Skip to main content
Research Base

The evidence is not
aspirational.

CLARO's design is grounded in peer-reviewed research. Every architectural decision β€” Socratic questioning, fading scaffolding, no answer generation, bilingual dialogue β€” is a direct response to what the evidence says about how AI affects learning.

7objections
addressed
6key studies
reviewed
Ξ·Β²=.32largest Socratic AI
effect size found

CLARO's Research Position

The question is not "Is AI good or bad for learning?" The question is "What does this specific AI do when a student encounters a hard idea?"

The research literature on AI and learning is not monolithic. It contains strong evidence for harm (cognitive offloading, homogenization, dependency) and strong evidence for benefit (Socratic tutoring, fading scaffolding, one-to-one dialogue). These findings are not in conflict β€” they describe different tools operating on different principles.

Answer-generating AI does the thinking for the student. Socratic AI requires the student to do the thinking. CLARO is built on the second principle β€” and every design decision flows from that distinction.

Objections & Responses

Every serious objection, answered.

We take every objection seriously. Some are valid concerns about AI in general. Some are category errors. All deserve a research-grounded response.

Key Studies

The studies that shaped CLARO's design.

These are not cherry-picked endorsements. They include studies that document real harms from AI β€” and explain why CLARO is designed the way it is.

CRITICAL2025

The "Cognitive Debt" Study

Kosmyna et al. Β· MIT Media Lab

Finding: ChatGPT users showed the weakest brain connectivity of any group in EEG analysis. Deficit persisted after tool use ended. NLP analysis revealed homogenized essay output in AI-assisted group.

CLARO implication: Answer-generating AI causes measurable neurological and cognitive harm. Socratic AI inverts this β€” it is the antidote, not the cause.

FOUNDATIONAL2025

AI Tools & Critical Thinking

Gerlich Β· Societies Journal

Finding: n=666 study found significant negative correlation between AI tool use and critical thinking scores across age groups.

CLARO implication: Passive AI use degrades critical thinking. Active Socratic engagement with AI β€” where the student does the reasoning β€” is the structural remedy.

Societies 15(1):6
SUPPORTING2025

Socratic Math Tutor Study

MIT RAISE Β· MIT

Finding: Socratic AI math tutor produced large learning gains (partial Ξ·Β²=.20) in K-12 settings.

CLARO implication: Socratic AI produces among the largest intervention effects in recent education research β€” directly contradicting the cognitive offloading narrative.

MIT RAISE (2025)
SUPPORTING2025

K-12 Science Socratic AI Study

Research Square Β· Research Square

Finding: Socratic AI in K-12 science classrooms produced very large effect size (partial Ξ·Β²=.32) β€” among the largest intervention effects in recent education research.

CLARO implication: Effect sizes of this magnitude are rare in education research. Socratic AI is not merely "not harmful" β€” it is actively beneficial at a level that exceeds most classroom interventions.

Research Square (2025)
SUPPORTING2025

Socratic Design & Epistemic Agency

Degen Β· arXiv

Finding: Socratic AI design preserves epistemic agency; fading scaffolding structures transfer the questioning role to the student over time.

CLARO implication: Design β€” not technology category β€” determines whether AI builds or erodes student agency. Fading scaffolding is the key mechanism.

FOUNDATIONAL1984

The 2 Sigma Problem

Bloom Β· Educational Researcher

Finding: One-to-one tutoring produces +2 standard deviation learning gains over classroom instruction β€” the most powerful intervention in education research.

CLARO implication: CLARO scales the most powerful known educational intervention to every student, regardless of family income. This is the equity argument for Socratic AI.

Educational Researcher, 13(6), 4–16
Research β†’ Design

How the evidence shaped every decision.

🧠

Research finding: Answer-AI causes cognitive debt (MIT, 2025)

CLARO never generates answers β€” only questions.

πŸ“‰

Research finding: Scaffolding addiction is a real risk (Lee et al., 2025)

Fading scaffolding: questioning intensity decreases as student independence grows.

🎯

Research finding: One-to-one tutoring = +2Οƒ (Bloom, 1984)

Every student gets a private Socratic dialogue partner, regardless of resources.

🌿

Research finding: AI homogenizes output (MIT NLP analysis, 2025)

No content generation means no homogenization β€” student ideas stay the student's.

πŸ”’

Research finding: General chatbots hallucinate 58–82% (Stanford HAI, 2025)

Bounded curriculum-aligned knowledge base; tool asks questions rather than asserting facts.

⚑

Research finding: Socratic AI produces Ξ·Β²=.32 effect (Research Square, 2025)

Socratic method is the core engine β€” not a feature, not a mode, not optional.

See the research in action.

Every session with Chispa is a live demonstration of what the evidence says should happen when a student encounters a hard idea.