Stanford research confirms: AI tutoring only works with humans at the helm

Stanford SCALE published a research article this week with a clear message for everyone involved with AI in education. AI tutoring is not a monolith. The way in which a pupil interacts with AI determines whether it leads to learning gains or, conversely, poses risks.

The researchers Toni Turano, Perpetual Baffour Pihl, Grant Agnew, Laurie Ziegler, and Susanna Loeb present a traffic light typology in their article "AI Tutoring Is Not a Monolith".

Green, amber, red

• Green: a human guides the pupil, with AI in the background as support. This is the form for which there is the strongest evidence of learning gains.

• Amber: the pupil interacts directly with AI, but a human supervises and remains ultimately responsible.

• Red: fully automated AI, without human supervision. Unproven and, according to the researchers, the most risky.

A striking figure from the research: 40 to 47 percent of pupils already use AI tools such as ChatGPT for homework help, often without any guidance. Precisely the red zone where the evidence is lacking.

What this means for EduGPT

This research confirms the approach with which we have built EduGPT. Not AI that replaces the teacher, but AI that strengthens the teacher. Teachers remain in control, can see where pupils are struggling, and continue to act as guides. AI remains in the background, exactly in the green zone where Stanford finds the strongest evidence.

Where pupils use a chatbot for homework without supervision, the red-zone risk described by Stanford arises. EduGPT is deliberately designed so that the school and teacher remain involved—not as a separate AI subscription alongside education, but integrated within it.

Source

• Turano, T., Pihl, P. B., Agnew, G., Ziegler, L. & Loeb, S. (2026). AI Tutoring Is Not a Monolith. Stanford SCALE.