AI Detection Disclaimer

The short version

• Our AI-likelihood score is a probability, not a certainty.
• False positives and false negatives are possible. No detector — ours or any competitor's — can guarantee accuracy.
• Do not use our score as the sole basis for academic discipline, hiring, firing, immigration, credit, insurance, or any other decision with significant consequences for a person.
• Treat the score as one signal among many. Combine it with human judgment, edit history, voice evidence, and a chance for the subject to respond.

1. What the score means

The Service outputs an AI-likelihood probability score (eg, 0.78), a categorical bucket ("possibly AI", "likely human"), and supporting layer-level signals (rhythm, vocabulary, rule fires, paragraph breakdown). The score is the output of a statistical model and a hand-tuned rule system. It is an estimate of likelihood under our model's assumptions, given the text we received and the language we detected.

A score of 0.85 does not mean "85% sure this text is AI." It means "under our model, this text scores in a range we generally observe in AI-generated samples." Different models would give different scores. Our score is calibrated to resonate with comparable detectors on extreme cases but is not a ground-truth measurement.

2. Sources of error

• False positives. Some human writing is unusually polished, formal, or structurally regular and looks AI-like. Non-native English speakers, technical writers, students using grammar tools, and writers who use templates can score high.
• False negatives. AI text run through a paraphraser or "humanizer" tool, AI text edited by a human, very short passages, and mixed AI+human documents can score low.
• Domain shift. Our model was tuned on broad domains. Highly specialised technical text, poetry, code, song lyrics, transcripts, and translations may score outside our envelope.
• Length sensitivity. Documents under 200 words are noisy. Documents over our tier word limit are truncated.
• Language. English performs best. Other languages are translated internally before some checks; translation introduces noise.
• Adversarial input. An attacker who knows the model can craft text that scores low even when AI-generated.
• Model drift. Generative AI models improve; our calibration is updated regularly but lags real-world output.

Independent academic studies (Liang et al. 2023, Sadasivan et al. 2023, Weber-Wulff et al. 2023, OpenAI 2023 classifier withdrawal note) have found that all current AI detectors have non-trivial error rates and that detector reliability varies across writer demographics. We agree.

3. Do not use the score as the sole basis for…

• Academic discipline, expulsion, plagiarism findings, or grade adjustments.
• Hiring, firing, or contract decisions.
• Immigration, visa, or asylum decisions.
• Credit, lending, insurance, or housing decisions.
• Criminal-justice outcomes, sentencing, or evidence presentation in court without expert testimony.
• Content moderation that imposes lasting reputational harm.
• Anything else that meaningfully changes a person's life.

Our score may inform — never decide — such matters. The decision must come from a human reviewer, applying domain expertise and the local procedural protections the subject is entitled to.

4. Recommended workflow when the stakes matter

1. Run the detection and review all the supporting signals (not just the headline score).
2. Compare against a baseline of the subject's known prior writing if available.
3. Look at edit history, draft revisions, voice notes, or process artefacts.
4. Hold a conversation with the subject — explain the result and let them respond, supply evidence, and ask questions.
5. Document the human review and the basis for any decision.
6. If the decision goes against the subject, give them a route to appeal.

5. EU AI Act & comparable laws

The Service is an "AI system" under the EU AI Act (Regulation (EU) 2024/1689). It is a limited-risk system in our intended-use scenario (informational signal for human reviewers). It becomes a high-risk system if you deploy it in certain prohibited or high-risk contexts (eg, education access, employment screening) without meeting the Annex III obligations yourself. Read our full EU AI Act Disclosure before deploying in those contexts.

6. No medical, legal, or financial advice

The Service does not provide medical, legal, financial, or other professional advice. Any output should not be relied on in place of professional consultation.

7. No guarantee, "AS IS"

To the maximum extent permitted by law, the Service is provided "AS IS" and "AS AVAILABLE" with no warranty of accuracy. See Terms §13 for the full limitation of liability.

8. Contact

If you have a result you believe is wrong and want a human re-review, email [email protected] with the detection ID. We respond within 5 business days.