AI and Learning: How ChatGPT and Claude Are Reshaping How We Think, Read, and Remember

The Great AI Learning Paradox

Here's a paradox that defines our moment: AI makes us more productive in the short term while potentially making us less capable in the long term.

A landmark 2024 study from the Wharton School of Business demonstrated this perfectly. Roughly 1,000 high school math students in Turkey were split into three groups: one using ChatGPT freely, one using a specially designed "GPT Tutor" with guardrails, and a control group with no AI access.

During practice sessions, the unrestricted ChatGPT group performed 48% better than the control. They breezed through problems. They felt confident. But when it came time for the exam (with no AI access), they scored 17% worse than students who had never used AI at all.

The AI hadn't taught them math. It had done the math for them.

Meanwhile, the GPT Tutor group (which gave hints and guided questions instead of direct answers) performed 127% better during practice and matched the control group on exams. Same AI technology. Radically different outcome based solely on how it was used.

This single study encapsulates everything you need to know about AI and learning: the tool isn't the variable. Your relationship with it is.

---

What the Research Actually Says

The debate about whether AI helps or hurts learning has moved beyond opinion. Here's what multiple peer-reviewed studies have found:

Studies Showing AI Can Harm Learning

Studies Showing AI Can Help Learning

The pattern is clear: the same technology produces opposite outcomes depending on how it's deployed.

The MIT Brain Study: A Wake-Up Call

Perhaps the most striking finding comes from MIT Media Lab's 2025 study, "Your Brain on ChatGPT." Researchers had 54 participants write essays over four months using either ChatGPT, Google Search, or no tools, while measuring brain activity with EEG.

The results were sobering:

• ChatGPT users showed the weakest neural connectivity across all measured bands, including those associated with creativity, memory formation, and semantic processing
• Over the four months, LLM users "consistently underperformed at neural, linguistic, and behavioral levels"
• They produced near-identical essays lacking original thought
• They couldn't accurately quote their own work afterward

This isn't about intelligence. It's about what happens when you outsource the process of thinking. The brain, like any other organ, adapts to the demands placed on it (or the lack thereof).

---

The Cognitive Offloading Problem

Cognitive offloading is the practice of using external tools to reduce the mental effort required for a task. Writing a shopping list instead of memorizing items is cognitive offloading. So is using a calculator. And so is asking ChatGPT to summarize an article instead of reading it yourself.

The concept isn't new. But the scale and ease of AI-powered cognitive offloading is unprecedented.

What the Data Shows

Michael Gerlich's 2025 study of 666 participants across diverse age groups found:

• Significant negative correlation between frequent AI tool usage and critical thinking abilities
• Younger participants showed higher AI dependence and lower critical thinking scores
• Higher trust in AI correlated with more cognitive offloading. The more you believe AI is right, the less you check.

A Microsoft and Carnegie Mellon University study confirmed this: workers who trust AI the most check it the least. This creates a dangerous feedback loop: frequent AI use builds trust, which reduces verification, which weakens critical thinking, which increases dependence. The cycle repeats.

Researchers have coined terms for this phenomenon:

• "Cognitive atrophy": the weakening of mental capabilities from disuse
• "AICICA" (AI Companion-Induced Cognitive Atrophy): a specific term for AI-related cognitive decline
• "Metacognitive laziness": diminished self-monitoring and self-regulation when AI handles the thinking

Why This Matters for Knowledge Workers

The implications extend far beyond education. A Frontiers in Psychology paper (2025) identified a "cognitive paradox" in AI-assisted learning: while AI enhances personalized learning, excessive reliance reduces cognitive engagement and long-term retention.

For anyone who reads, researches, or learns professionally, this creates a genuine tension: the tools that make us most productive today may be eroding the skills that make us valuable tomorrow.

---

When AI Helps Learning: The Augmentation Model

Not all AI usage is created equal. Research consistently identifies patterns where AI genuinely enhances learning:

1. AI as a Socratic Tutor

The Wharton study's most important finding wasn't that AI hurts learning. It was that properly designed AI dramatically helps it. Their GPT Tutor provided:

• Hints instead of answers
• Guiding questions that prompted thinking
• Scaffolded support that decreased as the student progressed

Result: 127% better practice performance, with no penalty on exams.

Dartmouth's 2025 "NeuroBot TA" study showed similar results with medical students. The key was that the AI pulled from curated expert sources and guided students through reasoning, rather than simply providing answers.

2. AI for Knowledge Construction

A 2025 study in Studies in Higher Education distinguished between two modes of AI use:

• Mastery approach: Using AI to construct and augment knowledge (asking it to explain concepts, challenge your understanding, generate examples)
• Procedural approach: Using AI to complete tasks mechanically (copy-paste, generate drafts without engagement)

Higher-level learning occurred only in the mastery approach. The mere presence of AI tools didn't matter. What mattered was the learner's intent and engagement.

3. AI for Personalized Pacing

Meta-analysis data (Ma, 2025) shows a combined positive effect size of g = 0.68 across studies, with particularly strong effects in:

• Cognitive dimension (g = 0.795): understanding and applying concepts
• Competency dimension (g = 0.711): developing skills

AI tutoring shines when it adapts to individual pace, identifies knowledge gaps, and provides targeted practice. These are things that are difficult to scale with human instruction alone.

---

When AI Hurts Learning: The Dependence Trap

The same research identifies clear patterns where AI degrades learning outcomes:

1. The Copy-Paste Spiral

MIT Media Lab tracked a disturbing trend: copy-paste behavior increased over the four months of their study. As participants became more comfortable with ChatGPT, they engaged less with the material, not more. They stopped paraphrasing. They stopped adding their own analysis. They became transcribers of AI output.

2. The Illusion of Understanding

When ChatGPT explains a concept clearly, you feel like you understand it. But there's a critical difference between recognizing an explanation and generating understanding.

This is the same phenomenon that makes re-reading textbooks feel productive but ineffective. The familiarity of the material creates a false sense of mastery. AI amplifies this illusion because its explanations are often clearer and more accessible than original sources.

3. Erosion of Productive Struggle

Learning requires effort. Cognitive science calls it "desirable difficulty": the productive struggle that forces your brain to form new neural connections. When AI removes this struggle, it removes the learning itself.

The Wharton study's unrestricted ChatGPT group experienced exactly this: they breezed through practice (no struggle) and failed on exams (no learning).

---

The Two Archetypes: Centaurs vs. Cyborgs

The Harvard/BCG study of 758 consultants identified two successful models for human-AI collaboration:

Centaurs

Like the mythical creature, Centaurs maintain a clear boundary between human and AI work. They strategically decide which tasks to delegate to AI and which to handle themselves.

• Approach: "I'll do the creative thinking and analysis. AI will handle the data processing and formatting."
• Strength: Preserves human judgment for complex decisions
• Risk: May underutilize AI capabilities

Cyborgs

Cyborgs integrate AI continuously throughout their workflow. They don't divide tasks; they collaborate on every step.

• Approach: "AI and I work together on everything, each contributing our strengths."
• Strength: Maximizes AI leverage across the entire process
• Risk: Harder to maintain independent critical thinking

The Critical Caveat

Both models worked well for tasks within AI's capabilities. For tasks outside AI's frontier, both models led to a 19 percentage point decrease in performance.

This brings us to one of the most important concepts in AI-augmented work: the "Jagged Technological Frontier." AI capabilities are uneven: brilliant at some tasks, mediocre at others, and terrible at a few. The most effective learners and workers are those who accurately map where the frontier lies for their specific domain.

---

A Practical Framework for AI-Augmented Learning

Based on the research, here's a framework for using AI that enhances rather than replaces learning:

The HEAR Framework

H - Highlight First, Ask Second

Before asking AI anything, engage with the source material directly. Read the article. Watch the video. Highlight the passages that matter to you. Form your own questions and initial understanding first.

Why this works: Research on active recall shows that the effort of forming questions and identifying key points is itself a learning activity. Skipping this step and going straight to AI summaries removes the most valuable part of the process.

E - Engage, Don't Extract

When you do use AI, use it as a thinking partner, not an answer machine.

• Instead of: "Summarize this article for me"

• Try: "I highlighted these key points from the article. What am I missing? What counterarguments exist?"

• Instead of: "Explain quantum computing"

• Try: "Here's my understanding of quantum computing [your notes]. Where am I wrong? What would deepen my understanding?"

A - Annotate and Connect

Don't let AI-generated insights exist in isolation. Connect them to what you already know.

• Add your own notes alongside AI explanations
• Create connections between new information and existing knowledge
• Share your annotated insights with others to reinforce understanding

This is where tools like Glasp become critical. When you highlight and annotate while reading, you're performing active processing that AI summaries alone cannot replicate. Your highlights become a map of your thinking, not AI's.

R - Retrieve and Verify

Periodically test yourself on material you've learned with AI assistance. If you can't explain a concept without AI, you haven't learned it.

• Re-read your highlights without the original context
• Try to explain concepts in your own words
• Use AI to quiz you, not to answer for you

Putting It All Together

---

The Role of Highlighting and Annotation in the AI Era

There's an irony in the AI learning debate: the most effective way to use AI for learning involves the oldest study technique there is: highlighting and note-taking.

Why Highlighting Still Matters

The research on cognitive offloading reveals that the problem isn't AI itself. It's the passivity it enables. When you read an article and immediately ask ChatGPT to summarize it, you skip the most valuable cognitive step: deciding what matters.

Highlighting forces you to make that decision. Every time you select a passage, you're asking yourself: Is this important? Why? How does it relate to what I already know?

This is active processing: the exact cognitive activity that AI-dependent learners skip and that research shows is essential for retention.

The Social Learning Advantage

The MIT study found that ChatGPT users produced "near-identical essays lacking original thought." This raises an important question: if everyone is learning from the same AI, where does originality come from?

The answer lies in diverse perspectives. When you see what other people highlight in the same article (what they found important, what questions they asked, what connections they made), you're exposed to thinking patterns that AI cannot generate, because they come from real human experience.

This is why social highlighting platforms create a learning environment that AI alone cannot replicate:

• You see multiple interpretations of the same content
• You discover insights you would have missed
• You engage with a community of learners, not just an algorithm
• Your highlights become part of a collective knowledge base that benefits everyone

AI Summaries + Human Highlights: The Best of Both Worlds

The ideal workflow isn't AI or human annotation. It's both:

1. Watch a YouTube video and use AI to get a quick summary and transcript
2. Read through the transcript and highlight the parts that resonate with your specific needs
3. Add your own notes explaining why those passages matter to you
4. Share your highlights so others can benefit from your perspective
5. Review your highlights periodically to reinforce learning

This workflow combines the efficiency of AI (fast summarization, transcript generation) with the effectiveness of human engagement (active selection, personal annotation, social sharing).

---

Frequently Asked Questions

Is AI making students dumber?

Not exactly. The research shows that unrestricted AI access can reduce learning outcomes (Wharton: 17% worse on exams), but guided AI tutoring can match or exceed traditional methods. The variable isn't intelligence; it's how AI is used. Students who use AI to shortcut thinking learn less. Students who use AI to deepen thinking learn more.

Should schools ban ChatGPT and similar AI tools?

The research suggests bans are counterproductive. 84% of high school students already use generative AI for schoolwork. Instead of banning, the Wharton study demonstrates that designing guardrails (hints instead of answers, guided questioning instead of direct responses) produces dramatically better outcomes. The goal should be teaching students how to use AI as a learning tool, not an answer machine.

How does AI affect critical thinking?

A study of 666 participants (Gerlich, 2025) found a significant negative correlation between frequent AI use and critical thinking. However, this appears to be mediated by cognitive offloading: the more people outsource thinking to AI, the less they practice thinking independently. The solution isn't avoiding AI but maintaining deliberate critical thinking practices alongside AI use.

Can AI replace human tutors?

For certain types of instruction, AI tutoring is remarkably effective. Dartmouth's 2025 study showed AI tutoring outperformed in-class active learning for medical students. However, AI tutoring works best when it's designed to teach (scaffolded questions, guided reasoning) rather than simply answer questions. Human tutors still excel at emotional support, motivation, and handling novel situations outside AI's capabilities.

What's the best way to use AI for studying?

Research points to the mastery approach: use AI to construct and augment knowledge, not to complete tasks mechanically. Specifically:

• Read and highlight material first, form your own understanding
• Use AI to challenge your thinking, not replace it
• Ask AI for alternative perspectives, counterarguments, and deeper explanations
• Test yourself on material without AI access
• Share and discuss your learning with others

Will AI eventually make traditional learning obsolete?

Current evidence suggests the opposite. AI highlights the irreplaceable value of active cognitive engagement: reading deeply, thinking critically, forming connections, and struggling productively with difficult material. AI tools are most powerful when combined with these traditional learning practices, not when they replace them.

---

Conclusion: How to Coexist with AI as a Learner

The research paints a clear picture: AI is the most powerful learning tool ever created, and also the most dangerous one. It amplifies whatever approach you bring to it. If you approach learning actively, AI accelerates your growth. If you approach it passively, AI accelerates your decline.

Here's what we know for certain:

• The "crutch" model fails. Unrestricted AI access without engagement leads to measurable declines in learning, critical thinking, and even brain connectivity.
• The "tutor" model works. AI designed to guide, question, and scaffold (rather than simply answer) produces outcomes equal to or better than traditional methods.
• Active engagement is non-negotiable. Reading, highlighting, annotating, and connecting ideas are not outdated practices. They are the cognitive foundations that make AI useful rather than harmful.
• Social learning adds a dimension AI can't. Seeing how others think about the same material exposes you to perspectives that no algorithm can generate.

The question is no longer whether we'll use AI. 84% of students already do. The question is whether we'll use it in ways that make us more capable, or less.

The answer starts with how you read, highlight, and think.

---

Want to build better learning habits in the age of AI? Glasp helps you highlight articles and YouTube videos, add your own notes, and learn from what others are reading. Your highlights become your digital knowledge base: a record of your thinking that no AI can replicate.