A photo is still evidence — it simply should not be the only evidence[2]
In e-commerce, images serve a concrete operational purpose. They document product condition, the type of defect, packaging, variant consistency and sometimes the way a parcel was protected. This lets companies resolve many cases without physical expert inspection.
The problem is that the cost of image manipulation has collapsed. Generative tools can change one local region without recreating the whole scene. A user no longer needs professional retouching skills; a prompt, a mask and several attempts may be enough.
The FBI notes that criminals use generated images to create believable profiles, identification documents and other visual material that supports fraud. That is not evidence that AI-faked e-commerce claims are already universal. It is, however, a strong reason to stop treating a single JPEG as an automatically neutral source of truth.
What can be forged in claim evidence?[3][4]
The simplest scenario is local editing: adding a scratch, crack, stain, tear or deformation. More advanced manipulation can replace part of a product, a label, a serial number or the background. A whole photograph can also be generated without the claimed object ever being in front of a camera.
From a process perspective, it matters less whether the tool was a generative model, Photoshop or a simple composite. The business question is whether the image is a reliable representation of the item's state at the claimed time.
That is why a detector that only answers “AI-generated” cannot solve the entire problem. Local-manipulation signals, provenance, metadata consistency and the rest of the case evidence also matter.
NIST applies a similar principle in a higher-stakes process: remote identity proofing[1]
NIST SP 800-63A discusses digital injection and forged media in remote identity proofing. Generative AI can create or modify images and video of people and identity evidence to defeat document validation, biometrics and human reviewers.
The guidance is useful beyond identity systems: submitted media should be analysed for manipulation and forgery indicators; automated algorithms need testing against both genuine and forged material; and algorithmic decision-making should be augmented by manual review because false positives and false negatives exist.
That is a sensible model for customer claims too. A detector should not “reject the customer”. It should identify risk that changes the route by which the case is handled.
High manipulation risk should trigger additional verification, not an automatic accusation of fraud.
Evidence handling starts before any AI analysis
The best detector cannot recover information that the business threw away. If the support system keeps only a resized screenshot, the original metadata, encoding structure and provenance may already be gone. Whenever possible, preserve the original upload, filename, receipt time, byte size and SHA-256 hash.
When a customer submits several images, treat them as a set. Relationships across views can be more informative than one score. Then add operational context: SKU, variant, warehouse photos, parcel weight, packing data, serial number, claim history and, when relevant, outbound or return documentation.
Technical image analysis should sit on top of that evidence base, not replace it.
Six layers are more useful than one AI score[3][4]
A practical system can combine distinct evidence classes: provenance and C2PA where available; EXIF/XMP and file-container structure; local pixel and compression forensics; several independent AI/deepfake detectors; consistency across multiple images; and business context.
The absence of one source should not automatically turn the rest into zero. C2PA may not exist, EXIF may have been stripped by a messaging app and one detector may be unavailable. What matters is how many independent layers remain and whether their conclusions align.
At the same time, missing layers must be visible. A four-layer report should not pretend to be a six-layer report. Transparency about coverage is part of forensic credibility.
Build a decision rule, not a detector threshold
The weakest policy is “AI > 50% = reject”. Detector outputs are not probabilities that a customer is guilty, and tools have different error profiles. A three-route workflow is safer.
- Low risk: no strong manipulation signals and consistent files/context — continue through the normal claim process.
- Unclear: isolated anomalies or disagreement between detectors — request original files, additional views or return the product for physical inspection.
- High risk: several independent layers indicate manipulation or direct provenance identifies generative processing — pause automatic decisions and route to manual review.
- Formal dispute: preserve files, hashes, the technical report and full case history; do not make claims about intent that the image analysis cannot establish.
Why additional images are disproportionately useful
One image can be edited convincingly. A series from different angles requires geometry, lighting, material texture and the claimed defect to remain consistent across several views. Asking for additional photographs therefore increases the information available for comparison rather than merely adding bureaucracy.
Requests should be specific: the whole product, a close-up of the issue, a label or serial detail, and a view under different lighting. The goal is not to make a legitimate claim difficult; it is to obtain material that can actually be evaluated.
When additional files come directly from a camera, they may also restore metadata and structural information lost in the first platform-processed attachment.
The technical and legal boundary: what a report should not claim[1][3]
A technical report may show strong signs of synthetic origin, local manipulation, inconsistent container structure or direct provenance. It cannot by itself establish who made the change, exactly when it happened or why.
That distinction matters in customer communication. “The file contains technical anomalies, so we need additional verification” is fundamentally different from “you falsified the image”. The first describes evidence; the second attributes conduct and intent.
A good anti-fraud system reduces error costs in both directions: it limits losses from fabricated claims while also reducing the risk of wrongly rejecting an honest customer.
A reasonable minimum standard for e-commerce in 2026[1][3][4][5]
Not every merchant needs a forensic laboratory. A relatively simple standard materially improves decision quality: preserve originals, request additional evidence in ambiguous cases, use several technical signals and document the reasoning behind the decision.
As Content Credentials become more common, more files will carry verifiable history. At the same time, generators and editing tools will improve, so detection will remain a moving target. The process needs to be robust to changes in individual tools rather than tied to one detector.
That is the biggest change AI brings to customer claims: images remain valuable evidence, but their reliability increasingly needs verification just like a payment, tracking number or order history.