Giving everyone free AI tools sounds fair, but do the old gaps just resurface somewhere new?

This explores whether giving people free AI tools actually narrows gaps between rich and poor, or educated and less educated, or whether the gaps just move. The corpus's short answer: access is the easy part. Even when everyone has the tool, the divides tend to reappear in three places: who actually uses it well, which tier of the product they get, and who owns the returns. A broad interdisciplinary review covering information, work, education, and healthcare finds that generative AI can push inequality in either direction. What decides it is access, integration, and incentives, not what the technology can do Does generative AI inevitably worsen or reduce inequality?.

The hopeful evidence is real. In a randomized experiment with 1,174 adults, AI shrank the performance advantage of more-educated participants by about three-quarters on a business problem-solving task. Less-educated participants also kept part of their gain after the AI was taken away Can AI narrow the education performance gap?. Among 758 BCG consultants, the bottom-half performers gained the most from GPT-4 Does AI help lower-skilled workers catch up to top performers?. One detail in that study matters a lot here: AI use was mandatory. The authors warn that when people choose for themselves whether to use it, the effect can reverse, because the people already ahead are often the ones who adopt fastest and use it most skillfully. Free access doesn't recreate the conditions of that experiment. It just makes the tool available.

The gap can also shift over time. A theoretical model by Agrawal, Gans and Goldfarb predicts a U-shape. Early AI helps weaker workers by covering for poor execution skills. As the tools improve, the payoff moves toward people with better judgment about what to ask for and what to trust, and the equalizing effect reverses Does AI's wage-equalizing effect actually reverse as it improves?. This hasn't been tested yet, but it suggests the divide may move from "can you do the task" to "can you direct the tool." Acemoglu's model points to the same kind of shift at the level of the whole economy: AI's effects on wages may spread more evenly across demographic groups than earlier automation did, but income from owning capital is still predicted to pull ahead of income from labor Will AI increase inequality less than past automation?. Brynjolfsson adds that if AI is built to replace workers rather than help them, bargaining power moves to whoever controls the systems, and that can lock in Does automating labor concentrate power more than augmenting it?. Free tools for individual users don't touch that ownership gap.

Then there's the question of what "free" actually means. OpenAI says ads on its free and $8/month tiers will pay for access for lower-income users, while the Pro, Business, and Enterprise tiers stay ad-free Can advertising revenue actually expand AI access equitably?. That widens access, but it also creates a two-tier system where some users' AI runs on advertising and others' doesn't. No data on usage or adoption was disclosed, so whether it is equitable in practice is still an open question. A separate argument says that because AI models are built from humanity's shared writing and other digital output, *restricting* access would itself create inequality by privatizing knowledge everyone helped produce Should restricting AI access create new kinds of inequality?. Seen that way, free access is the minimum, not a solution.

What you might not have expected: the strongest evidence that AI narrows gaps comes from settings where everyone is *required* to use it on the same task. Free access in the real world is almost the opposite: optional, unstructured, and different from tier to tier. So the corpus suggests that free access mostly shifts where the divides appear: to skill at using the tools, to which product tier people are on, and to who owns the systems. The corpus has little direct data on free-tier users specifically. Most of the evidence comes from controlled experiments and economic models, not from studies of how people use free AI products.