Access Is Not Enough: Why AI Tutoring Needs Human Support to Work for All Students

Access Is Not Enough: Why AI Tutoring Needs Human Support to Work for All Students

New research on implementation barriers, inclusive edtech adoption, and accessible AI learning platforms.

schools are investing in AI tutoring platforms with the expectation that access solves the problem. it doesn't. a new Stanford study of over 350 elementary students reveals a hard truth: nearly half never used the platform at all, even with dedicated time. among those who did? the students least likely to benefit—those receiving special education services—were the least likely to engage.

this research matters for anyone deploying edtech, designing inclusive learning environments, or evaluating AI tools for struggling readers.

The Access Paradox: Why Giving Students Tools Isn't Enough

access is a necessary precondition for technology-enabled learning. it's not sufficient.

the Stanford researchers ran two randomized controlled trials with elementary students across two school districts. the setup was straightforward: students were given access to an AI literacy platform with scheduled time to use it. the results were striking.

the baseline problem:

• 60.7% of students in District A used the platform at least once
• 53.3% of students in District B used the platform at least once
• 46–47% never used it, even once
• students who did use it averaged only 2–5 minutes per week
• the platform provider recommends 30 minutes per week for measurable reading gains

this is the implementation barrier in one sentence: access alone does not generate engagement. students will not use tools they lack motivation, accountability, or relationship around.

The Equity Gap Nobody Talks About: Special Education and Inclusive Edtech

here's the finding that should reshape how schools think about inclusive edtech deployment.

students with special education needs were significantly less likely to use the AI platform:

• 11% of special ed students engaged with the tool (District B)
• 35% of non-special ed students engaged (District B)
• the gap was consistent across both districts

this is not a coincidence. this is a design and implementation problem.

why this matters for accessibility:

the students who might benefit most from additional, personalized reading practice—struggling readers, students with learning disabilities, neurodivergent learners—are the least likely to access it independently. if we rely on student motivation or individual initiative to drive platform use, we will systematically exclude the students with the greatest need.

this is why "universal design for learning" and "belonging-first" approaches aren't optional extras. they're foundational to equitable edtech. a platform designed for struggling students needs to be adopted with struggling students' needs in mind: explicit relationship support, clear accountability structures, and adult accompaniment.

Human Support Changes Everything—But Not How You'd Expect

the researchers tested a hybrid model: students worked with a human tutor focused on engagement, not direct instruction.

the tutor's job was to:

• build relationship and trust (5–10 min check-in per session)
• support platform use (troubleshoot tech, encourage persistence)
• establish norms and accountability
• reflect on effort and growth (2–5 min at session end)

this meant students in the "human support" group had only 15 minutes per session on the platform itself, compared to the control group's full 30 minutes.

you'd expect human support to take away from learning time. instead, it multiplied engagement.

the results:

• 71% increase in engagement (stories read per week) in District A
• 80% increase in engagement in District B
• 1–4 additional minutes per week of platform use
• human tutors increased usage despite reducing platform time

the pedagogical insight: motivation, accountability, and relational connection remain the primary drivers of student participation in learning, even when instruction is delivered by AI. the human's role isn't to deliver content. it's to make the student want to show up.

Why Engagement Didn't Translate to Reading Gains (And What That Tells Us)

here's where the research gets sobering.

despite significantly higher engagement with the human support model, there were no measurable improvements in reading achievement in either district.

this isn't a failure of the intervention. it's evidence of a dosage problem.

in District A, students in the treatment group got an average of one additional minute per week on the platform—roughly 22 extra minutes across the entire intervention period. in District B, it was better but still insufficient: approximately 98 additional minutes total.

the platform requires ~30 minutes per week of consistent use for measurable reading gains. even with human support, students didn't reach that threshold.

the implication: relational support alone can't overcome weak implementation. you need both.

Implementation Varies Wildly—And That Matters More Than Tool Quality

one detail buried in Table A8 should change how schools evaluate edtech.

the same intervention, same tutors, same platform, different sites achieved wildly different results:

• one site: +5.5 additional minutes per week for treatment students
• another site: −1.4 minutes (treatment actually performed worse)
• the highest-performing site: +9.1 minutes per week

same study design. different outcomes.

this is the implementation story. tool quality ≠ outcomes. school culture, tutor expertise, leadership commitment, student cohort, and local context matter as much (or more) than the platform itself.

for inclusive edtech adoption, this means:

• audit your implementation readiness before you buy the tool
• train and support the humans in the loop
• expect variance by site; diagnose it
• don't blame the tool when implementation fails

The Spectrum of Relational Intensity: Where AI Fits in Personalized Learning

the researchers frame this beautifully: personalized instruction exists on a spectrum of relational intensity.

on one end: fully AI-led (scalable, low cost, low relational intensity, low uptake among struggling students).

on the other end: fully human-led 1:1 tutoring (expensive, high relational intensity, proven effective, not scalable).

in the middle: hybrid models where humans focus on relationship, motivation, and troubleshooting while AI delivers the instructional content.

the implication for inclusive pedagogy:

AI doesn't replace relationship. it supplements it. and relationship remains the load-bearing wall.

schools building inclusive edtech stacks need to ask: where on this spectrum are we sitting, and is that sustainable? a fully AI model might reach 60% of students. a hybrid model might reach 75%. but the hybrid model requires staff time, training, and coordination.

the tradeoff is real. the question is whether your school is willing to make it.

What This Research Means for Your School's AI Tutoring Strategy

if you're considering AI tutoring:

• audit your implementation readiness first. tool choice comes second.
• plan explicit human support for special ed students and struggling readers. they won't opt in on their own.
• set dosage targets (30 min/week is the research baseline). measure whether you hit them. if not, diagnose why.
• expect site-level variation. some implementations will work; others won't. be ready to adapt.

if you're already deployed:

• check your usage data by student subgroup. are struggling students using it? special ed students? English learners? if not, something in your implementation is failing.
• train tutors/staff on relationship-building and motivation, not just content delivery.
• set a minimum engagement threshold and intervene when students fall below it.

if you're designing curriculum or professional development:

• help teachers and staff understand the relational intensity spectrum. they need to know that access is not enough.
• build implementation readiness into your edtech evaluation rubrics, not just tool features.
• center the voices of students with disabilities and struggling readers in your deployment planning.

the bottom line: AI tutoring platforms offer real promise for scaling personalized instruction. but that promise only materializes if students actually use them. and students—especially those with the greatest need—are far less likely to use tools they lack motivation, relationship, and accountability around.

implementation and human support matter as much as tool quality. this is why accessibility and inclusive design aren't afterthoughts. they're the foundation of effective edtech adoption.

Source

Robinson, Carly D., David Gormley, Ana Trindade Ribeiro, and Susanna Loeb. (2026). Access is Not Enough: Human Support Improves Engagement with AI Tutoring. Stanford SCALE Initiative. https://doi.org/10.26300/pz7p-p388