IC-335GPT-4's detection performance as a scoring model is highly sensitive to the prompt, varying from 0.7289 to 0.9682 AUROC, far more than GPT-3.5 or Babbage

Guangsheng Bao, Yanbin Zhao, Juncai He, Yue Zhang

SourceGlimpse: Enabling White-Box Methods to Use Proprietary Models for Zero-Shot LLM-Generated Text Detection

The paper ablates five manually drafted prompts (from empty to complex) and measures the AUROC of fast-detectgpt when using different scoring models. GPT-4 shows the largest variation, with detection accuracy ranging from 0.7289 (empty prompt) to 0.9682 (prompt4), a spread of nearly 0.24. GPT-3.5 is less sensitive, ranging from 0.9071 to 0.9589, and Babbage is the least affected, ranging from 0.8921 to 0.9299. This indicates that GPT-4's predictive distributions shift substantially with the input context, making it the most prompt-dependent scoring model among those tested.

Evidence
correlational
Key metric
GPT-4 fast-detect AUROC: 0.7289 (prompt0), 0.8239 (prompt1), 0.8775 (prompt2), 0.9341 (prompt3), 0.9682 (prompt4); GPT-3.5: 0.9071 to 0.9589; Babbage: 0.8921 to 0.9299
Caveat
Only five manually drafted prompts were tested; the authors note they leave a systematic exploration of prompts for future work.
Model
GPT-4 / ChatGPT4 / GPT-4 Code Interpreter / GPT-4 Technical Report, GPT-3.5 / ChatGPT-3.5, GPT-3 / GPT base Babbage
Concepts
Scale-dependent behaviour
Datasets
XSum [eval], PubMedQA [eval]
Methods
Fast-DetectGPT [primary]
Related work
Taguchi et al. 2024 [context]
Related findings
IC-336, IC-337
Extraction
automatic-extraction