Does AI help lower-skilled workers catch up to top performers?

In the same 758-consultant field experiment, Dell'Acqua, McFowland, Mollick and colleagues split AI-condition subjects into top-half and bottom-half performers by their baseline assessment score and found both groups gained from GPT-4 on the within-frontier tasks, but unevenly: "the most significant beneficiaries of using AI are the bottom-half-skill subjects," while "the top-half-skill performers also received a significant boost, although not as much." The authors read this as generative AI serving "as a significant equalizer in the realm of high-end knowledge work, narrowing the performance gap between different skill levels," while noting the above-average group "also benefited meaningfully."

The paper places this result against a mixed literature it says its own data helps resolve. Brynjolfsson et al. (2025), in customer support, found gains "primarily among lower-skilled workers, with negligible effects observed among their higher-skilled counterparts." Choi and Schwarcz (2024), using law school exams, found bottom-of-distribution students "benefited enormously" while higher-skilled students "saw performance declines." Otis et al. (2024b), studying Kenyan entrepreneurs, found the opposite pattern: high performers gained from AI-generated advice while low performers "tend to fare worse." Against that spread, the BCG result is framed as a "universal benefit" case — in elite, high-end knowledge work, both halves gain and the bottom half gains most. The authors immediately qualify this with a caveat about adoption: their setting is one where AI use was strongly encouraged rather than discretionary. Citing Otis et al. (2024a), Humlum and Vestergaard (2025), and Blandin et al. (2026), they note that when adoption is discretionary, "early adopters are disproportionately male and more technologically confident, and the resulting performance gains can widen pre-existing gender- and skill-gaps rather than close them," with the risk of widening "existing promotion differentials."

This directly parallels Can AI narrow the education performance gap?, another setting where AI narrowed a skill-based gap rather than widening it. It also parallels, with a different shape, Does AI assistance help less experienced workers most?, where less-experienced workers gained most and the most experienced actually lost ground rather than merely gaining less. The explicit counter-case the paper itself cites is Does AI assistance help or hurt less experienced entrepreneurs? (Otis et al. 2024b), where the skill gradient runs the opposite way. See the companion note Does AI help or harm consultant performance on unfamiliar tasks? for this experiment's within-frontier-versus-outside-frontier result.

What the experiment establishes is the effect under mandated use in one firm's elite, assigned-task setting; it does not establish that the same equalizing pattern holds when consultants choose for themselves whether to use AI, and the authors themselves warn that discretionary adoption can push the other way. The practical implication the paper draws is conditional: AI narrowed this skill gap where its use was assigned and encouraged across the board, but left to individual choice, the demographic skew in who adopts first could turn the same tool into a gap-widener.