Making AI Access Fairer
The digital divide, concentrated AI development, biased training data, and the uneven distribution of AI's benefits are real problems — but they are not inevitable features of AI technology. They are the result of choices: choices about what to build, who to include, what to share, and how to regulate. That means different choices can lead to different outcomes. This lesson explores what those choices look like in practice.
Technical Approaches
Some of the most important work toward fairer AI happens at the technical level — in how data is collected, how models are trained, and how their outputs are evaluated. Diverse and representative training data is a foundational requirement. When training datasets systematically underrepresent certain groups, the resulting models perform worse for those groups. Addressing this requires deliberate effort: collecting data from underrepresented communities, funding datasets for low-resource languages, and auditing existing datasets for gaps. Fairness metrics allow developers to measure whether their models produce equitably accurate results across different demographic groups. A model that is 95% accurate overall but only 70% accurate for one group may not be acceptable, depending on the stakes of the decision. Open-source AI releases models and research tools publicly, allowing researchers, companies, and communities without large budgets to build on state-of-the-art work rather than starting from scratch. Meta's release of the LLaMA model family allowed researchers worldwide who could not afford to train large models themselves to fine-tune and adapt powerful AI for their own languages and contexts.
Linguists estimate that there are roughly 7,000 human languages. Only a tiny fraction of them have the digital text corpus needed to train quality AI language models. Initiatives like Masakhane in Africa and AI4Bharat in India are building datasets and language models for languages that large companies have no financial incentive to support, putting AI tools in the hands of communities that would otherwise be entirely left out.
Policy and Regulatory Approaches
Policy decisions shape who gets access to AI and on what terms. Broadband infrastructure investment — treating high-speed internet as public infrastructure like roads and electricity — can close the first-level access gap. In the United States, the Infrastructure Investment and Jobs Act of 2021 included over $65 billion for broadband deployment specifically in underserved communities. Similar programs exist in many countries. Public AI — government-developed AI services available to all citizens — is a model being explored in several countries. Estonia, known for its advanced digital government, provides AI-powered public services to all citizens regardless of income. India's Aadhaar system uses biometrics to provide identity services that allow citizens to access government benefits. AI in education: governments and nonprofits can fund AI tutoring tools, translation services, and educational resources specifically for under-resourced schools and communities, making the productivity benefits of AI available beyond those who can afford premium services. Requirements for language and accessibility: regulators can require that AI tools deployed in public services work adequately in the languages of all communities they serve, and meet accessibility standards for people with disabilities.
Community and Participatory Approaches
Technical and policy solutions can only go so far if the communities most affected by AI have no voice in how it is designed. Participatory design is a practice of involving the people who will be affected by a system in the process of designing it — not just consulting them after the fact, but giving them real influence over decisions. Community-based organizations have successfully advocated for limits on facial recognition in their cities, won requirements that hiring algorithms be audited for bias, and pushed for moratoriums on high-risk AI uses in criminal justice until better safeguards exist. Digital literacy education — teaching people not just to use technology but to understand how it works, evaluate AI claims critically, and advocate for themselves when affected by algorithmic decisions — is a form of empowerment that scales. A student who understands the digital divide and knows that AI systems can contain biases is better equipped to navigate, question, and influence the AI-infused world they are entering.
Learning about AI equity is not a passive exercise. Every student who understands these issues becomes a potential advocate, builder, researcher, policymaker, or voter who can influence how AI develops. The people shaping AI policy and AI tools a decade from now include students who are in middle school today.
Match each approach to making AI fairer to what it primarily addresses.
Terms
Definitions
Drag terms onto their definitions, or click a term then click a definition to match.
Why does releasing AI models as open-source help with global equity in AI access?
What is participatory design, and why is it important for AI equity?
Fairness Toolkit
- Step 1: Choose a specific inequity in AI access or outcomes from this module — for example: lack of broadband in rural areas, AI tools that work poorly in minority languages, biased hiring algorithms, or data labeling workers receiving low wages.
- Step 2: Identify three different approaches that could address this inequity — one technical, one policy-based, and one community-based.
- Step 3: For each approach, describe what it would involve, who would need to act, and what obstacle might prevent it from working.
- Step 4: Rate each approach on two dimensions: how quickly it could produce change (short-term vs. long-term), and how much power ordinary people have to make it happen.
- Step 5: Write a short explanation of which approach you would prioritize and why.