AI in Education, The Great Equalizing Promise
Summary
This is part two in a three part series on AI in education. The first part summarized the state of AI in education and privatization. This second part focuses on AI’s role in promoting equity in education (or the lack thereof). The final and third part will discuss privacy concerns, surveillance, power, and the lack of regulation in the AI space.
**While the term “artificial intelligence” is somewhat of a catch-all that describes a vast range of different technologies and applications, this article primarily focuses on Large Language Models (LLMs) such as ChatGPT, Gemini, Grok, and Deepseek.
The Great Equalizer Promise
In education, there are generally three types of AI powered tools:
Learner-centered AI tools designed to enhance the students’ learning experience
Teacher-led AI tools which help teachers in their instructional and administrative roles
Administrative AI tools which help improve operational efficiency and management.
A key aspect of learner-centered AI tools is the ability to adapt or personalize the learning of individual students. LLMs are said to be able to adjust content, pace, and difficulty in real-time depending on the students’ individual learner profiles. Example uses for equity include differentiating learning resources for different learning levels of students, or contextualizing the lessons based on student interests. A teacher may use an LLM to adjust a pre-made physics lesson plan so that it is framed using a sports game and simplified for a beginner learner. Lessons can also be translated to a student’s first language, or converted into an audio format with detailed descriptions of images or graphs for students with visual impairments or dyslexia.
How Does AI Know These Things Anyway?
Essentially, a LLM is a huge neural network trained on massive amounts of text. Its job is to predict the next token (word or part of a word) based on the previous ones. Each word the LLM spits out is the token with the highest probability of being correct.
Existing inputted data is heavily skewed towards higher income countries. 90% of data sets analyzed by researchers came from Europe and North America, while less than 4% came from Africa. English also remains the dominant language in training data, even though more than 80% of the world does not speak it.
Publicly, companies such as Open AI, Meta, Microsoft, and Google are not transparent about what data they use to train their models. One possible source of data comes from the Common Crawl foundation which “crawls” the entire internet and downloads it for anyone to freely use. The information amounts to about five trillion tokens and is likely to still contain copyrighted content. To access “closed” data, companies must reach agreements with repositories, such as media outlets or non-AI tech companies like Reddit or DeviantArt.
In 2025, it was revealed that Meta and OpenAI both trained their AI models from a massive trove of pirated books and articles. Emails from copyright lawsuits show that Meta employees knew they were using stolen intellectual property, but top Meta executives told them to press on. The year before, a whistleblower who worked at Open AI revealed that the output from ChatGPT corresponded to an estimated 73-94% of the information in the training dataset. This calls into question the ethics of using a tool trained on illegally obtained data and which produces content that includes copyrighted information.
AI replicates and exacerbates human prejudices
The future is already here. Even prior to AI, many math-powered applications and models encoded human prejudices, misunderstandings, and bias into the software systems which have increasingly become responsible for managing our lives. Algorithms were already used to determine credit offers, healthcare access, and court sentencing. Time and time again, these algorithms have been proven to discriminate actively against marginalized groups.
AI exacerbates these existing prejudices. In 2024, a study by UNESCO and the International Research Centre on Artificial Intelligence proved that LLMs reproduce and reinforce stereotypes. LLMs tested generated content that reinforced traditional gender roles and outputted negative content when referring to LGBT+ individuals.
In one experiment, LLMs were tasked with scoring and providing feedback on a piece of writing. The LLM was given different information about the supposed writer, such as race, class, or the type of school. The LLMs ended up producing biased evaluations depending on who the writer was, suggesting that if teachers use LLMs to assist in the grading process, there is a risk of discrimination, especially if student names or other identifying information are present in the work.
But bias in schools doesn’t only come from machines
There is a concern that both teachers and AI tools may be biased. While there is a possibility to address teacher bias through training, AI companies would have to take the initiative to reduce biases in existing datasets. This would require diverse and representative data collection and continual monitoring for bias throughout AI’s lifecycles; a cost and burden that these companies are unlikely to take on.
Does AI actually individualize learning for each student?
Yes, but not always. Studies have shown that AI provides more relevant results for some groups of students than others based on demographic characteristics. In other words, AI is more likely to effectively individualize lessons for privileged students from higher income countries, while struggling to generate relatable content that can reach others. This is because disadvantaged learners, who are supposedly the target groups of AI-based equity initiatives, are often from groups with underrepresented data sets. In Solomon Islands, culturally relevant resources were created by local people, digitalized, and published online for public access. This type of specific content, such as “Mautikitiki and the Giant Crab,” must still be made in partnership with local people to ensure accuracy and relevancy.
How do LLMs respond to language diversity?
Language models have been shown to perform better with languages that use Latin scripts than those that use non-Latin scripts. Furthermore, non-Latin languages often do not have preexisting datasets with the quality, scale, and domain-specific focus necessary for educational applications. The poor quality of these data sets result in a higher likelihood of AI hallucinations or misleading information, further compromising learning outcomes. Developers trying to build in these languages face greater limitations, higher operational costs, and increased processing times. LLMs are also poorly equipped to handle dialectal differences, regional variations, and languages that use non-linear syntax.
Even prior to the proliferation of LLMs, researchers discovered that a shocking amount of the web is machine translated. Not only was online content in low-resource languages more likely to have been (poorly) translated from another source, the material itself was more likely to be geared toward maximizing clicks. Training AI on this already incorrect and flawed data will result in even worse outputs. The lack of quality data sources for the world's thousands of languages has created an equity gap that is not easily solved.
Do LLMs exacerbate or reduce access issues?
The integration of AI in education may further the already large “digital divide”. The ability to use AI tools depends on a school having digital resources, internet access, and trained staff members. Popular AI enrichment tools such as game-based learning, Virtual Reality, and Augmented Reality, can widen the gap even more between resource-rich and resource-poor schools and families. Inadequate last-mile infrastructure remains a barrier to any type of educational technology. AI educational tools are only the most recent attempt in a long line of historical claims that technology can be the driving force for equity in education.
Some organizations, however, have adapted AI technologies to function in contexts with low connectivity and intermittent electricity. In India, several apps use AI to deliver content in regional languages across text, speech, and video, using WhatsApp. The tools also often offer offline capabilities. One of them, Doubtnut, is specifically built for low-end smartphones with minimal bandwidth, to reach learners in remote areas.
Old failures made anew
As highlighted by the failures of the One Laptop Per Child (OLPC) program, new technologies must be implemented alongside teacher training. However, the AI “revolution” is developing much faster than teachers are able to be trained on it. While in the US, half of teachers have received training on AI in 2025 (up from 39% in 2023); in sub-Saharan Africa only 24% have received training in just digital technology. This lack of training in the region has led to an underutilisation of education technology and a resistance by teachers to change. AI isn’t even on the horizon of possible integration.
Without proper infrastructure and strong education system support in place, these technology fads lose steam as the funding dries up and the teachers give up, often leaving schools with expensive equipment that no longer function and can’t be fixed.
The gap will only continue to increase. In 2025, Microsoft, Anthropic, and OpenAI committed to investing a total of 23 million dollars for AI training to the American Federation of Teachers. Similar partnerships have also been set up with the National Education Association. Not all countries are able to access this sort of training, especially if these large tech companies do not see a profitable future market, as they do with teachers from the United States. This brings up the issue of the “second digital divide” which focuses on disparities in digital skills and usage patterns. Though AI is a relatively easy tool to use “out-of-the-box,” new users without AI literacy will likely fall victim to its tendency to spout misinformation and hallucinations.
AI-ding in Education
There is potential for AI to be used to aid the education process. AI programmes (not chatbots or LLMs) which are fed data free of human biases do exist and have potential to help countless students.
In 2018, researchers developed AI tools to ease the diagnostic assessment of dysgraphia, a neurological condition that affects a person’s ability to write. The tool used handwriting data from children to achieve a 96% accuracy in dysgraphia detection. A similar tool has been able to detect dyslexia based on audio recordings.
In India, an AI-powered assessment tool called Smart Paper uses image processing to read paper worksheets with multiple choice answers. Rather than manual grading for a week, teachers can scan papers into the system and receive an automated aggregation of results. This is especially helpful for large-scale government assessments. At the administrative level, Smart Paper has been able to digitize millions of paper records that previously required weeks of manual data entry.
In order to mitigate the risk of students becoming overly reliant on LLMs, programmes such as Khamigo guide students through questions, helping them to arrive at the correct solution on their own. Unfortunately, this system is a commercial product. As of February 2026, the product is only available in the United States for students and their parents. A partnership with Microsoft has temporarily opened up a free pilot version to all teachers in 180+ countries.
We can’t stop it, so what can we do?
The integration of generative AI into education has already begun, bringing with it a set of new opportunities and challenges. If implemented with care, AI does have the ability to help bridge certain educational gaps within the classroom. However, on its current path, its implementation will likely further widen the gap between social stratas and higher/lower income countries.
Unfortunately, where we are right now in the journey of AI development is still far too soon for equitable implementation in global school systems. True equitable use of generative AI would necessitate large investments in technology infrastructure and training programmes for educators at the national level. Curriculums must also be re-worked with the consideration of AI tools. Educators must ensure that students understand how to critically engage with and use LLMs, while also working to ensure that students retain critical thinking, creativity, and problem-solving skills. Additionally, AI development teams must reduce biases and ensure AI technologies are safe for children to use. This is just the start of a large laundry-list of concerns and considerations that policymakers, administrators, educators, and AI developers must contend with. AI must be implemented into our education systems as a tool for equity, not controlled by private, for-profit companies, under the cultural narratives of the global north.
Author: Dorothy W.
A previous fellow with PEHRC, Dorothy currently works as a consultant for UNESCO doing education policy research while continuing to teach on the side.
Views expressed in this blog are those of the author/s’ alone. Publication on this blog does not represent an endorsement by PEHRC of the opinions expressed.