Is AI-generated learning content accurate?

For foundational, well-established topics, generally yes, since the underlying model has seen enormous amounts of accurate material. For niche, highly technical, contested, or rapidly evolving topics, hallucination risk is higher, and a language model can state an incorrect answer with the same confident tone as a correct one, which makes independent verification important for anything consequential.

What is the difference between AI explaining something and AI actually teaching it?

Explanation alone produces weaker retention than explanation paired with retrieval practice. Research on the testing effect found actively recalling information produces roughly 50% better retention than passive re-reading, regardless of whether the source was a textbook, a lecturer, or an AI. A chatbot answering questions conversationally is weaker for actual learning than a structured course with quizzes built in.

What can't AI teach you effectively?

Three categories in particular: certification-level technical mastery, which requires sustained engagement beyond any single session; hands-on physical skills like playing an instrument or a surgical technique, which need real-time physical feedback; and any high-stakes information where a confidently wrong answer needs independent verification before acting on it.

Does it matter if AI doesn't truly understand what it's teaching?

Not as much as it might seem. Learning outcomes depend on whether content is well-structured, engages retrieval rather than passive reading, and respects working memory limits, not on whether the source has subjective understanding. A well-structured AI-generated lesson with a quiz satisfies these conditions the same way a well-designed human-taught lesson does.

Can AI Really Teach You Anything? An Honest Answer

Q: Can AI actually teach you something new?

Yes, for foundational and intermediate understanding of most well-documented subjects. A 2025 Harvard study found students using AI tutors learned more than twice as much in less time compared to traditional classrooms, and a 2026 meta-analysis of 58,702 participants confirmed a consistent positive effect across studies.

Yes, for a specific and genuinely large category of learning. No, for a smaller but important category most people underestimate. The honest answer isn't a single word, and anyone giving you one is oversimplifying in one direction or the other.

The skepticism about AI teaching effectively is reasonable and worth taking seriously before dismissing it, not the reflexive "AI is just autocomplete" objection, but the more specific concern: can something that doesn't actually understand a subject teach it to you reliably? This post covers what the actual research says, where it holds up, and where the honest limits are.

Key Takeaways
A 2025 Harvard study found students using AI tutors learned more than twice as much in less time compared to traditional active-learning classrooms
A 2026 meta-analysis of 58,702 participants found a consistent, statistically significant positive effect of AI on learning outcomes across studies
AI teaches foundational and intermediate understanding well, particularly for well-documented subjects, and is weakest for highly technical, contested, or safety-critical information
The mechanism that actually works isn't the AI explaining things, it's AI-generated content paired with retrieval practice, the explaining alone does less than most people assume
The right question isn't "can AI teach," it's "for what, and paired with what verification"

What Does the Research Actually Show?

Skepticism about AI teaching effectiveness was reasonable to hold for a long time. It's gotten harder to hold in its strongest form since the research has caught up.

A June 2025 study published in Scientific Reports (Harvard University) compared students using AI tutors against students in traditional active-learning classrooms on the same material. The AI tutor group learned more than twice as much in less time, with effect sizes between 0.73 and 1.3 standard deviations, a genuinely large effect in education research, where anything above 0.4 is typically considered meaningful.

That's one study, however well-designed. A 2026 second-order meta-analysis in the Journal of Educational Computing Research synthesized 19 separate meta-analyses covering 58,702 total participants and found a statistically significant mean effect size of 0.67 for AI technologies on student learning outcomes. This is the most comprehensive analysis of the question published to date, and it points the same direction as the Harvard study: AI-assisted and AI-generated learning produces real, measurable gains, not a marginal or inconsistent effect.

There's an honest caveat worth including. A 2025 meta-analysis in Frontiers in Psychology found that while AI in blended learning produced a significant positive effect in short interventions, the effect size dropped substantially over longer interventions, one semester to one year. Some of the early gains likely reflect novelty, students engaging more actively with something new performing better, an effect that fades once the novelty does. The honest read: AI teaching works, particularly for getting started and building foundational understanding, and its advantage over traditional methods likely narrows over very long timeframes without good design underneath it.

Why Does the "It's Just Autocomplete" Objection Miss the Point?

The reflexive dismissal, that a language model doesn't actually understand anything so it can't really teach, treats teaching as requiring the teacher to hold understanding the way a person does. But teaching effectiveness and the philosophical question of whether a system "understands" are separable questions, and the research measures the former directly.

What actually determines whether a learner retains something isn't whether the source understands the material in some deep sense. It's whether the content is structured well, whether it engages the learner in retrieval rather than passive reception, and whether the sequencing respects working memory limits. Cognitive load theory and the testing effect both predict which teaching methods work regardless of whether the teacher is a human, a textbook, or a language model. A well-structured AI-generated lesson with a quiz at the end satisfies these conditions the same way a well-designed human-taught lesson does. The mechanism that produces learning doesn't require the source to have subjective understanding, it requires the content to be structured in a way that matches how human memory and attention actually work.

Where Does AI Genuinely Fall Short?

Being fair about the limits matters as much as being fair about the strengths.

Hallucination and Confident Inaccuracy

Language models generate plausible-sounding content whether or not it's accurate, and this risk is highest for niche, highly technical, contested, or rapidly evolving topics, exactly the areas where a wrong answer stated confidently is hardest for a non-expert to catch. This is different from a human expert being occasionally wrong. A human expert who doesn't know something usually signals uncertainty. A language model can generate an incorrect answer with the same confident tone as a correct one.

Hands-On Physical Skills

No amount of AI-generated explanation teaches you to actually play an instrument, perform a surgical technique, or throw a proper punch. These require physical practice with real-time feedback that text-based or even video-based instruction, AI-generated or human-generated, can't fully replace.

Deep Expertise Built Over Years

Deep expertise, the kind that lets someone generate new knowledge rather than reproduce existing knowledge, still requires sustained engagement well beyond what any single learning session, AI-assisted or not, provides. AI-generated content is genuinely strong for foundational and intermediate understanding. It doesn't compress years of deliberate practice into 30 minutes, and treating it as though it does sets up a disappointing outcome.

What Part Do Most People Get Backwards?

The common framing is "does the AI explain things well." That's the wrong question, or at least an incomplete one.

The research consistently points to a different mechanism as the actual driver of learning outcomes: retrieval practice. Roediger and Karpicke (2006) found that actively recalling information, being tested on it, produces substantially better retention than passive re-reading, roughly 50% better at the same total study time. This holds regardless of whether the original content came from a textbook, a human lecturer, or an AI-generated lesson.

This means the honest answer to "can AI teach you anything" depends heavily on whether the AI-generated content includes retrieval practice, not just explanation. An AI chatbot answering questions in a conversational format, with no structured follow-up testing whether you actually retained the answer, is weaker than an AI-generated course with quizzes built into the structure. The explaining is necessary but not sufficient. The mechanism that converts exposure into retention is the retrieval step, and that's a design choice, not an inherent property of AI-generated content. Spaced review across multiple sessions strengthens this further, since a single pass through even well-structured content produces weaker long-term retention than the same content revisited over several days.

So, Can It?

Yes, for building foundational and intermediate understanding of most well-documented subjects, provided the format includes active retrieval rather than passive explanation alone. The Harvard and 2026 meta-analysis data both support this directly, and the mechanism, structured content plus retrieval practice, is well understood independent of whether AI generated the content.

No, for certification-level technical mastery, hands-on physical skills, and anything where a confidently wrong answer carries real consequences without independent verification. These aren't edge cases to dismiss, they're categories where the honest answer is genuinely different, and treating AI as a universal substitute for expertise in these areas would be overselling it.

AI teaches this well	AI doesn't replace this
Foundational understanding of a new topic	Certification-level technical mastery
Well-documented, established subjects	Hands-on physical skills (instrument, surgery, sport)
Curiosity-driven, exploratory learning	Deep expertise built over years of practice
Quick orientation before a decision	High-stakes information needing independent verification

Morso is built around the version of this that actually works: structured courses with retrieval-based quizzes on any topic, generated in about 30 seconds, positioned for the foundational and curiosity-driven learning where the research supports AI genuinely delivering. For a fuller comparison of tools built for this specific use case, what is an AI course generator covers the category directly. For a broader look at where AI is delivering real results in education more generally, how AI is changing education covers the wider picture.

Sources

Henkel, L.A. et al. "Generative AI tutoring improves learning outcomes in a classroom study." Scientific Reports, June 2025. https://www.nature.com/articles/s41598-025-97652-6
Journal of Educational Computing Research. Second-order meta-analysis of 19 meta-analyses. 2026. https://journals.sagepub.com/doi/10.1177/07356331261424767
Frontiers in Psychology. "AI in Blended Learning Meta-Analysis." 2025. https://www.frontiersin.org/journals/psychology/articles/10.3389/fpsyg.2025.1691414/full
Roediger, H.L. & Karpicke, J.D. "Test-enhanced learning: Taking memory tests improves long-term retention." Psychological Science, 17(3):249-255. 2006. https://doi.org/10.1111/j.1467-9280.2006.01693.x
Sweller, J. "Cognitive load during problem solving: Effects on learning." Cognitive Science, 12(2):257-285. 1988.