1. Provenance
TSDIQ reads the provenance record already held for the content — who supplied it and under what class (business provided, customer submitted, GroundTruth, public source, and so on). Analysis never changes that record.
How TSDIQ Content Authenticity works
A multi-signal evidence system for image, text and video. TSDIQ presents the available evidence and its limits; you decide what it means.
TSDIQ does not claim that content is definitely AI-generated, and does not claim that content is definitely human-made. It reports which technical signals were readable, what they indicate, and what remains unknown.
This assessment describes available technical evidence about the content. It is not definitive proof of authorship.
Text analysis can identify signals associated with AI assistance, but those signals are not definitive proof of authorship.
2. Metadata
Container and embedded metadata are inspected where readable: creation and modification markers, software fields, export information. Location data, device identifiers, personal identifiers and file paths are never published.
3. Content credentials
Where a content-credential structure is embedded, TSDIQ records it as present, absent, invalid, unreadable or unsupported. The engine is provider-agnostic and treats no single credential issuer as universal truth.
4. Watermark signals
Detectors for known AI watermark and provenance schemes are pluggable. TSDIQ only reports a watermark result for schemes it can actually read; otherwise the signal is marked not supported rather than absent.
5. Content analysis
Structural and statistical analysis of the content itself: container structure for images and video, sentence-length variation, phrasing patterns and lexical variety for text. These indicate; they do not prove.
6. Transformation
Resizing, compression, editing, re-export, transcoding, screenshots, translation and rewriting are recorded as context. A transformed file is not an AI-generated file.
7. Source information
The declared source is preserved and clearly labelled as declared, not confirmed. Where the source cannot be established, TSDIQ reports the source as unknown.
8. Multi-signal assessment
Signals are weighed by class, never averaged: an embedded AI declaration outranks a statistical hint; conflicting evidence is routed to human review rather than resolved by arithmetic. The published methodology is TSDIQ Content Analysis Methodology v1.0.
9. Limitations
No single signal is universally definitive. Missing metadata is normal on most platforms. Absent content credentials carry no origin meaning. Statistical text analysis cannot establish authorship, and heavy editing or translation can both create and erase its signals.
Confidence
Engine v1.0 publishes qualitative confidence only — high, moderate, limited or insufficient. No numeric percentage is shown, because the methodology has no calibrated basis for one. Confidence describes how strong the technical signals are. It says nothing about the identity or honesty of the person or business that supplied the content.
Separation from the Trust Score
Content Authenticity is an independent transparency layer. It does not change the TSDIQ Trust Score, customer satisfaction, GroundTruth results, community intelligence or HERO recognition.
A business using AI-generated marketing images does not become less trustworthy on TSDIQ. A genuine photograph with no metadata does not become more trustworthy.
What TSDIQ does and does not claim
| Claim | Status | Note |
|---|---|---|
| TSDIQ analyses content for available AI-origin and provenance signals. | Supported | This is exactly what the engine does and reports. |
| TSDIQ provides evidence about content origin. | Partially supported | Only where readable signals exist. Where they do not, TSDIQ reports insufficient evidence. |
| TSDIQ can tell whether any image is AI-generated. | Not supported | No detector available to TSDIQ can do this reliably, and TSDIQ does not claim it. |
| TSDIQ proves who created an image, video or text. | Not supported | Authorship cannot be established from these signals. |
| Frame-level video forensics and licensed provider watermark detection. | Future | The architecture supports pluggable providers; none is enabled in engine v1.0. |
Signed-in users can submit content for analysis from the Content Authenticity workspace.