Grounded in research. Connected to classroom practice.

What the research found

In a randomized field experiment involving nearly 1,000 high school mathematics students, Bastani and colleagues (2025) found that access to a general-purpose GPT-4 interface improved performance during supported practice. However, when students subsequently worked without the tool, that group scored 17% lower than students who had practiced without AI.

An AI tutor designed with learning safeguards largely mitigated this negative effect, although it did not produce a positive effect on subsequent unassisted performance. These findings come from a specific mathematics setting and should not be generalized to every use of AI.

How we apply it

This distinction informs our principle:

Completion ≠ Development.

We help educators examine both what students accomplish with support and what they can understand, explain, and do independently. Our AI-use decisions begin with the learning purpose and the thinking students need to practice.

What the research found

Chi and colleagues (1994) studied eighth-grade students learning about the human circulatory system. Students prompted to explain the material to themselves showed greater gains in understanding than students who read the same text twice without those prompts. This was a small study involving 24 students.

How we apply it

We build opportunities for students to explain how they reached a conclusion, why they chose an approach, and what changed their understanding.

These explanations serve two purposes in our designs: they give students an opportunity to make meaning, and they give teachers additional evidence to examine.

What the research found

Dignath and Büttner (2008) analyzed 49 studies involving primary students and 35 involving secondary students. Their findings indicated that interventions can foster self-regulated learning at both levels, with outcomes varying according to the strategies taught, subject area, and implementation.

How we apply it

We embed planning, monitoring, and reflection into learning experiences.

Students set goals, select strategies, check progress, and consider what to change. In Human Impact Lab, these opportunities sit alongside routines that support organization and executive functioning.

What the research found

Wisniewski, Zierer, and Hattie (2020) synthesized 435 studies involving more than 61,000 learners. Feedback had a positive average effect on learning, but effects varied substantially. The information conveyed by the feedback was an important factor in its effectiveness.

How we apply it

We design feedback opportunities around a next decision: What needs attention? What could the student try? How will they know whether the change helped?

Students also practice evaluating feedback and explaining how they used it, including feedback generated by AI.

What the research found

Roseth, Johnson, and Johnson (2008) reviewed 148 studies involving early adolescents. Cooperative goal structures were associated with higher achievement and more positive peer relationships than competitive or individualistic goal structures.

How we apply it

The Activate Collaboration element of our HUMAN Framework focuses on shared thinking.

We design opportunities for students to consider different perspectives, explain ideas, negotiate decisions, and take responsibility for a shared outcome.

What the research found

Freeman and colleagues (2014) analyzed 225 studies comparing active learning with traditional lecturing in undergraduate STEM courses. Active learning improved examination performance and reduced failure rates on average. The study's scope was undergraduate STEM education.

How we apply it

This research contributes to our emphasis on students actively working with ideas.

Our designs ask students to interpret, question, discuss, test, and revise. We look for the intellectual participation within an activity and the understanding it is intended to develop.