AI Paper Atlas

Research map / Large language models / Language, safety and interpretability

DetectGPT: Zero-Shot Machine-Generated Text Detection using Probability Curvature

Eric Mitchell, Yoonho Lee, Alexander Khazatsky, Christopher D Manning, Chelsea Finn

ICML 2023 · Large language models › Language, safety and interpretability · approach: Zero-shot machine-text detection

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
replace log-prob thresholding with perturbation discrepancy (DetectGPT, T5-3B perturbations)zero-shot detection of model samples · XSum, SQuAD, WritingPrompts (Table 1)AUROC: +0.10 avg XSum, +0.05 SQuAD, +0.01 WritingPrompts over strongest baselineImproved
use perturbations from mask-filling model (T5) rather than no perturbation (log p baseline)GPT-3 and Jurassic-2 detection · PubMedQA, XSum, WritingPrompts (Table 2)average AUROC: 0.69 to 0.83Improved
nucleus (top-p=0.96) or top-k=40 sampling for generation (ablation)all zero-shot detectors · XSum, SQuAD, WritingPrompts (Table 3 and appendix)AUROC: detection easier on average for all methodsImproved
increase number of perturbations from few to 100 (ablation)DetectGPT · GPT-2 and GPT-J on three datasetsAUROCImproved
use a small mask-filling model (random or smaller T5) instead of larger T5 (ablation)DetectGPT · SQuAD contextsAUROCHurt
score with a surrogate model different from the source model (ablation)DetectGPT · XSum, SQuAD, WritingPrompts (GPT-J, Neo-2.7, GPT-2)AUROCHurt
replace a fraction of model text with T5 edits (human-edit simulation) (ablation)DetectGPT · XSum paraphrase experimentAUROCHurt
normalize perturbation discrepancy by std of perturbed log-probs (ablation)DetectGPT · main experimentsAUROC: around 0.020 increaseImproved
apply all T5 mask perturbations at once on very long passages (ablation)DetectGPT · passage length analysisAUROCHurt
use supervised RoBERTa detectors instead of DetectGPT on out-of-distribution domains (PubMed, German)supervised detectors · PubMedQA and WMT16 GermanAUROCHurt

Later papers that compared with this method

7 later accepted papers report DetectGPT as a baseline in their results tables; 7 of them report a better number on most of the results they share with it.

Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy