Epoch AI 2026: Pangram, GPTZero and Originality.ai on 495 human passages and 594 AI passages

The record joins the studies index as the first in which two detectors returned zero false positives on a human set of this size (epoch.ai data insight, results table, read 9 November 2026).

Three detectors, all with a card in this index, each at a named version. Pangram (version 3.3.2) and GPTZero (model 2026-05-11-base) were called through their APIs; Originality.ai (Turbo 3.0.2) "offers no API at the Pro subscription tier", so it was scored through the web bulk uploader, where two passages failed and were rescored by hand (epoch.ai data insight, Data section, read 9 November 2026).

Detectors and versions as named in the insight's Data section; scoring took place in June and July 2026 per the repository README.

The human corpus is 495 passages of roughly 500 words from 99 authors, 33 each in blogging, fiction and scientific writing, five passages per author, every one published before 30 November 2022 and extracted verbatim from a Wayback Machine snapshot or arXiv source. For each author and each of three models, Claude Opus 4.8, GPT-5.5 and Gemini 3.1 Pro, the study produced one passage after showing the model five of the author's real passages and asking for a new piece in that style, and one passage from a derived one-line prompt on the same topic, giving 297 passages per condition. Titles and headings were removed and passages trimmed at a sentence boundary (epoch.ai data insight, Data section, read 9 November 2026).

Sample as described in the insight; the repository README gives the 30 November 2022 cutoff and the 8-gram verification.

The metric is each detector's document-level verdict at its default setting. Only a clean "AI" verdict counts as a detection; a "Mixed" verdict from Pangram or GPTZero counts as a non-detection on AI text and would count as a false positive on human text, though neither returned one on the 495 human passages. Error rates carry 90% confidence intervals (epoch.ai data insight, "Learn more about this graph", read 9 November 2026).

Every figure below is printed in the insight's results table, its CSV and its Analysis tables (epoch.ai data insight and detector_error_rates.csv, read 9 November 2026).

Values as printed in the insight's tables and CSV. The insight's prose rounds 3.84% to 3.8% and the CSV prints 3.84; this record carries the CSV value. Counting "Mixed" as a detection would lower the imitation false negative rate to 6.1% for Pangram and 8.8% for GPTZero (Assumptions and limitations).

The false positive result is the headline for this index: across 495 pre-ChatGPT passages, two detectors returned no false positives and Originality.ai's 19 were concentrated in fiction, 16 of 165, against 2 in blogging and 1 in scientific writing; the author notes all 19 "were flags of writing that predates ChatGPT" (epoch.ai data insight, Analysis, read 9 November 2026). On plainly prompted AI text the three detectors missed 0, 1 and 2 of 297 passages.

Style imitation is the study's evasion method, and it worked unevenly: the three tools missed 10% to 18% of imitation passages overall but 24% to 29% of imitation passages in the scientific genre, and Gemini 3.1 Pro text evaded Pangram four times as often as Claude or GPT-5.5 text (20 of 99 against 5 of 99) (epoch.ai data insight, Analysis tables, read 9 November 2026).

The author lists five: the treatment of "Mixed" verdicts changes two detectors' imitation rates (10.1% to 6.1% for Pangram, 10.8% to 8.8% for GPTZero); genre-by-model cells hold only 33 passages; the five human passages per author come from one writer, which bears on the false positive rate only; detector versions were pinned in June 2026 and "can be updated by their makers without notice"; and the study tests "a single evasion method", few-shot style imitation, with no paraphrasing tools or manual editing, at default verdicts and one length of about 500 words (epoch.ai data insight, Assumptions and limitations, read 9 November 2026). Evident from the sources: the work is a web publication by one author, not a peer-reviewed paper; the repository README carries an earlier title than the published page; and the human authors are "well-known writers", not students or non-native writers.

All three detectors have cards, and each card takes the three error rates with their bases (epoch.ai data insight, results table, read 9 November 2026).

All three appear in the ranking of detectors by test record. The human-text rates are printed counts at a default verdict and qualify for the false positive rate ranking, where Originality.ai's 3.84% of 495 sits beside the 0.47% at a 0.25 threshold that RAID 2024 reports. The direct-prompt and imitation misses are printed counts of AI passages not flagged and qualify for the false negative rate ranking as rates, never as complements. The imitation drop is an evasion result on 297 passages for the adversarial robustness ranking, with the caveat that no paraphraser or humanizer was used. The home page lists which AI detectors exist and who has tested them.

Related records: Jabarian and Imas 2025 scored the same three detectors a year earlier on 1,992 human passages and reports false positive rates of 0.0000 for Pangram, 0.0011 for Originality.ai and 0.0071 for GPTZero at a 0.5 threshold; Liang 2023 found the false positives this study did not, on non-native student essays.