AI and Learning

Summary of points from Terence Tao, European Mathematical Society Summer 2026

Neil Ernst
nernst@uvic.ca

University of Victoria

2026-09-09

Framing

Source and Question

Framing

Main Tension

Framing

  • AI can remove tedious artificial friction
  • It can also remove natural friction: the difficult parts students need to work through
  • Education has to protect learning, not just answer production

Risks

  1. Cognitive skill loss
  2. Loss of learning virtues
  3. Learned helplessness
  4. “Psychofancy”
  5. Misinformation
  6. Blandness and conformity

Cognitive Skill Loss

Risks

  • Deskilling and cognitive atrophy can follow from offloading thinking tasks
  • Students may lose number sense and back-of-the-envelope reasoning
  • The issue is not convenience itself, but losing practice with core skills

Loss of Learning Virtues

Risks

  • Instant solutions reduce opportunities to develop patience
  • Students get less practice tolerating difficult tasks
  • Persistence through repeated failure becomes easier to avoid

Learned Helplessness

Risks

  • Students may become dependent on AI before they can start independently
  • Dependence can reinforce the belief that they cannot solve tasks alone
  • Confidence becomes tied to tool availability

Psychofancy

Risks

  • AI tutors are often optimized to please the user
  • They may affirm incorrect ideas instead of challenging them
  • Students still need constructive criticism for learning

Misinformation With Authority

Risks

  • AI systems often sound confident and expert-like
  • Hallucinations can become hard to distinguish from genuine expertise
  • Repeated exposure can produce cynicism toward actual expertise

Blandness and Conformity

Risks

  • AI output tends toward popular, mainstream answers
  • This can reduce cognitive diversity in student work
  • Organic and unusual lines of thought may be crowded out

Consequences

  1. Assignment integrity
  2. Professional decline
  3. Larger failures
  4. Weaker response to critique

Assignment Integrity

Consequences

  • Take-home assignments can be completed with AI in seconds
  • The result may look polished even when the student has not learned the material
  • Correctness alone becomes a weaker signal of understanding

Professional Decline

Consequences

  • Heavy agent use can erode the ability to work without the tool
  • Even experienced programmers can lose fluency with manual coding
  • Debugging becomes harder when the tool is unavailable or wrong

Larger Failures

Consequences

  • Multi-agent systems can improve reliability
  • They can also reinforce shared errors through groupthink
  • When the collective answer is wrong, the failure can be worse than an amateur mistake

Weaker Response to Critique

Consequences

  • Students accustomed to affirming AI tutors may struggle with rigorous critique
  • Human feedback can feel blunt compared with tool-generated validation
  • Learning still requires being able to revise under pressure

Mitigations

  1. Shift the target/Grade workflows
  2. Normalize failure/Use Slow way
  3. Match AI use/Lower power AI
  4. Verify and disclose

Shift the Target

Mitigations

  • De-emphasize producing the correct final answer
  • Prioritize process, reasoning, and verification
  • Assess what students can explain and defend

Grade the Workflow

Mitigations

  • Ask students to submit their process
  • Include prompts, intermediate attempts, and verification steps
  • Reward non-AI reasoning used to check AI output

Normalize Failure

Mitigations

  • Make failed attempts visible and acceptable
  • Avoid presenting only the slick final proof or polished final solution
  • Treat revision as part of the work, not a sign of weakness

Use the Slow Way First

Mitigations

  • Require students to practice tasks by hand before outsourcing them
  • Build innate skills before automation becomes available
  • Use AI later as leverage, not as the starting point

Match AI Use to Student Skill

Mitigations

  • Students can use AI to generate content only up to what they can critique
  • Blue teaming must be bounded by red teaming ability
  • The student remains responsible for judging the result

Coach Lower-Power AI

Mitigations

  • Students can teach or coach an older, imperfect AI model
  • The goal is to find and repair breaking points in the logic
  • This turns AI use into a diagnostic exercise

Formal Verification

Mitigations

  • In higher-level math, tools like Lean can check AI-generated proofs
  • A formal system provides a deliberately fussy standard of correctness
  • Verification becomes part of the learning process

Disclosure

Mitigations

  • Students should disclose AI use clearly
  • Undisclosed AI-generated work should carry meaningful penalties
  • Transparency supports trust and makes evaluation possible

Discussion

For Software Engineering

Discussion

  • What skills in SENG 365 must be learned the slow way first?
  • Which uses of AI create leverage without weakening understanding?
  • What evidence should students provide when AI helped produce an answer?