Reading with AI: Can Technology Foster Reading Acquisition and Promote Educational Equity?

Teachers in our treated group received a one-hour training about the use of the edtech platform Lalilo. The training emphasized how to create logins for their students, how to monitor their progress through a dashboard, and how to assign lessons to students. During this training, teachers were reminded that they volunteered to use the platform 40 minutes per week with first-grade students and were advised to divide this time between 2 x 20 minutes sessions.

Teachers received by email their logins to access the most recent version of the tool. This access needed a reliable internet connection, as the tool could not be downloaded. Teachers could then create logins for their students. Both students and teachers could then login to the platform, either at school (recommended) or at home (neither encouraged nor prohibited), on digital devices. We asked teachers to access the tool either through computers or through tablets with their students.

During their first connections, students were asked to complete an initial test, so that the algorithm behind the platform forms a prior about their academic skills. Once this first test is completed, students face exercises adapted to their skills: neither so hard that they would be discouraged to try again, nor so easy that they would not be challenged. And the algorithm keeps updating its estimate of the skills of students. During their use, students receive rewards (stars, stories, progress in a fictitious world) based on their duration of use and on their academic progress (difficulty of exercises), with the former carrying more weight than the latter to encourage perseverance and feeling of competence of all students, even the low-performing ones.

2024-11-01

2026-07-01

academic scores on the standardized French language exam at the beginning of grade 2 (short term outcome) and grades 3-6 (medium- and long-term outcomes). We will focus on the following items: writing, oral and written comprehension. Student scores on performance anxiety (Anderson index), self-efficacy (Anderson index), motivation, student engagement.

student wellbeing (Anderson index), growth mindset of students (Anderson index), scores in mathematics and reading aloud, score in comparison with peers for students (Anderson index), teacher score in differential instruction (Anderson index), in self-efficacy (Anderson index), in the use of formative assessment (Anderson index), in the use of cooperative practices, in openness to digital tools (Anderson index), growth mindset of teachers (Anderson index).

We partnered with an edtech company well-implemented in France, Lalilo. Our intervention consisted in randomizing the timing at which volunteer teachers would receive free access to the edtech platform as well as training and support for this use. Teachers could be either randomized to receive access in academic year 2024-2025 or 2025-2026.

Importantly, not all teachers in France were eligible to be part of this experiment. They needed to fulfill the following requirements
- The teacher can access at least five tablets or computers (or at least one device per four students) that can be used daily, together with an internet connection.
- The teacher teaches first‑grade students (CP) in the 2024‑2025 school year.
- The teacher does not actively use the platform in their class at the time of application and had no or low use of the platform over the past year.
- The teacher is willing to use the platform frequently and regularly and to comply with the year of treatment indicated by randomization (either the 2024‑2025 or 2025‑2026 school year, depending on the assigned group). It was recommended that each pupil uses the ed‑tech for at least 40 minutes per week, and the training advised two × 20‑minute sessions.

We assessed the eligibility of teachers in partnership with the edtech company, that verified their past use of the platform, and using data collected through a baseline survey completed by teachers before randomization.

Eligible teachers having completed the survey were then randomized in two waves (late October 2024 and November 2024). Wave of randomization depended on the date of completion of the baseline survey. The treatment assignment was clustered at the school level and stratified according to the following characteristics: past use of the edtech tool, above median openness to using digital tools, priority education status, academic region.

The randomization was performed on R software with a randomly drawn seed.

Not available

Stratified randomization done in office by a computer

School

Yes

223 schools, 258 teachers

2180 students

113 treated schools, 110 control schools