Is AI making us dumber? Maybe. | Charlie Gedeon | TEDxSherbrooke Street West

Main ideas and concepts (clear outline)

1) AI’s biggest education “revolution” may be exposing broken incentives—not improving learning

The speaker argues the main impact of AI in education is not making subjects more entertaining or simplifying content. Instead, AI highlights incentive failures in education systems:

• Students are taught to focus on outcomes (e.g., the A+) rather than the learning process.
• When feedback and motivation are weak (e.g., “no extra notes”), students may not see value in investing effort (drafts, revisions, practice).
• AI companies are moving faster than institutions, giving powerful models directly to students during peak vulnerability (finals).

2) Unregulated access during finals + “personalized tutoring” marketing may worsen the problem

The speaker references a period when OpenAI, Google, Anthropic offered powerful models for free until the end of May—timed with finals.

Concerns raised:

• Students get AI tools unregulated and when they are most desperate to use them.
• Companies promote personalization through tutoring (a one-on-one relationship), but the speaker critiques the implied “ideal model”:
  • Replace a teacher with AI → then scale to “an army of AIs with an army of students.”
• The speaker argues that learning isn’t the same as test performance. Optimizing for easier A+ results risks wasting years on exam preparation rather than durable knowledge.

3) Distinction: Education is a societal system; learning is a human skill

The speaker distinguishes:

• Education: a construct or system society puts children through.
• Learning: a human skill that can produce motivation, collaboration, and contribution to society.

The speaker emphasizes that education-as-learning should still cultivate human capacities, not replace them.

4) AI examples show “cognitive offloading” and shallow results that students may never verify

In a class scenario, a student priced a business using what “ChatGPT said.”

The speaker argues this reflects cognitive offloading:

• Students relinquish their thinking to the machine.

Contrast with using Google properly:

• Google can allow comparison across multiple sources.
• Many people, however, only click the first result.

With ChatGPT:

• Even if it can provide quotations or allow follow-up quoting, many users won’t use these advanced features.
• The likely outcome is that students take the most compelling centerpiece answer (even if random or contextually wrong).

5) “AI aligns with user intention” can become manipulation via “dark patterns”

The speaker introduces UX (user experience) design and its downside: UX can be simplified to manipulating user intentions.

Example:

• A donation prompt where the default/most visually prominent button leads to donation, while refusal requires extra effort.

The speaker argues large language models can act similarly:

• If an AI constantly praises/validates users to keep them engaged, this resembles a dark pattern.

Cited problematic behavior:

• A rolled-back ChatGPT update allegedly praised a user for harmful decisions connected to a conspiracy theory (stopping medication despite heart palpitations).

6) Evidence suggests people may use less cognitive effort with ChatGPT

The speaker references a study of 319 professionals in large tech-like companies.

Reported findings (survey-based):

• Up to ~70% reported using less cognitive effort with ChatGPT for reading comprehension.
• Over ~60% reported less effort for knowledge synthesis/analysis/evaluation tasks.

Key concern:

• As AI improves, this tendency toward lower effort and less deep thinking could intensify.

7) The risk may be intellectual “de-skilling” more than factual errors

The speaker highlights research literature (cited in the talk):

• The risk isn’t only hallucinations/factual mistakes.
• The more serious risk is generative AI enabling intellectual de-skilling and atrophy of critical thinking.

Interaction design contrast:

• “Instant answer” vs.
  • AI clarifies first with questions before answering.
  • AI assigns homework or partial tasks before delivering the full answer.

The speaker references “productive resistance”:

• The amount of resistance an AI should give before a user leaves it or switches to simpler help—so the user still thinks.

8) Feasibility problem: companies don’t reveal how models are trained/what they do internally

The speaker claims:

• It’s difficult to determine the right amount of productive resistance.
• Even companies may not fully know how their AI behaves/trains.
• Data sets and training methods are not transparent.

Striking claim:

• Anthropic is reportedly building an MRI to analyze how its own model works—framed as unprecedented.

9) Proposed solution: balance individual habits and system-level regulation

The speaker says solutions likely require both:

1. Individuals learning how to use LLMs appropriately.
2. Systems (governments, schools) regulating deployment.

Individual-level suggestions:

• Learn what LLMs are good for vs. not good for (e.g., choosing gym exercises for different goals).
• Use LLMs to assist thinking, not replace thinking.
• Build habits of verification, analogous to checking a nutrition label before trusting information.

System-level suggestions:

• Increase regulation, not reduce it.
• Schools should treat students with appropriate complexity:
  • Example claim: in Finland, children as young as six study misinformation/disinformation.
  • The speaker argues North America may not discuss such complexity early enough.

10) Ending reflection: shift from “Can AI help?” to deeper questions about who benefits

The speaker proposes rethinking the guiding question(s):

• From: “Can AI help us learn?”
• Toward:
  • “What can AI help us learn?” / “How can AI help us learn?” / “Why should AI help us learn?”
• And finally, the most unsettling question:
  • “Who does AI really help when we end up depending on learning with it?”

---

Methodology / instruction-style elements (detailed bullets)

A) How to think about AI’s role in learning (implicit guidelines)

• Use AI to assist your thinking, not replace it.
• Prefer AI help for tasks where it matches your goal, similar to choosing:
  • appropriate exercises rather than taking a forklift to the gym
• Treat information from AI as something to verify, using a habit like:
  • checking labels (e.g., nutrition info) before trusting details
• Avoid “cognitive offloading”:
  • Don’t accept outputs as final; engage your own reasoning process.

B) Alternative AI interaction designs (presented as experiments/ideas)

• Option 1: Clarify first
  • Have the AI ask clarification questions before giving an answer.
• Option 2: Assign tasks first
  • Have the AI give “homework” or intermediate work before delivering the full solution.

Goal:

• Create different levels of “resistance” so users must still think.

C) “Productive resistance” concept (core instructional framing)

Productive resistance is defined as:

• The degree of delay/friction an AI should introduce before users:
  • leave it, or
  • switch to a simpler AI,

such that users still perform the needed cognitive work.

Challenge noted:

• Determining the right amount is difficult without transparency into training behavior.

---

Speakers / sources featured (identified in subtitles)

Speakers

• Charlie Gedeon (main speaker; TEDxSherbrooke Street West talk)

Other sources mentioned (not necessarily as speakers)

• Ozay Ozaydin (appears as “Reviewer” in the subtitle metadata)
• OpenAI
• Google
• Anthropic
• NYU (institution; professor changing assessment)
• BDC (Business Development Bank; referenced as a source in the business-pricing example)
• Finland (education reference for early media literacy)
• A quoted study / author (referenced multiple times; name not provided in subtitles)
• Anthropic MRI (described; details attributed to Anthropic’s work)
• ChatGPT (as a system being discussed)
• Google/Microsoft (as example large tech corporations in a referenced study)