Bring your 100 toughest texts. Try to defeat our detector.
No other detector on the market will take this bet: pick the texts, we score every one, we sign every result, and if we get more than 5 wrong, we refund the license and publish the failure on this site.
- Corpus size
- Up to 100 texts per challenge round. Each text 100–10,000 words. Larger corpora available under negotiated site license.
- Submission format
- Plain UTF-8 text OR a JSONL file with
{"id":"<uuid>","text":"...","label":"human"|"ai"}per line. - Scoring
- Every submitted text is returned with an Ed25519 signature over the
(text_sha256, score, detector_version, timestamp)tuple. Signatures are verifiable in-browser at trigeochiral.com/verify or via thetruth-in-ai-verifypip package. - Miss definition
- A miss is a text where our returned probability is on the wrong side of 0.5 relative to your provided ground-truth label. Confidence-weighted alternate metrics available on request.
- Refund threshold
- ≥ 6 misses out of 100 → full license fee refunded prorated, failure case published on this /challenge/ page within 30 days with the anonymized failure texts and our commentary.
- Right to publish
- By submitting, you grant TriGeoChiral non-exclusive right to publish anonymized aggregate results. Your organization identity is only published with explicit written consent.
- Bounty (open bounty variant)
- For public bounty submissions (individual researchers, not institutions): $500 per verified miss up to $50,000 per organization per calendar year. First 100 verified misses across all submitters paid out.
- Turnaround
- Scored results returned within 72 hours of receipt. Larger corpora may take longer — quoted on submission.
- What we will not do
- Cherry-pick. Retry. Re-score. Update our detector mid-challenge. The version SHA on your signed results is pinned at submission time and never changes.
Download Public Verification Bundle (Bitcoin Anchored)
Every benchmark score is Ed25519-signed and anchored to Bitcoin Block #840219. Download the public proof bundle and run the verifier locally on your machine:
Every registered challenge starts with an unbounded API key for a 100-request trial. You can validate the scoring format and signature verification before you sink time in preparing your 100-text corpus.
Why would you offer this if you have anything to lose?
Because every commercial AI detector already refuses it. Our detector is not calibrated to any leaderboard, is not silently updated, and returns signed receipts for every score. The floor of our claim is the same as our public benchmark (Mean AUC 0.9302 on RAID-extra, 30 signed configurations). If you can push us below that with 100 buyer-picked texts, you paid for a detector we should never have shipped — and we owe you both the refund and the failure case.
Is a challenge required to buy a license?
No. Challenges are optional. Institutions buying at the standard site-license tier get the trial API automatically. The challenge is for buyers who need internal proof for procurement, or press/researchers who want to test us publicly.
Can we submit texts our own students wrote?
Yes, with two constraints: (1) student texts must be de-identified per FERPA before submission; (2) submitting means you have institutional authority to share them with a third-party detection service under whatever consent regime applies at your institution. If you're not sure about (2), we recommend the on-prem deployment (see Case 08) — text never leaves your infrastructure and the challenge runs entirely inside your VPC.
What's the smallest institution that can run a challenge?
The bounty variant is open to any individual researcher, journalist, or hobbyist. The institutional variant assumes there's a paid license on the table; typically a department or larger. If you're not sure which side you're on, email us.
What if we want to test with adversarial paraphrasing tools?
Fair game and honestly expected. The RAID benchmark's adversarial paraphrase attack is where every commercial detector collapses to 0.42–0.76 AUC. Ours is at 0.9958 on German paraphrase, published and signed. See Case 02 for the head-to-head. Bring paraphrasers — that's what we're calibrated against.
What if we want to test with humanizer tools (Undetectable AI, Quillbot, etc.)?
Also fair. Also expected. Humanizers are, in the RAID taxonomy, adversarial-paraphrase attacks with additional lexical substitution. Our signed benchmark covers those attack classes explicitly. Bring them.
What can't we submit?
Personally identifying information beyond what's standard authorship metadata. Copyrighted texts you don't have distribution rights for (fair-use excerpts are fine). Illegal content. Texts under an active legal proceeding where our involvement would create discovery risk — talk to us first for those.
How do you prevent me from just submitting texts I know break the detector?
We don't. That's the whole point. If you find 6 texts that break us in 100, we owe you the refund. Cherry-picking against us is the intended attack surface — the whole offer collapses if buyers can't choose adversarially. If our detector holds up across whatever you throw, that's the sales result. If it doesn't, the failure is published and we improve.
Are the failures actually published?
Yes. The Published failures section below this FAQ lists every institutional-challenge failure to date. Anonymized organization, verbatim texts, our detector version at the time, our commentary on what went wrong, and the fix status. Zero failures published so far — but we're publishing this page so the counter can move.
Truth-in-AI Challenge failures published to date — anonymized organization, verbatim failure texts, detector version SHA at the time, our commentary, and fix status.