The Biggest Questions About AI
The map · 5 Humanity · 5.2 Education and human cognition · 5.2.4

Cognitive inequality

Will personalized tutors equalize opportunity, or give already advantaged people far greater leverage?

Bloom's two-sigma promise is finally testable at scale, and early RCTs show large gains — for the students who engage. Whether tutoring compresses or stretches the distribution depends on who engages.

View on the map → · Open in Browse →

What changed
December 2025–August 2026 · swept August 3, 2026 · editorial review pending
01

The equalize-or-widen question got strong data pulling both ways. Purpose-built tutors matched or beat human tutors in a UK classroom RCT and produced large gains for disadvantaged students in Nigeria — the case that scalable tutoring could compress the distribution. But real-world use runs the other way in places: lower-income applicants lean more on AI for admissions essays and end up more homogenized, a case where unequal use widens rather than narrows gaps.

Recent thinking
3 pieces · December 2025–August 2026 · all 3 chronologically →
LearnLM Team (Google) & Eedi · arXiv · 29 Dec 2025 paper
AI tutoring can safely and effectively support students: An exploratory RCT in UK classrooms
students who received support from LearnLM were 5.5 percentage points more likely to solve novel problems on subsequent topics than those who received tutoring from human tutors alone.

A successor to the Nigeria and Harvard tutoring RCTs run in real UK classrooms: a purpose-built pedagogical model matched or beat human tutors on transfer — evidence bearing on whether scalable tutoring could equalize access to one-to-one instruction.

Martin De Simone & Federico Tiberti · VoxDev · 17 Mar 2026 article
How AI tutors improved learning in Nigeria
Students offered the programme achieved large learning gains in English and also performed better on end-of-year exams.

The World Bank authors' fuller writeup of the Nigeria RCT: large gains for disadvantaged students, the strongest data point for the 'AI tutors could compress the distribution' side.

Johanna Alonso · Inside Higher Ed · 8 May 2026 news
Low-Income Students More Likely to Submit AI-Generated Admissions Essays
If it's pushing all these applicants towards the same type of essay, the same template … it could be that students are inadvertently losing that opportunity.

A Cornell/CMU study of tens of thousands of admissions essays finding AI use skews toward lower-income applicants and homogenizes their essays — a concrete case where unequal AI use could widen rather than narrow gaps.

Foundational reading (3)The 2 Sigma ProblemBenjamin Bloom · 1984AI tutoring outperforms in-class active learningKestin et al., Scientific Reports · 2025From Chalkboards to Chatbots: AI tutoring in NigeriaWorld Bank · 2025
Previous5.2.3 AssessmentNext5.2.5 Intellectual agency