Findings from developing an AI-supported early literacy assessment for student self-assessment in Tanzania

• In Event: Developing and adapting voice AI for inclusive early grade reading assessments: insights from Sub-Saharan Africa and South Asia

Proposal

The presentation will share lessons and findings from the development and deployment of an offline AI-supported early literacy assessment approach for student self-assessment in Tanzania. The purpose of the effort was to increase the efficiency and adaptability of high-quality early grade literacy assessments to a diversity of administration approaches as well as at scale. The aim is to enable systematic adoption of early literacy assessment in low- and middle-income country contexts and facilitate data-informed, scalable student and teacher support, thus helping to improve student learning outcomes.

An organization partnered with an African tech company to develop the AI functionality to automatically score student early literacy assessments and to natively integrate it with an open-source technology platform. The platform is already used in 60+ countries to facilitate large-scale early grade reading assessments as well as teacher support for quality formative assessment and differentiated, inclusive instruction.

The development process was highly rigorous, conducted in conjunction with the elaboration of a psychometrically tested early literacy and numeracy assessment for student self-assessment in Kiswahili. The AI assessment and scoring functionality underwent systematic user testing with 40 grades 1-3 students in July 2025, field testing with 250 students in grades 2 and 4 in August 2025, and pilot testing with nearly 300 new grades 2 and 4 students in Tanzania (planned for September 2025). All data collection took place in real world, school and classroom contexts. The AI assessment functionality included tasks for letter and syllable identification, non-word identification, and oral reading fluency with scoring at letter/syllable and word level. Along the way, the team conducted concurrent validity analysis against an existing, Tangerine-supported self-assessment modality as well as against a traditionally administered, i.e., non-technology supported Early Grade Reading Assessment administration.

As a result, the AI model was trained with original audio files recorded, with parental permission, from over 500 students, with transcriptions done by local education specialists. The model achieved satisfactory character and word error rate levels. The presentation will share lessons learned as well as AI performance data.

Author

• Carmen Strigel, RTI International