AIFakeScan.Start a scan
OUR METHODOLOGY

A closer look.
A considered conclusion.

AI detection should give you a reason to trust the finding. Our approach brings source context and visual evidence into a report that makes its scope clear.

Evidence with context Clearly defined coverage Visible uncertainty
AI IMAGE & VIDEO INSPECTIONPublic methodology · Beta
THE BASIS OF A FINDING

Different evidence.
Different weight.

A verified source statement and a visual clue do not carry the same meaning. Each finding needs to be understood in relation to what it can actually establish.

SOURCE EVIDENCE

What the source declares

Valid, trusted Content Credentials can document declared AI generation or editing. The finding concerns the declaration and the media it covers; it does not independently verify every claim made about the scene.

Why trust matters

A credential needs more than a recognizable name or an embedded label. Verification and trust determine whether a source declaration can support a strong finding. An absent credential is neutral; an invalid one is a provenance issue, not proof of AI.

SUPPORTING CONTEXT

What the file preserves

Embedded information and visible AI labels can provide useful context about a file’s origin or editing history. They are assessed with their limitations, because sharing and editing can change what survives.

Why one clue is not enough

Camera fields can be edited. An export label can describe software without explaining how it was used. Visible marks can be added or removed. None of these alone establishes the authenticity of a file, and missing metadata does not imply AI generation.

VISUAL EVIDENCE

What the media suggests

Visual analysis looks for signals associated with generated media. These findings depend on the file, the content and the checks available. A convincing appearance is not an authenticity test.

Why validation comes first

Detector signals need evaluation against genuine and generated media before they can support reliable probability claims. Passive-only results remain Inconclusive in the current beta. An internal signal should not be read as a measured chance that a file is fake.

These are evidence categories, not a sequence of processing steps. Availability is specific to each scan.

PRECISION ABOUT THE QUESTION

A finding should have
a clear boundary.

“Was AI involved?” can mean several things. The report separates these questions so a finding about one part of the media does not speak for all of it.

GENERATION

The visual as a whole

Evidence associated with generated imagery. A whole-image signal does not reliably locate an edited region or establish what happened outside the frame.

MODIFICATION

A change within the media

AI editing and face replacement are different from whole-image generation. A source declaration may disclose editing; specialized face-swap classification is not included in Basic Scan.

AUDIO

The origin of the voice

The soundtrack can have a different origin from the picture. Basic audio inspection does not classify cloned or synthetic speech, and that coverage is reported separately.

Understand the report labels
OUR STANDARD FOR CLAIMS

Confidence needs
evidence of its own.

A detector’s output is only useful when its reliability is understood. We distinguish the ability to run a check from evidence that its conclusions are accurate.

CURRENT STATUS · BETA

Validation in progress

Our evaluation and calibration are not yet complete. Passive-only findings remain Inconclusive; no validated accuracy percentage or measured AI probability is claimed for this beta.

A trusted declaration of AI use can support a stronger source finding. It does not turn unvalidated visual signals into a calibrated probability.

What meaningful evaluation must address

  • False alarmsHow often genuine media is incorrectly flagged.
  • Missed AI contentWhere generated or modified media escapes detection.
  • Changed filesHow sharing, compression and editing affect the findings.
  • Unfamiliar contentHow well results extend beyond the examples used to develop a detector.

These are evaluation requirements, not performance claims. Published results should identify their test conditions and limitations.

WHERE THE ANSWER STOPS

Useful evidence.
Not unlimited certainty.

The limits are part of the method. Knowing what a result cannot establish is essential to using it well.

A shared copy can lose context
Crops, screenshots and re-encoding can change or remove useful information. A finding describes the submitted file, not an unseen original.
A sampled video is not every frame
Short edits and unsampled moments may be missed. A reported timestamp helps direct a review; it does not identify an exact edit boundary.
An untested feature remains unknown
Unavailable watermark verification, manipulation checks or AI speech classification cannot be counted as a clean result. Deep Scan is not publicly available yet.
Authentic media can carry a false story
A detector cannot establish someone’s identity, intent or the truth of a caption. Source research and contextual verification remain necessary.
QUESTIONS ABOUT THE METHOD

The details that matter
to your decision.

Our public methodology describes the standards behind the report. Internal implementation details remain private.

Do you publish the full detection process?

This page explains the evidence you can expect, how to interpret its strength and where the result stops. We keep implementation details private. That does not change the report’s responsibility to identify missing coverage, explain its finding and make uncertainty visible.

Can AI Fake Scan identify the exact generator?

Not reliably for every file. A source declaration may name a tool, but that is different from independently identifying a generator from the media alone. We do not claim universal attribution across generators or model versions.

Do you detect every invisible AI watermark?

No. Some markers require access to a vendor’s verification service. A watermark that could not be checked is outside the result’s coverage. No marker found is not proof that a file came from a camera or was made without AI.

Why might two copies of the same media produce different findings?

A screenshot, crop or re-encoded copy is a different file. Its metadata, source credentials and visual signals may differ from the original. Detector availability and future method updates can also affect coverage. Compare the evidence and limitations, not only the headline.

Does a signed source credential prove the event is real?

No. A verified declaration can support a statement about the file’s source or declared editing. It does not establish that a caption is accurate, an event was unstaged or a person’s claims are true. Media provenance and factual verification answer different questions.

Where can I see your measured accuracy?

We have not published a validated accuracy benchmark for this beta. Evaluation and calibration are still pending. Any future benchmark needs to explain the tested media, conditions, errors and coverage so its results can be interpreted meaningfully. We do not replace that evidence with an unverified accuracy percentage.

PUT THE EVIDENCE IN CONTEXT

Know what you’re looking at.

Choose a file, inspect the finding, and see its limits.

Check an image Check a video