Every trap added to this repository strengthens the research community's ability to detect AI agents in online surveys. Submissions are reviewed for completeness and added to the browsable collection. You do not need to have full deployment data; preliminary results from model testing or pilot studies are welcome.
Required
- Stimulus image (the visual task itself)
- Trap name
- Which architectural constraint it exploits
- Question text and correct answer
- Your name/affiliation (for attribution)
Optional (but valuable)
- Model testing results (which models, pass/fail rates, number of trials)
- Human validation data (pass rate, sample size, platform)
- Deployment results (agent failure rate, discrimination)
- Source paper documenting the constraint
- Response options used in the survey
Ready to submit?
Click the button below to open the submission form on GitHub. You can fill in the structured fields and drag-and-drop your stimulus image directly into the form. A free GitHub account is required.
Open Submission FormNo GitHub account? Submit via email with your stimulus image and trap details.
What makes a good submission?
Grounded in a known constraint
The best traps exploit a documented architectural limitation in vision-language models, such as data overfitting, spatial reasoning failures, cross-modal binding errors, or spatiotemporal processing limitations. Link to the computer science paper that documents the constraint if possible.
Easy for humans, hard for models
The task should have a human pass rate above 75%. If humans also struggle with the task, it cannot distinguish humans from AI and will produce too many false positives. Our Shape Overload trap (47% human pass rate) was excluded for this reason.
Tested on at least one model
Even preliminary testing (10 trials on one model) is useful. Report the model name, version, number of trials, and pass rate. Use standalone sessions (incognito/temporary chat) so each trial is independent.
Clear and unambiguous
The question should have a single correct answer that does not depend on subjective judgment. Avoid tasks where reasonable humans might disagree on the answer.