AI Will Widen the Gap Between Rich and Poor

The Great AI Divide: Why the Richest Will Get Richer

In 2023, a radiologist in Boston used a generative AI tool to read 40% more scans in a single shift. Her hospital paid her a bonus. That same year, a medical transcriptionist in Phoenix lost her job to the same kind of software. She had no bonus. She had no warning.

This is not a story about technology. This is a story about who owns the levers.

A new paper in PNAS Nexus by Valerio Capraro, Austin Lentsch, Daron Acemoglu, and Selin Akgun (2024) lays out the evidence: generative AI will not simply create winners and losers. It will amplify the distance between them. The rich will adopt first, train the models on their data, and capture the productivity gains. The poor will be left with the scraps of automation, misinformation, and a digital divide that widens by the day.

The authors are not alarmists. They are economists and behavioral scientists who have spent years studying inequality. What they found is colder than hype.

The Productivity Paradox: Who Actually Gets the Gains?

The standard story is that AI will make everyone more productive. That is true in the narrow sense. A lawyer using GPT-4 can draft contracts faster. A graphic designer using Midjourney can iterate concepts in minutes. But the paper shows that the benefits are not distributed evenly. They are concentrated.

Capraro et al. (2024) reviewed dozens of studies on workplace automation. The pattern is stark. High skill workers, those already in the top income brackets, see their output rise by 20% to 40% when using generative AI tools. Low skill workers, those in data entry, customer service, and transcription, see their jobs disappear or their wages stagnate.

The mechanism is simple. AI tools are expensive to implement. They require training, infrastructure, and organizational buy in. Wealthy firms and wealthy workers can afford the upfront cost. They then capture the efficiency gains. Poorer workers and smaller firms cannot. They fall further behind.

The authors found that this effect is not hypothetical. In a meta analysis of workplace studies, they observed that generative AI increased inequality within occupations, not just between them. Two people doing the same job, one with access to AI tools, one without, will see their earnings diverge. The gap is not temporary. It compounds.

The Education Trap

Education was supposed to be the great equalizer. Generative AI could personalize learning for every student, adapting lessons in real time. That is the promise.

The reality is different. Capraro et al. (2024) examined pilot programs in schools. Wealthy districts already have AI tutors, personalized dashboards, and teacher training. Poor districts have none of that. They lack the bandwidth, the hardware, and the expertise.

The result is a digital divide that grows with each new tool. Students in affluent schools learn faster because the AI adapts to their pace. Students in underfunded schools get the same generic worksheets they had before. The gap in test scores widens. The gap in college admissions widens. The gap in lifetime earnings widens.

The authors are careful to note that this is not inevitable. Policy could close the gap. But current funding models do not prioritize equitable access. They prioritize early adoption by those who can pay.

Healthcare: Better Care for the Rich, Worse Care for the Poor

Healthcare is where the inequality becomes literal. Generative AI can analyze medical images, predict disease progression, and recommend treatments. It can save lives.

But Capraro et al. (2024) found that the benefits flow to those who already have good insurance. Hospitals in wealthy areas adopt AI diagnostics first. They reduce errors. They catch cancers earlier. Patients in rural clinics and public hospitals get the same old human error rates.

The paper cites a study of dermatology AI tools. The algorithms were trained on images of light skinned patients. They performed poorly on darker skin. The result is a diagnostic gap that mirrors the racial wealth gap. The people who need better care the most get worse care because the AI was not built for them.

This is not malice. It is market logic. Companies train on the data that is easiest to get, which comes from wealthy, mostly white populations. The poor are invisible to the training set. They become invisible to the care.

The Misinformation Machine: Who Gets Fooled and Who Profits

Generative AI can create content at scale. That includes lies.

Capraro et al. (2024) reviewed the evidence on misinformation. They found that generative AI dramatically lowers the cost of producing false narratives. A single bad actor can now generate thousands of convincing fake articles, videos, and audio clips. The scale is unprecedented.

But the damage is not evenly distributed. The authors showed that people with lower digital literacy, often those with less education and less income, are more susceptible to AI generated misinformation. They are targeted by scams, political propaganda, and health fraud. The wealthy, who have better access to fact checking tools and critical thinking resources, are less affected.

The result is a two tier information ecosystem. The rich navigate a world of curated, verified content. The poor swim in a sea of AI generated garbage. The gap in trust, in knowledge, and in decision making widens.

The Policy Failure

The paper concludes with a sobering assessment of policy. Capraro et al. (2024) examined the regulatory frameworks in the European Union, the United States, and the United Kingdom. They found that each fails to confront the inequality problem directly.

The EU has the strongest data protection laws, but they focus on privacy, not on distribution of benefits. The US has a market driven approach that favors innovation over equity. The UK is somewhere in between, but none of them address the core issue: who gets to own the AI and who gets left behind.

The authors propose concrete policies. Tax incentives for companies that train AI on diverse data. Public investment in AI infrastructure for underfunded schools and clinics. Mandatory impact assessments that measure inequality effects before deployment. But they admit that political will is scarce.

What the Research Does Not Prove

This paper is rigorous, but it is not prophecy. The authors are clear about what they cannot know.

They cannot predict which jobs will be fully automated versus augmented. The history of technology is full of predictions that did not come true. Typewriters were supposed to disappear. They did, but only after a century of coexistence with computers.

They cannot measure the long term effects of AI on inequality because the technology is too new. The studies they reviewed cover months or a few years. The compound effects over decades are unknown.

They cannot account for human adaptation. People may find ways to use AI that the authors did not anticipate. A low skill worker might use a free AI tool to learn new skills and move up the ladder. That is possible. But the evidence so far suggests that the barriers to entry are high.

The paper is a warning, not a verdict. It is an invitation to act before the gap becomes a chasm.

The Hidden Mechanism: Data Ownership

There is a layer beneath the paper's analysis that deserves attention. The authors touch on it but do not fully develop it. It is the question of who owns the training data.

Generative AI models are built on vast datasets scraped from the internet. Most of that data was created by ordinary people writing blog posts, commenting on forums, uploading photos. The companies that train the models capture the value. The creators get nothing.

This is a wealth transfer from the many to the few. Every time a writer, a photographer, or a teacher contributes to the training set, they enrich a corporation without compensation. The rich get richer because they own the models. The poor get poorer because they gave away their work for free.

Capraro et al. (2024) note that this dynamic is not well studied. It is a gap in the research. But it is the logical conclusion of their analysis. If the benefits of AI flow to those who own the capital, and the capital is concentrated, then inequality is not a bug. It is a feature.

The Education Paradox Revisited

There is a cruel irony in the education findings. Generative AI could be the ultimate equalizer. A student in a remote village with a smartphone and a good internet connection could access the same personalized tutoring as a student in a private school in Manhattan. The technology exists.

But the paper shows that access is not enough. The student in the village needs a device, reliable electricity, and a language model that speaks their dialect. They need a teacher who knows how to integrate the AI into the curriculum. They need a society that values their learning.

None of that comes free. The digital divide is not just about hardware. It is about the entire ecosystem of support that wealthy students take for granted. The AI widens the gap because it amplifies existing advantages. It does not create new ones from scratch.

The authors recommend targeted investment in underserved communities. That is correct. But it requires political will that is currently absent.

What This Actually Means

• ▸If you are a policymaker, do not wait for the market to fix inequality. The evidence from Capraro et al. (2024) is clear: market forces concentrate AI benefits among the already wealthy. Regulation and public investment are the only tools that can redirect the gains. Tax AI profits. Subsidize access for low income schools and clinics.

• ▸If you are a worker, assume your job will be augmented or automated within five years. The paper shows that low skill tasks are most vulnerable. Start learning how to use AI tools now, even if your employer does not provide them. Free versions exist. The gap between those who use AI and those who do not will only grow.

• ▸If you are a company, invest in diverse training data. The healthcare example is a warning. Models trained on homogeneous populations fail on others. That is not just unfair. It is bad business. The market for AI products that work for everyone is huge and underserved. Build for the bottom of the pyramid, not just the top.

• ▸If you are a citizen, demand transparency in AI deployment. The paper shows that the effects on inequality are measurable. They should be measured before, not after, a tool is released. Support laws that require impact assessments. Ask your representatives why they are not already in place.

• ▸If you are a journalist, stop writing about AI as if it is a neutral force. It is not. It is a tool that amplifies the power of those who wield it. The question is not whether AI will change the world. It is whose world it will change and who will be left behind.

The gap is not inevitable. But it is the default. Closing it will require more than good intentions. It will require a fight.