The AI Class Gap Has a Number. It Is 76 and 38.

The AI class divide is not a metaphor. It is a number: 76 versus 38.

Those are the percentages of employed AI users who tap the technology for work purposes, depending on how they access it. Among workers with an employer-provided subscription — typically a company that has negotiated a site license or handed out credentials — 76 percent use AI as part of their job. Among free-tier users, the share is 38 percent. That gap, documented in a new Epoch AI/Ipsos survey of more than 2,000 American adults, is the clearest signal yet that the AI revolution is not arriving uniformly across the economy. It is arriving first for people who already have the most leverage in it.

The data, drawn from a probability-based KnowledgePanel with none of the self-selection bias that infects most online polling, lands at a moment when AI has become routine: half of all American adults report using an AI service in the past week. ChatGPT leads at 31 percent, followed by Google Gemini at 21 percent, Microsoft Copilot at 11 percent, and Meta AI at 8 percent. These numbers are not the story. The story is what they conceal.

Anthropic's own quarterly economic index, which tracks how Claude is actually used across the economy, shows concentration in high-income countries and in American metros where knowledge workers cluster. Early adopters — people with six months or more of tenure on the platform — use it 10 percent more for education-oriented tasks and 10 percent less for personal queries than new users. They also get meaningfully better results: a 10 percent higher conversation success rate. That is what compounding looks like. The people who arrived early are getting better at using the tool, which means they will continue to extract more value from it, which means they will continue to arrive early.

This is the mechanism that turned computer literacy into a labor-market differentiator in the 1990s. Workers who had access to PCs and training early built skills that employers later demanded of everyone. Workers who were already behind fell further back. The technology was the same in both cases. The outcome was not. Something structurally similar appears to be happening with AI, except the feedback loop may run faster and the gap between the best tool and the default tool may be wider.

Meta AI, which appears to attract a younger and less work-oriented user base, occupies a different position in this landscape. It is free. It is embedded in applications people already use — Instagram, WhatsApp, Facebook. That is precisely the point. The workers using Meta AI most heavily are not the ones whose employers have decided AI is worth buying. They are the ones using it on their own time, for personal tasks, without the institutional scaffolding — training, site licenses, workflows — that converts a chatbot into a productivity tool.

One important caveat: the specific claim that Anthropic's Claude skews higher-income while Meta AI skews lower-income, which appeared in an Epoch AI summary that has since gone offline, could not be independently verified. What can be verified is the work-use split between paid and free tiers, the knowledge-worker geography of Claude adoption, and the compounding advantage documented among early adopters. Those three facts, taken together, paint a coherent picture without relying on the missing data point.

The question worth sitting with is not whether AI is good or bad. It is whether the current adoption pattern is the kind that eventually equalizes or the kind that calcifies. Universal smartphone adoption eventually closed the mobile internet gap. There is no obvious reason AI must follow the same arc. The cost structures are different. The capability differences between a free tier and a professional subscription are meaningful. And the people who are not getting access to the professional tier are precisely the people who would benefit most from the leverage it provides.

The 76-and-38 gap will not stay frozen. Either paid AI access becomes so universal that the tier distinction dissolves, or it becomes a permanent feature of how the economy allocates technological advantage. Right now, there is no policy, no product strategy, and no regulatory framework seriously aimed at closing it.