High-impact tutoring has been heralded as an evidenced-backed solution for addressing the significant academic gaps that schools are facing. But scaling and sustaining these programs are challenging for schools because they require a lot of human tutors, time, and money.
The rise of artificial intelligence-based tutoring software has given schools a way to expand access, lower costs, and improve personalization of a key learning-recovery approach.
However, research on the effectiveness of AI tutoring models is limited and still emerging.
To help educators navigate the AI tutoring landscape, the National Student Support Accelerator and the AI Hub for Education at Stanford’s SCALE Initiative recently published a research brief examining the current research on the effectiveness of using AI in tutoring.
As schools deal with an uncertain funding environment, they might consider using AI tutoring to reduce costs, and “we wanted to make a more actionable tool for education system leaders on what the evidence says about AI tutoring,” said Chris Agnew, the director of the AI Hub for Education, which researches AI’s impact on student learning.
Part of the challenge with researching the effectiveness of AI tutoring is that the term could refer to a wide range of tutoring models, Agnew said.
“People hear the term ‘AI tutor,’ and they think that is a young person engaging directly with a chatbot,” he said. “What our brief outlines is that AI tutors exist on an AI-human spectrum, and that spectrum exists on a range of relational intensity. Depending on where you fall on that range, there is either strong evidence to support [it] or very little evidence.”
According to the research brief, evaluations of AI tutoring tools suggest that:
- Many students don’t engage with AI tutors.
- Adding human oversight helps engagement, but not achievement.
- The strongest results of effectiveness so far come from AI tools built for human tutors to use, not for student use.
Level of human involvement is key factor when considering AI-assisted tutoring tools
The brief categorizes tutoring models by their level of human involvement, then maps them against existing evidence:
- In-person or remote tutoring: This is when a human tutor is responsible for all instruction and student interaction. There is robust evidence that this can be a highly effective tutoring model.
- Human tutoring with AI support: This is when a human tutor is responsible for interacting with the student, and an AI tool assists the tutor behind the scenes. Research on this model is still emerging but indicates it could be as effective or more effective than human-only tutoring.
- AI tutoring with human support: This is when a student engages directly with AI, while a human tutor oversees, guides, and intervenes when needed. Research on this model is still emerging but indicates implementation determines effectiveness.
- AI-only tutoring: This is when a student works directly with AI for all aspects of the tutoring experience without direct human oversight. There isn’t a lot of research on this model; a study found that when left to work independently, 40-47% of students never used the AI platform, making it challenging to gauge effectiveness.
“Tutoring research tells us that relationships are a key part in student persistence, student engagement, and hitting dosage that allow sustained improvement in student outcomes,” Agnew said. “Access to a tool is not enough.”
Sustained tutoring relationships can enhance student motivation and allow for more effective personalization, said Matthew Kraft, a professor of education and economics at Brown University who was not involved in the Stanford research brief.
“However, we also know personalized learning can be successful without that element,” he said in an email. “Evidence from more structured computer-adaptive learning software programs is promising when schools can structure implementation successfully.”
The brief also outlines AI tutoring use cases that enhance human-led instruction and augment educators’ capacity. Some examples include using AI to:
- Streamline master scheduling and allocation of students and educators, which can reduce schedule-building time and preserve core instructional time and intervention blocks;
- Enhance tutor training and professional development through realistic practice simulations that allow tutors to hone their instructional techniques before leading live sessions; and
- Generate targeted student practice materials with a tutor reviewing them before implementation.
Before adopting AI tools in tutoring, Stanford researchers recommend school, district, or state education leaders prioritize apps that augment human-led instruction and operational capacity and establish strict student data-privacy safeguards.
AI is not a “fix-all,” Kraft cautioned. “AI cannot ensure you meet the basic elements of successful tutoring—high attendance, sustained tutoring over time, and strong relationships.”
