🔍 Deepfake Detector

Triage an image, video, voice clip, or passage of text for manipulation signals — and get an honest suspicion band, not a verdict. All analysis runs in your browser, on your device. Your files are NEVER uploaded to CertSight or anywhere else; the server only serves this page.

$29 one-time.

Buy Deepfake Detector — $29 one-time

Deepfake Detector is a gated tool: sign in with an account that owns it to use the triage tool below. The tool section on this page is shown only to owners.

What it checks — and what it honestly cannot

  • Images — EXIF/metadata inconsistencies, JPEG error-level analysis (ELA) suspicious regions, quantization/format anomalies, C2PA manifest presence (absence proves nothing), dimension/encoding anomalies.
  • Video — container metadata the browser exposes, sampled-frame compression-artifact analysis (not every frame), audio-track spectral check where a track decodes, duration/format anomalies.
  • Audio (voice) — spectral flatness and rolloff, unnatural silence gaps, clipping, noise-floor patterns, encoder/duration/sample-rate anomalies.
  • Text — style signals only, high false-positive rate. Sentence-length variance (burstiness), lexical repetition, punctuation/casing patterns, over-regular structure. Text results are never as reliable as image/video/audio bands.

This is not proof that media is real or fake. A suspicion band — low / medium / high / inconclusive — is a triage priority for human review, not a finding of fact. No result here claims a person is lying, a crime occurred, or content is "confirmed fake" or "confirmed real." There is no neural-network detector and no accuracy claim anywhere on this page.

⚠️ Honest limits — read before you triage

  • ELA is sensitive to innocent recompression. Social platforms recompress everything, so a high ELA response on a social-media download is weak evidence. Regions flagged by ELA are suspicious regions, never proof of editing.
  • Metadata can be stripped innocently (privacy tools, messaging apps) and can also be forged. Metadata signals are corroborating only. A fully AI-generated image may carry perfectly consistent metadata.
  • Voice-clone signals are weak. A good clone through a phone codec is often indistinguishable from a real voice by these signals. Clean studio recordings of real voices also show low noise floors. Compressed audio skews toward inconclusive.
  • Text has a high false-positive rate. Careful human writing, edited copy, legal/technical prose, and non-native writing all trigger style signals. Short texts (< ~100 words) are inconclusive by default.
  • Re-encoding washes out video artifacts (every platform re-encodes). Short clips and low-resolution video give too little signal and report inconclusive rather than low.
  • When in doubt, this tool prefers inconclusive over low. A low band is never worded as "likely authentic," and "no signals found" reads as "No manipulation signals detected — this does not prove authenticity."
  • No face tracking, lip-sync, blink-rate, gaze, or speaker-identification analysis. Not in this version, not implied by any result.
  • No automated reporting, takedown, or accusation workflows. If a result worries you, it is a reason for human review — not an accusation.

Desktop-only checks (not available in this browser tool)

These need the full desktop engine. They are listed here, not faked with buttons that do nothing:

  • Full-container video forensics — encoder strings, muxing history, and per-track encoder/settings inconsistencies. The browser does not expose container internals; it exposes file type, duration, and dimensions only.
  • Sampled-frame video analysis from the file bytes — the browser tool samples a few frames at reduced size; the desktop engine samples up to 5 frames from the actual file for blockiness analysis (not every frame).
  • Saved plain-language reports — the browser tool triages one file at a time and keeps nothing after you leave the page; the desktop app can save a .txt report for the file you checked. It does not triage batches of files — one file at a time, like the browser tool.

For those checks, use FakeForge, the desktop deepfake triage app — files never leave that device either; the only network use is license activation/deactivation.