Image checks that do not pretend to be certainty
We assess whether a submitted image fits your case. Instead of a confidence score from a detector, your team gets a documented judgement they can act on.
No detector can tell you an image is real. A sound process can tell you whether it is plausible.
Claims photos, delivery proofs, application documents and ID copies arrive digitally and are usually glanced at rather than checked.
AI-generated and edited images can no longer be reliably spotted by eye, not even by experienced case handlers.
Detector tools create false confidence: they flag genuine photos as fakes and miss good fakes, with high error rates in both directions.
Provenance and metadata
- Reading Content Credentials based on the C2PA standard where they are present
- EXIF and metadata analysis: capture device, timestamps, editing traces, software chains
- An honest reading of what missing metadata means and, more importantly, what it does not mean
Reuse and image history
- Reverse search against images already published or submitted elsewhere
- Cross-referencing your own earlier cases to find the same picture used more than once
- Detecting crops, mirrored versions and re-encoded copies of the same source image
Consistency with the case
- Checking the image against the reported location, time and weather conditions
- Matching against contract, asset and master data from your own systems
- A documented assessment with reasoning, so your decision holds up under audit
Suspicious cases get prioritised instead of every image being treated alike.
Every assessment is reasoned and documented, including for audit and internal review.
Your people know how solid a signal is before they act on it.
Deepfake detection
Can you tell for certain whether an image was generated by AI?
No, and nobody can credibly promise that. AI image detectors have high error rates in both directions. They flag genuine photos and miss good fakes. We therefore treat them as one signal among several and never as the sole basis for a decision.
Are C2PA Content Credentials not proof of authenticity?
Content Credentials are useful, but they are not proof. The data is lost when an image is edited, sent through a messenger or screenshotted, and it can be stripped deliberately. Missing credentials are not evidence of a fake, and present credentials are not a clean bill of health.
So how do you reach a defensible conclusion?
By combining methods. Technical analysis of metadata and image history, cross-referencing your case data such as location, time and weather, and a human decision at the end. That combination holds up. A single detector score does not.
What does the EU AI Act require from 2 August 2026?
Article 50(4) requires disclosure when content has been generated or manipulated by AI. Read honestly: the obligation does nothing against fraudsters, because criminals ignore it. It does show that regulators recognise the risk, and it applies directly to you if you deploy AI-generated content yourself.
Does a system decide, or a person?
A person. The system sorts, explains and presents. Approval, rejection and follow-up questions stay with your team.
Talk to us about your image workflow
Send a short description of your review process to info@cavrix.de. We will tell you plainly what is technically possible and what is not.