AI writing in computer science · 2016–2025
Given AI, how much
Page one asks how many papers are AI-written. This page asks, of those papers, how much of each one is AI? Is it a flagged sentence here and there, or most of the text? And where in the paper does it sit? The cutoff below is the same slider as on How much AI, and it carries over between the pages.
How AI-written papers are distributed, by year (2016–2025)
Each bar is all full-text papers of a year, split by what share of their passages is flagged: under 10%, 10–24%, 25–49%, 50% or more. The lowest band is “within detector noise”: the detector flags about 1% of human-written passages, so a few flagged passages per paper are expected without any AI. The bars do not move with the cutoff; the cutoff outlines the bands at or above it.
Each school
Data table: share by band, school and year
Given AI, how much (2023–2025)
Papers are chosen for having at least the cutoff share flagged, so the level of this number is tied to the cutoff: raise the slider and it rises. Read it as how far above the cutoff typical AI papers are, not as an absolute amount. Only 2023–2025 is shown; earlier years have too few real AI papers for a mean to mean anything.
The distribution (choose the years; default 2023–2025)
Labs, ranked at this cutoff (one year at a time, default 2025)
A lab appears when it has at least 5 full-text papers that year. A high rank is a measurement about papers, not a judgement of the people.
Where in the paper (papers of 2023–2025, with 2016–2021 as the floor)
Each paper’s body text is cut into ten equal parts, from the opening (0) to the end (9). The chart shows what share of passages in each part is flagged, among papers at or above the cutoff. Section headings are not recoverable from the stored text, so position is by tenths of the text, grouped as opening (first three tenths), middle (next four) and end (last three). Papers are selected for having many flagged passages, so the overall level is high by construction; compare the opening, middle and end with each other, and the two periods only in shape.
How this is measured
Dose and cutoff are as on How much AI: a paper’s dose is the share of its passages the detector flags, and it is AI at cutoff t when the dose is at least t%. Everything here uses full-text papers only; abstract-only papers have no dose.
Bars. Each year’s full-text papers are counted in five bands of dose (bins 0–9, 10–24, 25–49, 50–100) from the 101 stored bins; the shares sum to 100%. Given AI. Among 2023–2025 papers with dose at least t, the mean of their doses and the median (read from the stored one-point bins, valued at the mean of its bin). Earlier years are not summarised this way: their flagged papers are false positives.
Lab list. Each lab-year with at least 5 full-text papers has its own stored distribution. The lab’s share AI at t is the papers at or above t over its papers. Small labs are shrunk toward the group’s mean share with weight n/(n+5) for ranking only; the table shows the raw share and the number of papers.
Position. A passage is flagged at its paper’s own cut; shares are over all passages of papers at or above the cutoff. Papers from 2016–2021 show the detector’s flag rate on human writing as a reference; their flagged passages are expected to be false positives.