A plain-English glossary of research words

The short version

Preprint

A preprint is a research paper that the authors have put online themselves, often on a site called arXiv. Nobody has to approve it first, so mistakes may not have been caught yet. Preprints are often revised.

Example. The MIT "Your Brain on ChatGPT" study is a preprint. It was posted in June 2025 and revised in December 2025. A working paper is a similar thing in economics and education research.

Peer review

Before a journal publishes a paper, it sends it to other researchers in the field. They look for errors and can ask for changes or reject it. It is a filter, not a guarantee.

Example. The field trial with about 1,000 students was peer reviewed and published in the journal PNAS in June 2025. A correction followed in August 2025. It fixed an author's affiliation and changed no results.

Randomised trial

In a randomised trial, people are put into groups by chance. One group gets the thing being tested and another does not. Chance makes the groups alike in everything else, so a difference in results can be put down to the thing tested.

Example. In three experiments with 1,222 adults, chance decided who worked with an AI assistant and who worked alone.

The opposite is an observational study, where researchers only watch what happens. The colonoscopy study and the school records study are observational.

Correlation

A correlation means two things tend to go up and down together. It is often written as r, a number from -1 to +1. Zero means no link. It does not tell you why: A might cause B, B might cause A, or something else might cause both.

Example. A UK survey of 666 people (Gerlich, 2025, peer reviewed) found r = -0.68 between AI use and critical thinking scores. The survey cannot say whether AI use led to the lower scores. More in what surveys can and cannot tell you.

Cognitive offloading

Cognitive offloading is using something outside your head to make a thinking task easier. Writing a shopping list is offloading. So is asking a chatbot. The term was set out in a 2016 review by Risko and Gilbert in a peer-reviewed journal.

Offloading is normal and often sensible. The review also notes that people offload when they believe it is more efficient, and that those beliefs are not always right.

Effect size

"Statistically significant" only says a difference is unlikely to be chance. Effect size says how big it was. A common measure is Cohen's d, which counts the gap between two groups in standard deviations, a unit of spread. As a rough guide, 0.2 is small, 0.5 is medium and 0.8 is large.

Examples. In the persistence experiments, the gap in problems solved alone was d = -0.45 in one experiment and d = -0.20 in another. In a Harvard trial of an AI tutor with 194 students, the effect size was 0.63 in the tutor's favour.

Sample size

Sample size is the number of people or cases in a study, often written as n. Small samples give shaky results, because a few unusual people can swing the average. But a large survey still cannot show cause.

Examples. The MIT preprint had 54 people, and 18 in its last session. A set of experiments on AI summaries against web search had 10,462.

What these words cannot do

No single label settles whether a study is right. A peer-reviewed survey still shows only a link. And no study of any kind has shown that AI lowers intelligence: none has measured it before and after. Ask four questions together: was it reviewed, was it randomised, how many people, and how big was the difference?

What to do with this

Our evidence page lists every study we cite with its year, size and status.

Questions