AI writing in computer science · 2016–2025
How much AI
A detector scores each passage of a paper’s text for signs of language-model writing. A paper counts as AI-written when the share of its passages that are flagged reaches a cutoff you choose. Drag the slider: every number and chart on this page recomputes from the stored distributions. Because any detector flags some human text, the page also shows its false-positive floor, the same measure on papers from 2016–2021, before these tools were in use.
At this cutoff
All schools, by year
Data table
Each school
All panels share one vertical scale. The shaded band is the 95% interval; the dashed line is that school’s false-positive floor. Intervals are bootstrap intervals over labs at the stored cutoffs (interpolated between them), or Wilson intervals from counts when abstract-only papers are included.
Data table: share by school and year
How this is measured
Dose. Each full-text paper is split into passages. A passage is flagged when the detector’s score passes its threshold, set so that it flags 1% of passages in text written before 2022 (the 1% false-positive rate). A paper’s dose is the share of its passages flagged. Papers with no flagged passage have dose 0.
Cutoff. A paper is AI-written at cutoff t when its dose is at least t%. The share on this page is the number of such papers divided by the papers counted. The slider moves t from 1% to 100%; at a low cutoff the measure is sensitive and the floor is high, at a high cutoff only mostly-AI papers count.
Abstract-only papers. Some papers have no usable full text. With the toggle on, they are counted by their abstract flag (a separate detector rule, also set at a 1% false-positive rate), which does not change with the cutoff. They are added to both the count and the denominator. Georgia Tech currently has almost no full-text papers, so its panel appears only with the toggle on.
False-positive floor. The share at the same cutoff over papers from 2016–2021, using only papers the detector never saw in training. It is what the measure reports when no AI writing is present; the excess above it is the signal.
Intervals. 95% bootstrap over labs (a paper’s first lab as the cluster, 300 draws) at the stored cutoffs 1, 5, 10, 25, 50, 75 and 100%; in between they are interpolated linearly. Schools or cells with fewer than 20 full-text papers have no band. With abstract-only papers included the band is a Wilson interval from the combined counts.
Years. 2016–2025 only; the year after is incomplete and is not shown anywhere. Next: among AI papers, how much of each paper is AI →