These approaches should be compared by the support the family needs, not by a blanket claim that one is always superior. Stanford’s November 2026 evidence review reports mixed learning effects and warns that positive experiences with AI do not necessarily translate into stronger learning outcomes. It also cautions against broad generalization across US K–12 settings. (scale.stanford.edu)
Explore the comparison in this briefing
Route
Human mentor
- What the household should expect to supply
- Goals, consent, scheduling decisions and periodic review
- When we would consider it
- The child needs explanation, project scoping or sustained feedback
- Condition for continuing
- The mentor can show growing learner independence
Compare each option against the same criteria. The complete table and qualifications remain in the briefing.
Our recommended division of responsibilities is:
| Route | What the household should expect to supply | When we would consider it | Condition for continuing |
|---|---|---|---|
| Human mentor | Goals, consent, scheduling decisions and periodic review | The child needs explanation, project scoping or sustained feedback | The mentor can show growing learner independence |
| AI tutoring software | Product selection, eligibility checks, supervision and progress review | Bounded practice within a suitable learning workflow | The learner can subsequently perform a related task without the tool |
| Parent-led curriculum | Preparation, explanation, encouragement and technical troubleshooting | A parent genuinely wants and can sustain that role | The arrangement remains enjoyable and manageable for both |
| Hybrid | Clear boundaries between human teaching and software-supported practice | The family wants human judgment with additional between-session practice | The mentor reviews reasoning, not merely completed exercises |
A 2025 randomized study in Turkish high-school mathematics found that unrestricted GPT-4-style assistance improved supported practice but reduced subsequent unassisted performance; a guarded tutoring version largely mitigated that negative effect. This is evidence about a particular mathematics setting—not a comparison of the private AI-mentorship providers in this report. (pmc.ncbi.nlm.nih.gov)
Purchasing implication: Require both supported work and a small independent transfer task. Do not use project polish or the learner’s enthusiasm as the only renewal evidence.
Sources cited on this page
- Al Hub for Education ↗scale.stanford.edu
- Generative AI without guardrails can harm learning: Evidence from high school mathematics - PMC ↗pmc.ncbi.nlm.nih.gov