The Generative AI Generation in Higher Education · Evidence dashboard, Nov 2022 to Nov 2026

Sheet 08 of 17 · Outcomes

Learning outcomes

Effects depend heavily on design. Scaffolded tutoring helps; unrestricted answer access can raise assisted scores while lowering unaided ones.

Table 8.1: Four outcomes, measured separately

Better work with AI is not the same as learning. Each study below is sorted by what it actually measured. Only column D shows whether students can do the work on their own.

A Assisted performance Quality of work produced with AI Details +

B Speed Time taken to complete the task Details +

C Perceived learning How much students feel they understood or engaged Details +

D Independent competence Performance without AI, ideally after a delay Details +

StudyA · AssistedB · SpeedC · PerceivedD · Independent
Bastani et al. 2025 · RCT+48% general GPT; +127% tutor——−17% after general GPT; tutor largely avoided the loss
Kestin et al. 2025 · RCT—Less time on taskHigher engagement and motivationMore than 2× median gains on post-test
IZA 2025 · experiment———+0.23 SD exam score (check whether AI was available)
Middlebury · experiment———+0.27 SD knowledge test; more durable for augmentation
Chen & Cheung 2025 · meta-analysisMixed; not separated—g = 0.62 affective-motivationalNot separated from assisted results

— = not measured or not reported. Classification is this dashboard's reading of each study; check the originals.

Fig. 8.2: Effect sizes by outcome (Hedges' g)

0 0.5 1.0 1.5 2.0 2.5

Open marker = not statistically significant. 57 studies, 97 effect estimates. Source: Chen & Cheung, 2025 meta-analysis.

Fig. 8.3: Assisted gains vs. unaided performance

Change relative to control, ~1,000 high-school maths students in Turkey. Assisted practice improved; once AI was removed, the general-chatbot group did worse. Tutoring safeguards largely prevented the loss.

Source: Bastani et al., PNAS 2025.

RCT · Harvard physics · n=194 ↗

>2×

median learning gains from a pedagogically scaffolded AI tutor vs. active-learning class, in less time.

Details +

IZA · Dec 2025 · n=334 ↗

+0.23 SD

test performance from AI tutoring; strongest for low prior knowledge and strong self-regulation.

Details +

Middlebury experiment ↗

+0.27 SD

immediate knowledge scores; more durable for augmentation than automation users.

Details +

Reading note: Meta-analytic effects come mostly from structured interventions, not everyday chatbot use. Separate four outcomes: a better answer with AI, faster completion, feeling that one understands, and retained competence without assistance.