Tellcheck for GitHub 作者: Munzzyy
Flags likely AI-generated pull requests and issues on GitHub. A signal for maintainers, not a verdict.
擴充套件後設資料
畫面擷圖
關於此擴充套件
Your issue tracker is filling with AI slop. Codeberg, Godot, OpenJDK, Rust, and Ghostty all restricted AI-generated contributions this year. Whichever way a project lands, enforcing it means a human still has to read every submission.
Tellcheck for GitHub does the first read. On any GitHub pull request or issue, click scan. The description and comments get badges saying whether the text statistically reads as AI-generated. Each badge opens a breakdown of the exact signals that fired: buzzword density, stock phrases, uniform sentence rhythm, chat-UI artifacts, and more. On a repo's pull request list, scan up to 25 open PRs in one click. Anything past a scan's limit is labeled "not scored", never silently skipped.
Honest about what it is. It surfaces tells for a human to weigh and never returns a verdict. Nobody can prove AI authorship after the fact, this tool included. Every result is a signal at a stated false-positive rate. On a held-out set of 4,300 real non-native-English essays that rate is 7.8%, not zero. That is the honest ceiling on trusting any single flag. It is why the advice never changes: judge the contribution, not the author. The full measurement page with the held-out numbers and the named failure modes is linked from the site.
Private by design. Nothing leaves your browser until you click scan (auto-scan exists, ships off). Scanned text is scored over HTTPS and discarded, never stored. No accounts, no tracking. Optional GitHub token for batch scans stays on your device.
Tellcheck for GitHub does the first read. On any GitHub pull request or issue, click scan. The description and comments get badges saying whether the text statistically reads as AI-generated. Each badge opens a breakdown of the exact signals that fired: buzzword density, stock phrases, uniform sentence rhythm, chat-UI artifacts, and more. On a repo's pull request list, scan up to 25 open PRs in one click. Anything past a scan's limit is labeled "not scored", never silently skipped.
Honest about what it is. It surfaces tells for a human to weigh and never returns a verdict. Nobody can prove AI authorship after the fact, this tool included. Every result is a signal at a stated false-positive rate. On a held-out set of 4,300 real non-native-English essays that rate is 7.8%, not zero. That is the honest ceiling on trusting any single flag. It is why the advice never changes: judge the contribution, not the author. The full measurement page with the held-out numbers and the named failure modes is linked from the site.
Private by design. Nothing leaves your browser until you click scan (auto-scan exists, ships off). Scanned text is scored over HTTPS and discarded, never stored. No accounts, no tracking. Optional GitHub token for batch scans stays on your device.
由 1 位評論者給出 0 分
權限與資料
更多資訊