is limited; overload it and learning stalls.
  • Principles of Instruction — Barak Rosenshine (2012). Small steps, worked examples, scaffolding.
  • Retrieval practice — Agarwal et al. (retrievalpractice.org). Testing beats re-reading; the engine under "The Flip."
  • Effective learning techniques — Dunlosky et al. (2013). Spaced and retrieval practice ranked highest by evidence strength.
  • The 85% rule — Wilson et al. (Nature Communications, 2019), with Bjork & Bjork on desirable difficulty.
  • Dual Coding Theory (Clark & Paivio, 1991) and Mayer's Cognitive Theory of Multimedia Learning — how format affects what sticks.

Primary sources for Sachdeva herself: her Science of Learning newsletter, her dissertation on microlearning, and the explainer she co-wrote, "Let's focus on 'Learning' in MicroLearning" on The Learning Scientists blog.

Why It Stuck With Me

The reason this one shaped my own writing: it gives names and research to a thing every one of us pairing with AI is quietly negotiating — am I getting smarter, or just feeling smarter? Three lines I'm keeping:

  • The goal isn't to use AI less. It's to use it so that you still do the retrieval.
  • "Feels easy" is a warning light, not a win. Frictionless usually means nothing landed.
  • Amplifier or substitute is on you, not the tool — it depends entirely on what you bring to the session.

If you want the hands-on version of all this, I turned Framework 2 into a concrete routine: make your AI quiz you at the end of a session, complete with the prompt I use. That post is The Flip, applied to shipping code.

Use AI to think better, not to think less.

The rest of the Upper Bound 2026 series:

Header photo by Ashique Anan Abir on Unsplash.

Content on this blog was created using human and AI-assisted workflows described here. Original ideas and editorial decisions by Justin Quaintance.