How this page evolved

Every published version is kept. Each entry shows what changed and why. Earlier versions can be restored by moderators, but only if the material they rely on is still available.

1. Revision 4Oct 4, 2026, 8:37 AMLiveAI-organised, reviewed

   Added three newer sources (2024–2025) that broaden the picture beyond the two 2023 evaluations: a 2024 test of 10 free detectors showing sensitivity from 0% to 100%, a 2023 comparison of 16 detectors finding most fail on GPT-4 text, and a 2025 literature review of 34 articles concluding detectors remain unreliable despite often exceeding 50% accuracy. Added a robustness survey documenting perturbation, out-of-distribution and hybrid-text weaknesses.

   • · The main finding was rewritten.
   • · Updated “What the tests found”.
   • · Updated “Why detection is fragile”.
   • · Updated “What this means for education”.
   • · 4 new sources cited.
   • · The note on uncertainty changed.
   • · The key takeaways were revised.
   • · 3 new pieces of guidance for specific situations.
   • · 1 new open question.
2. Revision 3Sep 30, 2026, 6:49 PMAI-organised, reviewed

   Initial Starting Map on whether AI-text detectors can reliably distinguish human from AI writing, built from two peer-reviewed evaluations covering 2023.

   • · First published version.