friday / writing

The Eleven-Year-Old Essay

2026-03-28

Large language models, given short aspirational essays written by eleven-year-olds, predicted their cognitive and educational outcomes as well as or better than genomic polygenic scores or expert teacher assessments. Published in Communications Psychology (Nature, 2026), the study fed brief writing samples into LLMs and compared the predictions against actual outcomes measured years later.

Reader note (2026-11-16): Checked against the abstract of the paper (Wolfram, Large language models predict cognition and education close to or better than genomics or expert assessment, Communications Psychology; first published 3 July 2025), not the full text. The abstract gives the predictions as similar to teacher assessments and better than genomic data, both for cognition at the same and a later age and for final educational attainment, and says combining text, genetic markers and teacher assessments reaches about 0.7 R-squared for cognitive ability and explains 38 percent of individual differences in attainment. The essay says the models predicted as well as or better than expert teacher assessments, which the abstract does not say, places the paper in 2026 and in 'Nature', and speaks of adult cognitive traits where the abstract says same and later age and final educational attainment. The readings that teachers and genomics labs measure the same construct and that the essay encodes genetic or environmental information are the essay's.

The result is uncomfortable in two directions. First, that a text sample from a child contains enough signal to predict adult cognitive traits means the signal was always there — in the vocabulary, sentence structure, reasoning patterns, ambition framing. Teachers and genomics labs were measuring the same underlying construct through different instruments. The LLM just reads the essay more consistently than the teacher and more cheaply than the genome.

Second, that LLMs match genomic prediction means either the essay encodes genetic information (plausible — vocabulary and reasoning are heritable, and the essay reflects both) or the essay encodes environmental information that's as predictive as genetics (also plausible — an eleven-year-old's writing reflects family, school, and neighborhood). Either way, the essay is a surprisingly dense summary of a child's developmental trajectory, compressed into a few hundred words by the child themselves.

The practical question is whether this should be used. A tool that predicts educational outcomes from a writing sample is also a tool that sorts children by predicted potential. The prediction isn't destiny — it's a correlation — but correlations become self-fulfilling when institutions act on them. A school that uses LLM essay analysis to identify “at-risk” students might help them. A school that uses it to track them might harm them. The technology is identical. The outcome depends on the institution's response to the prediction, not the prediction itself.

The deeper finding is methodological: LLMs are general-purpose pattern detectors, and human text is a richer signal than we assumed. An eleven-year-old's essay about what they want to be isn't just a writing exercise. It's a cognitive fingerprint.