Preamble

At the co-creation workshop for the Africa Foundational Learning Assessment Initiative (AFLAI) in Nairobi earlier in September, representatives from 13 ministries of education across Africa converged to discuss data and assessments in foundational learning. This blog focuses on artificial intelligence (AI) in education assessment. It is the first output of AFLAI Knowledge Exchange Series which will address different topics identified during that workshop.

Introduction

There are different views on the use of AI in educational assessments within the African context. At the Nairobi workshop, Ministry of Education officials from 13 African countries emphasized one central message: AI can help fix pain points in assessment if African countries lead the agenda and keep children and teachers at the center.

“We cannot push AI away! We need education on the best ways to use it,” said a country representative from Nigeria, capturing the room’s mood and the shift from externally driven innovation to homegrown leadership. Across the region, ministries are exploring how AI can make assessments faster and more equitable. AI can make assessments better and more equitable by reducing testing bias, improve scoring accuracy, and accommodate multiple languages and learner’s abilities. To achieve this, ministries emphasized the importance of starting small, learning fast, and setting national rules for responsible use.

The use of AI to support assessment is particularly attractive to many ministries because current approaches are expensive, slow, and difficult to scale, leaving teachers overstretched and systems lacking timely, useful feedback on performance. Several ministries expressed interest in testing AI through targeted pilots that put classroom needs first and generate evidence to inform policy.

The potential of AI in foundational learning assessments

Ministries and partners across a growing number of countries are beginning to test how AI can strengthen existing assessment systems, from capturing oral reading fluency to analyzing handwritten responses. Some early use cases include:

Speeding up early reading checks with automated speech recognition (ASR). Governments and research teams are piloting ASR to measure oral reading fluency in local languages. Early work in South Africa, Morocco, and Tanzania shows the promise of dramatically faster and more accurate data collection, making routine reading checks far more feasible. “It’s possible to have AI assess oral reading fluency in real-time,” said a representative from Senegal. “It reduces subjectivity and bias, and cuts the time needed to score assessments.” In addition, the EGRA-AI consortium is working to build voice-recognition models for early reading in African languages to reduce the cost and time burden of one-to-one assessments.

Building the language resources Africa needs. As one representative from Côte d’Ivoire noted, “The challenge with vocal recognition is that it needs to be trained in phonetics,” highlighting the importance of language and context in developing accurate models. Researchers across several African countries are responding by expanding the range of languages represented in speech technology, from large-scale efforts like African Next Voices (9,000+ hours across 18 languages) and AfriSpeech-200 to community-driven initiatives such as Masakhane African Languages Hub, Lacuna Fund, DVoice, and NaijaVoices. While most large-scale datasets still focus on adult speech, they lay the foundation for future child-focused applications such as reading assessments.

Marking written work with computer vision and Optical Character Recognition (OCR). “Why are we not talking about AI tools to measure math?” asked a representative from Madagascar, a reminder that assessment innovation goes beyond literacy. Computer vision is now being tested to mark handwritten work, including math. At the same time, natural language processing (NLP) and large language models (LLMs) are being used to evaluate student writing for accuracy and coherence. In Rwanda, Rising Academies is experimenting with AI tools that help teachers analyze student problem-solving patterns and tailor instruction.

Turning data into decisions. “Can we have AI present data to us in ways that are usable for us?” a Kenya representative asked, an important reminder that the value of assessment lies in what educators do next. Beyond scoring, AI can synthesize results into clear dashboards and reports, so teachers can identify which students need help now and tailor instruction accordingly. It can also automate data cleaning and analysis for school and district teams, while supporting national reporting on progress toward global learning indicators such as SDG 4.1.1.

These efforts point to a future where assessments are faster, fairer, and more consistent, delivering timely insights that help teachers adjust instruction, leaders target support, and ministries monitor progress.

Guardrails for responsible use of AI

As countries across Africa begin experimenting with AI in assessment, clear guardrails are essential to build trust and protect learners.

Start with trust and quality assurance. AFLAI participants agreed that AI should be used to strengthen the integrity of assessments rather than to run or score assessments directly. “We’re not ready to use AI to assess students. We should use AI to quality-assure assessments, to catch errors by enumerators, both accidental and deliberate,” said a civil society representative from Kenya. Building confidence through transparency, human review, and quality assurance is the first step, especially before any high stakes use.

Protect children’s data and uphold sovereignty. Ministries at the workshop emphasized the deeply personal nature of children’s voices, writing, and records. However, only 36 of Africa’s 55 countries have data-protection laws, and even fewer include child-specific provisions. This highlights the need for strong safeguards as voice data collection expands. Countries such as Angola, Ghana, Kenya, and South Africa have already reviewed or are currently reviewing their policies to strengthen protections while still enabling responsible AI innovation.

Build the capacity of people in the loop. Participants highlighted the importance of human oversight and transparency in AI-assisted assessment, as well as the need to upskill those involved in implementation. As a Senegalese representative noted, “We have to accept that AI is here. It’s up to the government to enable the community to gain the necessary skills to know how to advance. The world is not waiting for us.” The boundary is clear: AI should support, not replace, teachers, enumerators, and assessment specialists.

The way forward

A shared understanding emerged at the AFLAI workshop: ministries across the continent are ready to harness AI to support assessment where it adds the most value, but the way forward must be measured and deliberate.

1. Pilot carefully, compare to human scoring, and contextualize. Participants agreed on the importance of starting small and learning fast through low-stakes pilots, benchmarking against human scoring, and adapting models to local languages and child speech. This approach can de-risk innovation while generating country-owned evidence. “Are we not going too fast?” asked an official from Benin, a useful reminder to pace adoption responsibly.
    
2. Prioritize early, low risk wins. Many participants highlighted quality assurance (e.g., flagging inconsistent enumerator entries) and progress tracking (e.g., routine reading checks or quick math exit tickets) as practical starting points for using AI to support assessments. As a representative from Zizi Afrique noted, “AI is not ready to run assessments; it needs to be contextualized.” These applications offer immediate value to teachers and leaders without introducing high-stakes stakes use too soon.
    
3. Publish national guidance and set clear rules. Ministries asked for practical advice on establishing guardrails for data collection and retention standards, child data protection, expectations for human oversight, and clear requirements for vendors on transparency, bias testing, localization, and model updates. Templates for equitable partnerships can help countries negotiate access to data, intellectual property, and capacity-building commitments.
    
4. Focus on use, not just data. The real value of AI in supporting assessment lies not only in testing technology or data collection, but in how results are used to improve decision-making. Effective pilots can help teachers adjust instruction, guide ministries on where to direct resources, and give communities clearer insight into how children are learning. Pilots can also provide contextualized information on unforeseen potential hindrances to the effective use of AI.

As evidence on the use of AI to support assessment grows, so does the imperative to deepen regional and country-level expertise, ensuring that ministries, researchers, and practitioners lead in interpreting results, setting priorities, and scaling what works. This is more than a moment of experimentation; it is a turning point. With ministries defining their own pathways and shaping partnerships on their own terms, African countries can set a global example of how responsible, equitable, and locally grounded AI can transform learning for every child.

Postscript

This blog was co-authored by Clio Dintilhac (Gates Foundation), Ahmad Jawad Asghar (Gates Foundation), Sara Cohen (Woodspring Advisory) and Jacqueline Jere Folotiya (ADEA) based on outcomes from the Africa Foundational Learning Assessment Institute (AFLAI) workshop in Nairobi on September 11^th and 12^th, 2025. It is the first in a series of publications as part of the AFLAI Knowledge Exchange Series.