Frequently asked questions
Straight answers about AI detection
Detection is useful, but it is not a polygraph. Here is what these checks can and can't tell you — including the parts most tools would rather not say out loud.
- How accurate is AI detection?
- No detector is reliable enough to be the sole basis for a decision — ours included. Independent testing of major tools regularly finds false-positive rates well into the double digits on genuine human writing. We treat every signal as evidence, not proof, and show you the underlying measurements so you can judge for yourself instead of trusting a single number.
- What does the score actually mean?
- A score such as “72% AI-likelihood” is an estimate of the probability that the content was AI-generated — not a claim that 72% of the text is AI and 28% is human. It is one weighted summary of several forensic signals, each of which we display alongside the verdict.
- Can the result be wrong?
- Yes, in both directions. Polished, templated, technical, or non-native human writing can read as “uniform” and trip a false positive, while AI text that has been lightly edited or paraphrased can slip below the threshold. Never use a single automated result to accuse someone.
- Is my content uploaded or stored?
- No. The text, image, audio, and video checks run entirely in your browser — your content is decoded and analyzed locally on your device, and is not uploaded to a server or stored by us.
- Which AI models can you detect?
- We look for general statistical and forensic artifacts rather than a fingerprint of one specific model, so there is no per-model guarantee. Newer generators and human editing both reduce the available signal, which is why we report confidence honestly and flag low-confidence inputs.
- Does paraphrasing or a “humanizer” defeat detection?
- Often, yes. Tools that rewrite AI text specifically raise its burstiness and unpredictability — the very properties statistical detectors rely on — so they can push a sample below the detection threshold. This is a fundamental limitation of any statistical approach, not a quirk specific to us.
- How much text or media do I need for a reliable check?
- More is better. A couple of sentences or a few seconds of audio simply don’t carry enough signal to score with confidence, and we say so rather than guessing. Longer, unedited samples produce more trustworthy results.
- Is a verdict proof that someone used AI?
- No. A verdict is the start of a closer look, not the end of one. False accusations based on detector output have caused real harm — treat the result as one input alongside context, earlier drafts, and conversation.
- How is this different from a watermark or Content Credentials?
- Watermarks and C2PA Content Credentials are forms of provenance: when present and valid they are strong, cryptographic evidence, and we read them directly. Our forensic signals are heuristics for the common case where no such provenance exists. Provenance, when available, always outranks heuristics.
- Do you offer an API or batch processing?
- A public API and batch processing are planned but not yet available. Today the detectors run interactively in your browser.