Technology

How AI and LLMs Are Transforming Course Learning

2025 research shows AI tutoring can outperform classroom instruction—but only when designed right. Here's what the evidence actually tells us.

LMSMore TeamUpdated January 19, 202516 min read
AI and artificial intelligence concept - representing AI-enhanced learning
Photo by Merakist on Unsplash

In June 2025, researchers at Harvard published a finding that sent shockwaves through education: AI tutoring didn't just match classroom learning—it crushed it. Students learned more in 49 minutes with an AI tutor than in 60 minutes of active learning with experienced instructors. The effect size? Up to 1.3 standard deviations.

For context, educational researchers consider anything above 0.4 standard deviations "educationally significant." This study hit triple that.

For 40 years, Benjamin Bloom's "2 sigma problem" has haunted education. His 1984 research showed that one-on-one tutoring produced learning gains of two standard deviations—but the cost made it impossible to scale. Educators have been searching for a solution ever since. We may finally have found one. But the story is more nuanced than the headlines suggest.

This article examines what 2025's research actually tells us about AI in learning—the breakthroughs, the caveats, and what it means for L&D professionals building the next generation of training.

1.3σ
Effect Size
Harvard RCT: AI tutoring vs active learning
Nature Scientific Reports, 2025
92%
Student Adoption
Students using AI for learning (up from 66% in 2024)
DemandSage, 2026
49 min
Time to Learn
AI group vs 60 min classroom (18% faster)
Harvard RCT, 2025
43%
Multi-Model Users
Employees using 2+ LLMs for different tasks
Harvard Business Review, 2025

The Harvard Breakthrough: What the Study Actually Found

The Harvard randomized controlled trial employed a crossover design with 194 undergraduate physics students. Each student served as their own control, experiencing both AI tutoring and classroom active learning. The methodology was rigorous—and the results were striking.

Learning Outcomes

Effect sizes of 0.73 to 1.3 standard deviations—far exceeding the 0.4 threshold for educational significance.

Time Efficiency

Median AI group finished in 49 minutes vs. 60 minutes for classroom—18% time savings with better results.

Engagement

Students reported engagement scores of 4.1/5 for AI tutoring vs. 3.6/5 for classroom learning.

Critical Insight

The AI tutor was designed using the same pedagogical best practices as classroom instruction—design mattered.

Source: Kestin et al. (2025), "AI tutoring outperforms in-class active learning," Nature Scientific Reports

Why AI Tutoring Works (When It Works)

The Harvard study's success wasn't accidental. Here's what the research tells us about why well-designed AI tutoring produces results:

The Bloom Effect at Scale

One-on-one attention without one-on-one cost. Patient, judgment-free repetition that adapts to each learner's pace and knowledge gaps—the tutoring approach Bloom proved worked, now available to everyone.

Harvard RCT showed 0.73-1.3 sigma improvement over classroom instruction

Immediate, Contextual Feedback

The Harvard study found AI tutoring delivered feedback while thinking was still fresh—a critical factor. Traditional LMS feedback often arrives days later, after the mental context has faded.

Students reported engagement scores of 4.1/5 vs 3.6/5 for classroom learning

Dialogue Over Broadcast

Traditional e-learning is broadcast: one-way content delivery. AI enables conversational learning that mirrors how humans naturally learn from each other—through questions, discussion, and exploration.

Conversational AI mimics the Socratic method at scale

Time Compression

Same learning outcomes in less time. The Harvard study showed median AI learners finished in 49 minutes what took classroom students 60 minutes—an 18% efficiency gain with better results.

18% time savings with superior learning outcomes (Harvard RCT)

The Uncomfortable Findings

The research isn't all good news. Several studies have identified significant risks that L&D professionals must address:

Cognitive Offloading

Zhai et al., 2024

Regular AI dialogue use associated with decline in cognitive abilities and information retention. Students let AI do the thinking instead of struggling productively.

Implication

Design must require learner effort before AI assistance

Design Trumps Technology

Education Next, 2025

Harvard's AI worked because it followed pedagogical best practices. Generic ChatGPT conversations don't produce the same results—LLMs alone 'quickly get lost when trying to teach common math concepts.'

Implication

AI + knowledge base + instructional design = results

Personal vs Personalized

Psychology Today, 2023

AI offers personalization, not personal connection. Critics note that 'only a human can see students and engage with them in the context of a relationship.'

Implication

Position AI as complement to human instruction, not replacement

Assessment Integrity

Industry Analysis

If learners can use AI to complete assessments, what are we actually measuring? Traditional recall-based testing becomes meaningless.

Implication

Shift to process-based, synchronous, application-focused evaluation

What Corporate L&D Should Actually Do

Based on the research, here's a phased approach to implementing AI-enhanced learning that maximizes benefits while mitigating risks:

1

Foundation

Months 1-2

  • Audit current content for AI-readiness
  • Connect your CMS (Contentful/Sanity) to enable grounded AI responses
  • Establish baseline metrics for comparison
2

Pilot

Months 2-4

  • Start with low-stakes use cases: pre-learning prep, clarification during modules
  • Design AI interactions that guide rather than answer directly
  • Implement 'attempt first' requirements before AI assistance
3

Measure

Ongoing

  • Track learning outcomes, not just engagement
  • Monitor for cognitive offloading patterns
  • Compare time-to-competency vs traditional approaches
4

Scale

Based on Evidence

  • Expand what works; cut what doesn't
  • Build organizational capability in prompt engineering
  • Train facilitators to work alongside AI, not compete with it

The LMSMore Approach

Based on this research, we've built LMSMore with specific principles in mind:

Grounded AI responses

Connect to your Contentful or Sanity CMS—not hallucination-prone general knowledge

Designed for productive struggle

AI guides rather than answers directly—preventing cognitive offloading

Outcome-focused metrics

Track learning outcomes, not just engagement or completion rates

Human-AI complementarity

Position AI to free up instructors for high-value interactions

What Remains Irreplaceably Human

As we integrate AI into learning, it's worth being clear about what it cannot—and perhaps should not—replace:

Motivation Through Relationship

Learning is social. Accountability to humans drives persistence. AI can inform, but humans inspire. Cohort dynamics and peer learning remain irreplaceable.

Tacit Knowledge Transfer

Experts know things they've never written down. Observation, mentorship, shared practice—AI can only access what's been documented.

Ethical Development

AI can present ethical frameworks, but moral reasoning requires human guidance and lived experience. Judgment, wisdom, discernment develop through human relationships.

Emotional Resilience

Learning is often emotionally challenging. AI patience isn't human empathy. Struggle is part of growth; humans help us persist through it.

The goal isn't AI instead of humans, but AI that frees humans to do what only humans can do.

Looking Forward

The Harvard study represents a genuine milestone—but it's not the end of the story. The researchers themselves emphasize that their AI tutor worked because it was carefully designed around pedagogical best practices, not because AI is inherently superior to human instruction.

For L&D professionals, the message is clear: the technology is ready, but the implementation matters enormously. Generic AI chatbots won't produce Harvard's results. Thoughtfully designed AI learning experiences, grounded in verified content and built around how humans actually learn, can.

The 2 sigma problem may not be fully solved. But for the first time, we have tools that can meaningfully close the gap—if we use them wisely.

Sources & Further Reading

Research Studies

Ready to Build AI-Enhanced Learning That Actually Works?

LMSMore connects your Contentful or Sanity content to intelligent learning experiences. Ground AI in your verified content—the approach the research shows actually produces results.