Who gets AI's benefits — and who gets left out
Two students. Same city, same school district, different zip codes. One had a laptop at home, fast broadband, and parents who knew how to use AI tools for homework help. She used AI to get immediate feedback on her writing, practice math problems with explanations, and research her history project with curated summaries.
The other did his homework on a phone with a data cap, in a house where nobody had used AI tools before. The same AI tools existed for both of them. But they didn't have the same AI. Access isn't just about whether the door is open. It's about whether you can walk through it.
The digital divide has always had multiple layers, and AI adds new dimensions to each. The infrastructure gap: reliable broadband is still not universal. Rural areas, low-income urban neighborhoods, and most of the Global South lack the connectivity that AI tools require. Mobile data caps and shared devices impose real constraints on AI use that are invisible to those without them.
The device and compute gap: the most capable AI tools — local AI models, image generation, video production — require hardware that low-income users often lack. Cloud-based tools partially mitigate this, but premium features increasingly sit behind paywalls. The free tier is not the full tier.
The skills and literacy gap: even with access to tools, effective AI use requires knowing how to prompt, evaluate outputs critically, and integrate AI into workflows. These skills are unevenly distributed — strongly correlated with education level, professional context, and exposure. People who most need AI's productivity benefits often have the least support in developing AI literacy.
The Compounding Effect
Each layer of the access gap compounds the others. Without infrastructure, skills don't matter. Without device quality, infrastructure doesn't help fully. Without literacy, tools are underutilized. And the people facing all three gaps simultaneously are typically those who would benefit most from AI's productivity amplification — because they have the fewest alternative resources to draw on.
When a powerful productivity tool is unequally distributed, it doesn't just fail to reduce existing inequality — it actively widens it. Workers who can use AI effectively become more productive, more valuable, and better compensated. Workers without AI access or literacy fall further behind in relative terms. Students who use AI well produce better work and learn more efficiently. Those without access don't.
This amplification dynamic means that inaction on AI access is not a neutral choice. Allowing AI benefits to accrue primarily to already-advantaged populations while doing nothing actively increases inequality — even without any discriminatory intent.
Choose a specific population or context where the AI access gap is documented — rural communities in the US, low-income students, Global South countries, elderly populations, or workers in a specific sector. Map all three layers of the gap for that population. Then assess: what would closing the gap actually require in terms of resources, policy, and time?
Start with: "I want to analyze the AI access gap for [population/context] — here's how the three layers of the gap manifest: [your analysis]"
How AI systems reproduce and amplify existing inequalities — even without intent
The healthcare risk scoring system was deployed across a hospital network serving millions of patients. It predicted which patients needed additional care management — a resource-limited program, so targeting mattered. The system used healthcare cost as a proxy for health need.
Healthcare cost and health need are strongly correlated for most populations. But for Black patients, they diverged: due to documented historical barriers to healthcare access, Black patients with the same level of illness had historically incurred lower costs. The algorithm interpreted lower cost as lower need. The bias wasn't in the algorithm's math. It was in what the data measured — and what it didn't.
Algorithmic bias is not primarily about programmers having prejudiced intentions. It enters through several structural pathways. Biased training data: if historical data reflects discriminatory decisions, AI trained on that data learns to replicate those decisions. A hiring algorithm trained on historical hires learns that men were historically promoted more — and continues that pattern. Proxy variables: when protected attributes (race, gender) are excluded, AI systems learn to use correlated proxies (zip code, name, school) that serve as effective substitutes. Excluding the variable doesn't remove the discrimination. Measurement gaps: when training data measures outcomes (like healthcare cost) rather than underlying conditions (like health need), the gap between the measure and the reality carries any existing disparity directly into the model.
The 2019 Healthcare Algorithm Study
A 2019 Science study by Obermeyer et al. documented exactly the healthcare cost-proxy problem at scale: a widely used commercial algorithm was found to assign the same risk score to Black patients who were significantly sicker than white patients. The algorithm wasn't designed to discriminate — it used cost as a proxy for need, and historical cost data carried the disparity. The researchers estimated the bias affected millions of patients.
Anti-discrimination law distinguishes between disparate treatment (intentional discrimination based on protected characteristics) and disparate impact (neutral-seeming policies that disproportionately harm protected groups). Most algorithmic bias is disparate impact: no intent to discriminate, but outcomes that fall disproportionately on specific groups.
This distinction matters for governance. Disparate treatment is straightforwardly illegal in most jurisdictions. Disparate impact is legally more complex — and AI systems that produce disparate impact can often be deployed legally even when they cause significant harm to protected groups. The absence of discriminatory intent does not make discriminatory outcomes acceptable; it just makes them harder to remedy through existing legal frameworks.
Choose a documented case of algorithmic bias — the COMPAS recidivism algorithm, Amazon's hiring AI, facial recognition accuracy gaps, or the healthcare cost-proxy case. Identify the specific mechanism through which bias entered (biased training data, proxy variable, measurement gap). Then assess whether the bias was disparate treatment or disparate impact, and what remedy would address the root cause rather than just the symptom.
Start with: "I want to analyze bias in [system/case] — the specific mechanism I think caused it is [your analysis]"
How AI development concentrates power — and what that means for the rest of the world
The largest AI models in 2024 were trained by a handful of companies — all headquartered in the United States, with significant research presence in the United Kingdom and China. The compute required to train them ran on data centers consuming gigawatts of power. The data used to train them was drawn primarily from English-language internet content.
A researcher in Lagos building a medical AI for Nigerian patients found that the foundation models available to her had been trained on almost no data from sub-Saharan Africa. They performed poorly on local languages, local disease patterns, and local clinical presentations. The tools were available. But they hadn't been built for her, her patients, or her problems.
AI development is geographically and economically concentrated in ways that have structural consequences for the rest of the world. The compute, capital, talent, and data required to train frontier AI models are accessible to a tiny number of organizations in a small number of countries. This concentration creates several dynamics that disadvantage most of the world.
Training data representation: AI models are trained on data that reflects the linguistic, cultural, and economic contexts of their creators. English dominates; high-income-country contexts dominate; Global North problems and solutions dominate. Models trained on this data perform worse in other contexts and embed assumptions that may not apply globally.
Value extraction: AI systems deployed globally may extract economic value from users worldwide while concentrating that value at the headquarters of the developing companies. Data generated by users in low-income countries contributes to AI improvement; the benefits primarily accrue elsewhere.
The Language Gap
There are approximately 7,000 languages spoken in the world. Most large language models are fluent in fewer than 100. The majority of the world's linguistic communities have minimal representation in the training data for the most powerful AI systems. For users of underrepresented languages, AI tools are less capable, less accurate, and less useful — a form of the access gap that persists even when connectivity and devices are available.
As AI becomes critical infrastructure — powering healthcare systems, educational tools, government services, and economic activity — countries without domestic AI capacity become dependent on AI developed elsewhere. This dependency has sovereignty implications: the rules embedded in AI systems (what content is allowed, what languages are supported, what values are encoded) are set by the developing organizations, not the countries using them.
Several Global South countries and blocs are attempting to develop domestic AI capacity specifically to avoid this dependency. The challenge is that the compute, capital, and talent requirements for frontier AI development are enormous — making genuine independence difficult for most countries. The middle path being explored: developing locally-adapted versions of open-source models, building domain-specific AI for local needs, and participating in international governance processes that might shape global AI norms.
Choose a specific country or region in the Global South and analyze its relationship to global AI development. Map: What AI tools are available and how well do they work in local languages and contexts? What local AI development capacity exists? What dependencies on foreign AI have emerged? What strategies is the country pursuing to address the gap?
Start with: "I want to analyze AI inequality for [country/region] — here's what I know about how AI development affects them: [your analysis]"
What equitable AI requires — and what governance approaches come closest
The city had deployed an AI system to predict which children were at risk of lead poisoning — a genuine public health problem with serious consequences. The model was accurate in aggregate. But an audit found it was less accurate in the most affected neighborhoods — the ones where historical data was patchiest, where children had moved most frequently, where records were least complete.
The children most at risk of lead poisoning were also the children the AI was least capable of identifying. The intervention had been designed to help the most vulnerable. Its limitations fell hardest on them.
Equitable AI is not just AI that produces equal outcomes across groups — it is AI that serves all affected groups effectively, that does not systematically harm those with least power, and that distributes its benefits in ways that address rather than amplify existing disparities. This is a higher bar than "non-discriminatory" and a different bar than "accurate in aggregate."
Four requirements for equitable AI in practice: Disaggregated performance evaluation — measuring accuracy and error rates by demographic subgroup, not just overall. Systems that work well on average can fail systematically for specific populations. Affected community inclusion — involving the people most affected by AI decisions in system design and evaluation. Those with lived experience of the problem the AI addresses often identify failure modes that developers miss. Benefit distribution assessment — asking not just whether the AI works, but who benefits from it working, who bears the costs of its errors, and whether that distribution is equitable. Ongoing monitoring — equity is not a one-time certification; systems drift, contexts change, and ongoing monitoring is required to catch emerging disparities.
The Equity-Accuracy Tradeoff Myth
A common objection to equity requirements in AI is that they reduce accuracy — that requiring equal performance across groups forces a worse overall system. Research does not consistently support this tradeoff. Improving performance for underserved groups often requires better data and better models, which can improve overall performance. The framing of equity versus accuracy typically reflects whose accuracy is being counted as "overall."
Several governance approaches specifically address AI equity. Algorithmic impact assessments: requiring pre-deployment evaluation of AI systems for disparate impact on protected groups — analogous to environmental impact assessments. Adopted in some US cities and jurisdictions. Disaggregated reporting requirements: mandating that AI developers and deployers report performance metrics by demographic group, making disparities visible. Procurement standards: governments requiring equity criteria in AI procurement — particularly important given government's role as a large deployer of AI in high-stakes domains. Community benefit requirements: conditioning AI deployment approvals on demonstrated benefit to affected communities, not just to deploying organizations.
Choose a public sector or social impact AI application — predictive policing, benefits eligibility, school resource allocation, medical diagnosis, or child welfare risk scoring. Design an equitable AI framework for that application covering: disaggregated performance evaluation (what metrics, for which groups), affected community inclusion (how, at what stage), benefit distribution assessment, and ongoing monitoring. Then identify the most significant barrier to implementing your framework.
Start with: "I want to design an equitable AI framework for [application] — here are the four components of my framework: [your design]"
15 questions. Complete all to finish the module.