Table 6 Mapping of textual results to classification labels
Tool | Result | 01-Hum, 02-MT | 03-AI, 04-AI, 05-ManEd, 06-Para |
|---|---|---|---|
Check for AI | “very low risk” | TN | FN |
“low risk” | PTN | PFN | |
“medium risk” | UNC | UNC | |
“high risk” | PFP | PTP | |
“very high risk” | FP | TP | |
GPT Zero | “likely to be written entirely by human” | TN | FN |
“may include parts written by AI” | PFP | PTP | |
“likely to be written entirely by AI” | FP | TP | |
OpenAI Text Classifier | “The classifier considers the text to be …” | ||
“… likely AI-generated.” | FP | TP | |
“… possibly AI-generated.” | PFP | PTP | |
“Unclear if it is AI-generated” | UNC | UNC | |
“… unlikely AI-generated.” | PTN | PFN | |
“… very unlikely AI-generated.” | TN | FN | |
DetectGPT | “very unlikely to be from GPT-2” | TN | FN |
“unlikely to be from GPT-2” | PTN | PFN | |
“likely to be from GPT-2” | PFP | PTP | |
“very likely from GPT-2” | FP | TP |