How our AI detection works.

Text

The text analyzer measures sentence and paragraph structure, repetition, vocabulary distribution, transitions, hedging, templated phrasing, citation-shaped patterns, stylistic shifts, and document consistency. A frozen paired-pattern rule is available only for sufficiently long essays, articles, and reviews. Other genres and short samples abstain.

Research update · October 13, 2026

What our paid writing pilot taught us

Twenty participants completed a small Prolific feasibility pilot: one assignment asked for original writing, the other for a revision of a supplied AI draft. All twenty received $5, including the participant who withdrew their writing. We evaluated nineteen retained samples with the existing local engine; the withdrawn text was excluded.

• Original-writing assignment: 3 of 9 showed the paired pattern, 4 did not, and 2 received no paired-pattern assessment.
• AI-draft-revision assignment: 4 of 10 showed the paired pattern and 6 did not.

These are assignment groups, not verified human and AI labels. Assistance disclosures were mixed, some task instructions were misunderstood, and participant clarifications could not all be matched to individual samples. Both groups used the same prompt. These counts do not measure accuracy or error rates, and cannot tell us who wrote a particular sample.

Changes informed by the pilot

The analyzer now shows its own word count and explains the minimum input requirements before a local review. The pilot revealed that a study word count could pass while the analyzer counted fewer than 180 words. Another sample lacked enough sentence structure. Descriptive review remains available when the paired-pattern comparison cannot run.

We also clarified the revision task’s assistance question and made missing submission requirements visible. The pilot is closed. Participant writing remains private under the original withdrawal and retention terms; it was not published, used for model training, or sent to an external detector in this evaluation. The engine thresholds remain unchanged.

Try the text analyzer

Detector evaluation · October 27, 2026

What the detector did with our writing samples

Fifteen retained, paid participants had separately consented to GPTZero processing: six from the original-writing assignment and nine from the AI-draft-revision assignment. Four other retained samples were excluded because external processing was declined. No participant identifiers were sent with the writing.

• Original-writing assignment: 3 AI-like, 1 mixed, 2 human-like.
• AI-draft-revision assignment: 4 AI-like, 1 mixed, 4 human-like.

Using the provider’s current response fields, a fixed 95% confidence requirement, and the preceding available model version did not change any of the fifteen classifications. All scans completed. These are self-reported assignment groups using one prompt, not independently verified authorship labels or a held-out accuracy benchmark. Human editing can substantially change an AI draft.

The tested changes did not demonstrate improved accuracy. We do not turn these counts into a public accuracy percentage. Our detector now preserves uncertainty as an inconclusive result and records the provider model version; neither change is claimed to solve the pilot’s separation problem. Original participant writing stays private and was not used for model training.

For a new detector, we require a separate evaluation of original human writing, unedited AI output, and edited AI output. Unavailable scans and inconclusive results remain in the evaluation record. A new provider needs its own consent before receiving participant writing.

URL

The website tool retrieves one public page and lets you review its extracted writing before separately consenting to GPTZero screening. Local structural declarations remain supporting context. It does not inspect source-code authorship or crawl an entire site.

Image, audio, and video

The local observation tools organize compression, pixel, metadata, waveform, frame, and temporal indicators. These local measurements have not been validated as origin detectors. The separate Sightengine image scan provides a provider model score. The experimental video workflow sends three sampled still frames to the image provider only after consent. Those frame scores do not measure the entire video, temporal manipulation or deepfake authenticity. Audio has no AI classifier.

Volunteer Video research

The planned study keeps preliminary checks in the browser, then uses private quarantine and complete no-network offline validation only after explicit adult, rights, people, research, and product-development consent. Related versions are source-grouped. Review states, declared production evidence, measurements, contradictions, and provenance remain separate; none is an origin verdict.

Evidence model

Reports keep observations, supporting evidence, contradicting evidence, alternative explanations, limitations, and provenance checks separate. Detector-style ideas such as regularity or burstiness are limited indicators - not verdicts or probabilities.

Research boundary

Research counts describe fixed, bounded studies and counterexamples. They are not advertised as population accuracy, sensitivity, specificity, or proof. Frozen evaluation results are not used to silently retune the evaluated rule.