Our method

BA-METHOD 1.0

Bureau Attest does not attest that an image represents the truth: we document what the examination of the file reasonably allows to conclude about its origin and possible transformations. Every case follows the five phases below.

1. File identification and integrity

Format, dimensions, relevant technical characteristics, and immediate computation of the SHA-256 fingerprint. The report is then bound to this precise file. The fingerprint does not prove the file is authentic: it allows verifying that the report concerns the very same file.

2. Available technical information

Metadata, container, EXIF/XMP information where present, profiles, encoding traces, observable software history, and provenance data such as C2PA / Content Credentials when they exist. Missing metadata is not evidence of AI generation; present metadata is not evidence of authenticity — it can be forged and is treated as a signal.

3. Visual and structural examination

Search for anatomical, geometric, lighting, typographic, texture, depth, repetition, seam or compositing inconsistencies; for video, temporal consistency. An isolated artifact is never presented as proof.

4. Provenance and complementary tools

Search for earlier versions or sources when the case warrants it, comparisons, and specialised analysis tools. Essential principle: automated outputs are signals, never the verdict on their own.

5. Human synthesis

The examiner weighs the elements for and against each hypothesis, looks for contradictions, and retains the most cautious conclusion compatible with the whole. The report presents the conclusion, its confidence level (limited, moderate or high), the key findings and any reservations.

No single signal ever determines the conclusion on its own.

Possible conclusions

Consistent with authentic content — the examined elements do not reveal sufficiently probative signs of generation or substantial modification.
Likely AI-generated — several concurring elements are consistent with synthetic generation.
Likely modified or composite — concurring elements indicate substantial modification, composition, or the introduction of synthetic content.
Inconclusive — the available elements are insufficient, contradictory or too degraded; no charge is made.

The confidence level expresses the strength and convergence of the observed elements. It is not a statistical probability of accuracy.

A method built to evolve

Techniques for generating, modifying and distributing digital content evolve quickly. The methods used to examine them must evolve too.

Bureau Attest uses a versioned methodology, based on the confrontation of several families of signals rather than a single tool or detector. The method is regularly re-evaluated as new tools, formats, provenance standards and generation techniques appear. The version applied is stated on every report.

Significant protocol changes are documented. Certain operational details, tools and internal criteria are however not published, to preserve the integrity of the method.

Errors and revisions

Any documented dispute sent within 14 days of delivery (reply to your delivery email) leads to a free re-examination, processed within 48 hours. If we were wrong, the report is revised — the revision is dated, never silent — and you are refunded in full. A commercial refund never changes a conclusion.

Reports are documented opinions delivered under an obligation of means. They are neither judicial expertise, nor official certification, nor a guarantee. They cover the examined file only — not the identity or intentions of its author, nor the truth of the depicted event.