If using AI at work is optional, does it mainly help people who are already ahead — by skill or gender?

This explores whether letting workers decide for themselves whether to use AI ends up helping people who were already ahead. The corpus answers the skill half more directly than the gender half. Its clearest finding is a contrast. When AI use is required, gaps shrink. In a 758-consultant BCG experiment where everyone had to use GPT-4, the bottom half of performers gained the most and moved closer to the top performers Does AI help lower-skilled workers catch up to top performers?. That same note warns that discretionary adoption could reverse the effect. The tool narrows the gap only if the people who would benefit most actually pick it up.

The corpus suggests several reasons they might not. One is that many workers don't see their own tasks as things AI could do. Making tools easier to use doesn't fix that, because the problem is noticing the opportunity, not technical friction Does easier tool-building actually solve enterprise adoption problems?. Another is a social cost. Across four experiments with more than 4,400 people, AI users expected colleagues to rate them as less competent and less hardworking, and they were less willing to tell managers they used it Do people fear judgment when they use AI at work?. When use is voluntary, people who already worry about being seen as less capable have the most reason to hold back or to hide their use. Those same workers may have the most to gain. This is where a gender effect would plausibly show up. However, none of the notes here measures adoption or penalties by gender, so the corpus can point to the mechanism but can't confirm the gap.

Self-selection also makes the evidence that adoption pays off harder to read. Anthropic found that people who delegate the most to Claude are also the most optimistic about their careers and their skills Does delegating work to AI actually damage worker skills?. That pattern fits just as well with confident, already-advantaged workers being the ones who adopt heavily. The finding is a correlation within one company's own users, which is exactly the trap discretionary adoption sets for anyone measuring its effects.

Zoom out from individuals and the same widening shows up at larger scales. Workers have actually handed tasks to AI mostly in information-heavy occupations, and the wage pattern reverses at advanced-degree levels Where have workers actually delegated tasks to AI?. Job-posting data from ten countries shows demand for AI skills clustering around a technical core (Python, SQL, machine learning). Non-technical occupations are drifting away from that core rather than toward it Is AI creating common skills across jobs or deepening divisions?. Firms split the same way: companies more exposed to AI replace contract labor with it faster and more cheaply, which points to compounding advantage rather than even spread Do firms substitute labor for AI at different rates?.

Put together, the corpus suggests the tool itself tends to equalize, but the choice to adopt it tends not to. Who adopts, who feels safe being seen using it, and who works in occupations where AI fits all lean toward people who were already advantaged. If you want the thing you didn't know to look for, it's this: the gap may come less from AI's capabilities than from the social penalty for being seen using it. That makes how employers frame and normalize AI use a fairness question, not just a productivity question.