Can AI help bridge the gap in inclusive education?

In a classroom in Uruguay, students follow a lesson about the yacaré – an alligator species native to the country. Some focus on the colorful visuals on their laptop screens, while others listen to the audio narration coming through the device’s speaker. Even though the students have different learning preferences and face different barriers, they are all engaging in the same lesson together.  

The lesson happening in this classroom is remarkable for two reasons.  

First, children with and without disabilities are learning side by side. Children with disabilities are one of the most marginalized groups in society and face multiple barriers that make it hard – or even impossible – to attend school and learn. For most of the 240 million children with disabilities worldwide, an inclusive classroom like the one in Uruguay is still out of reach.  

To address this disparity, UNICEF launched the Accessible Digital Textbooks (ADT) initiative in 2014. These curriculum-aligned books are designed using Universal Design for Learning (UDL) principles: they offer content in multiple formats, such as audio narration or simplified text, and give students flexible ways to learn. Textbooks are a central pillar of learning, offering a structured and reliable foundation developed by education experts. Making them accessible is essential for inclusive education. Today, ADTs are being developed in six countries across Latin America and the Caribbean, three countries in Eastern and Southern Africa and two in South Asia. However, creating a single ADT can still take months, making it difficult to scale and meet growing demand. 

This brings us to the second reason this classroom in Uruguay is remarkable: the ADT was created with the help of Artificial Intelligence (AI).  

In 2024, UNICEF began exploring how generative AI can make the ADT development process faster, more cost-effective, and easier to scale. By using AI responsibly, countries can more efficiently convert standard curriculum materials into accessible formats – while respecting the content holders’ intellectual property and adhering to review processes led by Ministries of Education (MoE), stakeholders, and education experts. The goal is to accelerate the creation of accessible learning materials like ADTs, removing barriers to educational content for all children and making inclusive education not only possible, but sustainable.   

How does this actually work in practice? It begins by rethinking how accessible textbooks are made – and who’s involved in making them. 

Incorporating the AI tool into the production model 

To convert a printed book into an accessible digital one requires a multidisciplinary team. This team brings together experts from different fields to make sure the final product is inclusive, easy to use, pedagogically sound, technically functional and suited to the local context. It usually includes MoE officials, teachers, writers, publishers, representatives from organizations of persons with disabilities (OPDs) and technology developers. Together, they select the content to adapt into an ADT, based on the specific needs and priorities of the country.  

In the traditional workflow, without AI, the team manually creates and adapts every part of the ADT. They begin by designing each page through a detailed storyboard, deciding how it should look, sound and function to support all learners. They then produce all the accessibility content in-house – including audio narration, image descriptions, sign language videos and adapted learning activities – while also defining the user experience, layout and navigation. The process is collaborative and  rewarding but also complex and time-consuming, requiring a wide range of skills and constant coordination. 

As part of Uruguay’s effort to innovate and scale inclusive tools, AI is now part of their ADT creation process. This means the team no longer builds everything from scratch. Instead, they upload a PDF of the selected material to a secure and rights-protected AI tool, which extracts the original content and generates accessible features like audio narration, simplified text (Easy Read), image descriptions, glossaries, explanation of activities (Explain Me) and layout suggestions based on UDL principles. It creates a substantial first draft of the ADT. The team’s role has shifted from building to reviewing, refining and validating. They decide what to keep, adjust or add to ensure a high-quality ADT. They aim to strike a balance between efficiency and ensuring the content is accessible for all students.

The real game-changer is that what used to take months, can now be done in days or weeks, opening the door to scaling up accessible content production more efficiently and sustainably. The process is still led by people – but now they can spend less time on initial, repetitive tasks and more time improving the final product. As the tools improve, so do efficiency and accuracy, leading to continued growth in both the quality and quantity of accessible content.  

Adopting AI also means adopting a new mindset – one that sees inclusion not as an add-on, but as the starting point. It also means a shift for teachers and book authors: from holding full control over content creation to collaborating with technology as a creative partner. In many societies and educational contexts, AI is still met with resistance. However, the results helped overcome many of the team’s initial concerns. For many team members, it was a chance to rethink assumptions about how learning tools are created and for whom. 

Building a tailored AI tool 

Uruguay is the first country to co-create an ADT with the help of AI. Thanks to its strong education policies, expertise in inclusive pedagogy and universal implementation of digital technologies in schools, it offered the right conditions to explore how AI can support the creation of ADTs. 

A wide range of partners came together to build the tool. This included the National Administration of Public Education (ANEP), Ceibal, teachers from ANEP, AI experts and several UNICEF teams at all levels – including those working on innovation, education, disability inclusion, ICTD, research, partnerships and communication. The work kicked off in July 2024, when ANEP selected Cuaderno para leer y escribir en quinto (CLE 5), a fifth-grade Spanish language textbook, as the pilot content to be adapted with the help of AI. 

To train the AI tool and ensure quality, context-appropriate content, the team used a PDF of CLE 5, earlier ADTs developed in Uruguay (Zorrito and Mulita, created without AI), UDL guidelines and established accessibility standards. As part of the training process, the team also developed a set of “few-shot” examples: a small number of carefully selected and structured exercise templates that illustrated the expected format, tone, and pedagogical approach. This combination of inputs guided the AI tool in producing the first prototype – but that was only the beginning of the real test. 

Using expert feedback to shape and refine the AI tool 

As expected, the first accessible textbook generated by the AI tool was far from perfect. A number of key aspects needed to be adjusted and improved.  

UNICEF Innocenti – the Global Office for Research and Foresight and UNICEF Uruguay coordinated the effort to gather expert feedback. Contributions came from ANEP, including curriculum experts from ProLee, accessibility specialists from CERESO, both teams coordinated by the Office of Language Policies, and tech teams from Ceibal. Additional validations were carried out by UNICEF’s ICTD team through their Accessibility Helpdesk. At the same time, UNICEF’s Global Learning Innovation Hub, Office of Innovation and ICTD teams worked closely and continuously with AI development teams to improve the AI tool based on this expert feedback. 

The feedback process included four rounds of review, conducted in August, October, and November 2024, and again in March 2025. In each round, expert teams assessed the content, design and accessibility features, offering structured input to improve both the tool and the final ADT product. 

Several strengths became clear: 

• The AI tool could generate a wide range of accessibility features simultaneously and in a synchronized manner.
• The tool proposed multiple UDL-aligned layout options in minutes – a task that previously took days using the manual storyboard process.
• The process also helped teams develop new skills in prompting, quality assurance, accessibility and integrating AI tools. 

But the process also revealed some challenges: 

• The quality of the AI tool’s output varied significantly depending on how well the input materials reflected local context and how clearly they were structured.
• Small changes in task phrasing could produce inconsistent results, making iteration less predictable, even with well-crafted prompts and “few-shots”.
• The AI sometimes produced more solutions than needed, which could disrupt the intended pedagogical flow or skip over key steps in the learning sequence. 

Technical teams also raised an important question: how can improvement be measured across iterations? 

• They began designing an internal tool to track progress and spot patterns across versions.
• Clearer communication between teams was also flagged as a priority – acknowledging that, much like the AI tool itself, collaboration is still evolving. 

Finally, with the core components in place, the ADT generated by the AI tool was tested in real classrooms between May and July 2025. This pilot was a key step in understanding how an AI-powered ADT performs in practice – and what adjustments may be needed to support inclusive, effective learning. 

How did the AI-powered ADT perform within the dynamics of the classroom? 

Launching the pilot and learning together 

In May 2025, Uruguay launched a pilot of the AI-powered ADT, CLE 5, in 30 schools. Field visits included both inclusive and special schools to see how teachers and students – with and without disabilities – interacted with the new ADT. 

The pilot began with training sessions for 30 teachers, introducing them to the ADT’s features and supporting them in planning how to use these in their classrooms. Most teachers reported feeling comfortable with the ADT afterward, even those with limited experience in inclusive education or digital learning tools. 

In the classroom, students and teachers were learning together. While many students needed some help navigating the ADT at first, they stayed engaged and made strong progress. Teachers often used it as a traditional textbook – guiding reading, leading discussions and using activities to reinforce learning. Features like text-to-speech and simplified text were especially valued, helping all students work toward the same learning goals, together. 

Both teachers and students also had ideas for enhancements. Students imagined voice commands, mini-games and emojis in the glossary. Teachers suggested ways to better meet their students’ needs – like font toggles, different types of questions and making sure activities match the level of the simplified content. 

Alongside the pilot, Ceibal’s UX team conducted user testing in schools across Uruguay to better understand how students interact with the ADT. Overall, students showed strong engagement and enthusiasm, especially for features like audio narration and visual elements. The user testing also highlighted areas to improve, such as simplifying navigation and refining features like “Easy Read” and “Explain Me.”  

Applying this early feedback is already improving the AI tool. It is a reminder that usability, engagement and student ownership matter just as much as smart technology. 

Learning from Uruguay’s experience  

As the AI tool continues to improve, Uruguay’s experience is already offering valuable lessons – both for future work in the country and for others exploring similar paths. 

First, cross-team collaboration was essential. This process brought together a multidisciplinary team, including curriculum and accessibility experts from government partners like ANEP and Ceibal. Working closely helped ensure that the new technology truly supported Uruguay’s goals for inclusive education and can be scalable nationally and internationally. 

Second, inclusion starts with a mindset. For many team members – especially those less familiar with inclusive education – this was a chance to rethink what learning can look like. Moving away from a ‘one-size-fits-all’ approach and toward flexible, student-centered tools became a powerful idea – one that needs to be shared and reinforced from the start. 

Third, using AI requires human knowledge and guidance – not once but continuously. Countries interested in using AI for ADTs benefit from first working with the traditional, non-AI approach. It helps teams understand the underlying work – what the AI is automating; where human judgment is still essential; and how to measure quality throughout. As people interact with the AI tool, even through  a few examples or well-crafted prompts, they also can help improve the outputs and learn how to “instruct” the tool. This ongoing exchange strengthens both the tool and the skills of those using it. 

Finally, Uruguay’s experience is already making a difference beyond its borders. The AI tool developed through this pilot will now serve as a base for other countries. In this way, Uruguay’s experience is not just an example, it is a model that will help shape more cost-efficient inclusive tools and adaptive, accessible learning approaches around the world. 

The AI tool will be open-sourced and LLM agnostic before the end of 2025. In parallel, UNICEF is partnering with ministries of education, national libraries, and publishers globally, in Latin America and in South Asia to establish sustainable hosting for the pipeline, testing different business models so it can live beyond pilots. This is the beginning of a very exciting journey: we are building a community of practice around the open-source pipeline, stimulated by hackathons with developers and educators contributing new exercise types, richer content, and additional languages. Learnings from this process will ensure that we recommend a replicable, scalable model that new countries can adopt quickly and confidently. 

Uruguay’s case shows how AI can help speed up the creation of accessible digital textbooks while keeping inclusion at the heart of learning. A full report with more insights from the pilot will be published soon. 

This article was authored by Sophia Torres Cantella, Marta Carnelli (UNICEF Innocenti), and Julie de Barbeyrac (UNICEF Learning Innovation Hub).  

UNICEF does not endorse any company, brand, organization, product or service. 

Learn more about the ADT: Accessible Digital Textbooks for All | UNICEF Accessible Textbooks for All,  

ADT Research: Accessible digital textbooks | Innocenti Global Office of Research and Foresight