Compare · Provenance · Research notes
Training-data checkers — scored
“Was my art used to train AI?” and “Did a student write this?” are different jobs. This page keeps them apart and gives you an evidence habit instead of a vibe ranking.
Last reviewed: 2026-08-15. Re-run the protocol after major model or pricing changes.
How do you test a training-data or detector tool?
- Name the job: dataset search for your files, or AI-text detection on a document.
- Run one known-positive and one known-negative if you have them.
- Save timestamps, URLs, and what the tool claims vs what it can prove.
- Treat a miss as incomplete information, not a clean bill of health.
What is the scoring rubric?
| Criterion | Weight | How to score |
|---|---|---|
| Job match | Gate | Fail a detector if you needed a dataset search, and the reverse. |
| Evidence quality | 1–5 | Can you export a URL, thumbnail, or score you could show someone else? |
| False confidence | Pass/Fail | Fail if the UI implies legal proof or a complete index. |
| Privacy | Pass/Fail | Do not upload private or regulated files to a public search. |
Which checker matches the job?
| Tool | Best fit | Job | What a hit means | Note |
|---|---|---|---|---|
| Have I Been Trained? | Art / image dataset search | Index match | Appears in a known set — not “model X used this” | Start here for creator provenance questions. Absence of a hit is not proof. |
| GPTZero | AI-text detection | Text classifier | A score, not authorship proof | Use as a signal in education or editing — never as a sole disciplinary trigger. |
Learn the craft, not just the tool
Practice the transferable habit in Deepfakes and synthetic media, then AI copyright basics, Have I Been Trained? overview, Voice consent checker.
FAQ
Can Have I Been Trained? prove a model used my art?
No. A match means a file or near-duplicate appears in a searchable index. It is an investigation aid, not legal proof that a named model trained on your work.
Are AI detectors accurate?
They produce scores with false positives and false negatives, especially on short or edited text. Use them as one signal beside process evidence, not as a verdict.