Abstract of Study 5

65 JC1 students participated in 3 Chemistry lessons over 5 weeks, taught by the same teacher. In the first and third lessons, students completed similar graded titration-planning exercises without scaffolding, providing baseline and post-intervention measures of performance. During the intervention, students were assigned to one of three groups:

Control group (C) received static scaffolds in the form of hints.
Experimental group 1 (E1) used an AI chatbot prompted to provide contingent, step-by-step coaching without giving direct answers anchored to the same hints.
Experimental group 2 (E2) received the same chatbot support as E1, but the scaffolding was faded: students were informed that they would lose access to the chatbot during the latter half of the lesson.

E2 achieved the highest adjusted mean post-test performance, followed by E1 and then C. This pattern suggests that gradually removing AI support may encourage students to internalise the learning process and perform more independently.

However, the differences among the groups were not statistically significant (p = .254), possibly because of the short intervention period. Interestingly, students' perceptions showed the reverse pattern.

They rated the static scaffolds (C) as most helpful, followed by the non-faded contingent scaffolds (E1) and then the faded contingent scaffolds (E2).

The condition perceived as least helpful produced the highest adjusted mean performance. This finding aligns with research suggesting that students may prefer less effortful learning approaches—even when these approaches are not the most effective for learning.

Write-up: Effect of Faded and Non-Faded Contingent Scaffolds Provided by an AI Tutoring Bot on Student Learning (875 KB)

Potential Implications

(Practice) Teachers should implement faded scaffolding deliberately when using suitably prompted AI chatbots, to further improve learning outcomes. For Singapore MOE teachers, you can limit the total number of interactions when students use the Learning Assistant in SLS, or use Lesson Pacing to switch off access to it after a certain time to encourage students to build independent mastery.

(EdTech products) In light of the unreliability of student perceptions regarding usefulness of different learning approaches as opposed to learning performance, Edtech developers should enhance product decision making by using accurate measures of learning outcomes instead of depending solely on engagement measures.

Cycle 1: January to June 2026