4 Signs a Damage Photo May Be AI-Generated
Generated damage photos often break the rules of light, text and physics. Learn four signs worth checking, why none is proof on its own, and how to ask for more evidence on the platform.
A buyer sends a photo of a shattered mug, and something about it feels wrong: the shadows don't quite agree with each other, and the crack looks painted on. Generative image tools can now produce convincing pictures of damaged products from a short text prompt, so sellers on Shopee, Lazada and other marketplaces are starting to ask a new question about some damage claims: was this photo taken, or generated?
This guide covers four tells of AI fake damage photos: signs that can suggest an image is generated or edited, why each one happens, and how to respond politely and on the platform. None of these tells proves anything on its own. Real photos taken in poor light, compressed by a chat app or cropped on a phone can show some of the same signs. Treat them as reasons to ask for more evidence, never as a verdict about the buyer.
Key Takeaways
- Four signs are worth checking on a suspicious damage photo: inconsistent lighting and shadows, garbled text or serial numbers, damage with impossible geometry, and file details that don't fit a phone photo.
- Each tell is a signal, not proof. The right response is a neutral request for more evidence (another angle, a close-up of a label you know, or a short video), made inside the platform's chat.
- Your strongest counter-evidence is your own: packing photos and videos captured before dispatch.
Why AI Fake Damage Photos Are Hard to Spot
Text-to-image systems don't model the physics of a scene; they predict pixels that look plausible. Research by Hany Farid at UC Berkeley explored this for DALL-E 2, noting that such systems are "not based on explicit geometric or physical models" and that the human visual system is generally insensitive to lighting inconsistencies (Lighting (In)consistency of Paint by Text, arXiv 2022; see also Perspective (In)consistency of Paint by Text). In other words, a generated image can look fine at a glance while breaking rules a careful look can catch.
Two cautions apply. First, image generators improve quickly, and checks that work on one generation of tools may not work on the next. Second, most damage claims are genuine. Use these tells to decide when to ask for more evidence, and keep your tone the same for every buyer. For a broader overview of this problem, see our guide to AI-generated fake damage claims.
1. The Lighting Doesn't Agree With Itself
What to look for. In a real photo, light comes from one source or a consistent set of sources. Shadows fall in compatible directions, highlights sit on the sides of objects facing the light, and shiny surfaces reflect things that exist in the scene. In a generated or heavily edited image, you may see shadows pointing in different directions for objects side by side, packaging lit from above while the product inside looks lit from the side, or a reflection of a window that appears nowhere else.
Why it happens. Because generators predict appearance rather than simulate light, different parts of an image can carry different implied lighting. Edits that paste damage onto a real photo can create the same mismatch.
Innocent explanations. Mixed lighting is common in real homes: a window on one side, a ceiling lamp on the other. A phone's HDR processing can also produce odd-looking highlights.
How to respond. Ask for one more photo in neutral language, inside the platform chat: "Sorry to hear about this. Could you send one more photo showing all the pieces together next to the original packaging? It helps us match the item to our shipment record." A buyer with a real broken item can usually do this quickly.
2. Text and Serial Numbers Look Almost Right
What to look for. Labels, batch codes, serial numbers, warranty stickers and ingredient lists that are slightly wrong: letters that bend or merge, extra or missing digits, a logo with the right colours but the wrong letter spacing, or words that don't exist.
Why it happens. Image generators have historically struggled to render long, precise text, because they learn what text looks like rather than which characters belong where. Newer tools are better at short words and familiar logos, so this tell is weaker than it used to be, but long serials and dense small print are still worth checking.
Your advantage. You know what the real label looks like, because you shipped the product. Compare the buyer's image with your own product photos or a unit from the same batch.
How to respond. Ask for a close-up of a specific area you can verify: "So we can match this to the right unit, could you send a close-up of the batch code on the back label?" Be honest about why you're asking. Don't invent reasons, such as a courier or insurer requirement, that aren't true; a request that is simply about matching the item to the order is reasonable on its own.
3. The Damage Has Impossible Geometry
What to look for. Damage that doesn't behave like a physical event: cracks in glass that float on the surface without depth or stop neatly at a boundary they should cross, a dent in a box with no matching shadow, a tear in fabric with no loose fibres, or a "spill" with a glossy highlight but no wet edge where it meets the surface.
Why it happens. Generators produce damage that looks typical rather than damage caused by a specific impact on a specific object. Farid's perspective research explored the related problem that these systems don't model geometry explicitly, which is why structure and perspective can drift.
Innocent explanations. Some materials break in unusual ways, and phone photos of glass are hard to read because of glare. Low resolution can make real cracks look flat.
How to respond. Ask for a short video showing the damage from more than one angle. Shopee's own guidance already asks buyers for "a clear video demonstrating that the product is not functioning as intended" for faulty-product returns (Shopee Help Centre), so asking for video is consistent with how platforms already handle evidence. A video of real damage shows depth and movement that a single still image can't.
4. The File Doesn't Look Like a Phone Photo
What to look for. If you can access the original image file, check its details. Photos from a phone camera usually carry metadata such as the camera model and the time the photo was taken. Some generated images carry no camera metadata at all, and some tools add provenance information. The C2PA standard, for example, describes Content Credentials as "a nutrition label for digital content" that records a file's origin and edits (C2PA), and some tools and devices now attach them. Other signs worth noting are a perfectly uniform, noise-free look or a crop and resolution that don't match the buyer's other photos.
Why this tell is weak on its own. Metadata is easy to lose. Messaging apps, screenshots and upload processes often strip it, so missing metadata is common for perfectly genuine photos. Treat file details as supporting context, never as the deciding factor.
How to respond. Keep everything on the platform. Don't ask the buyer to send files by email or outside messaging apps, and don't share your own contact details. Shopee's seller chat policy, for example, treats directing transactions off the platform as a chat violation and censors messages containing details such as contact numbers (Shopee chat violations). Evidence submitted through the platform is also what reviewers see. If the file details raise a question, ask for another photo or a short video through the platform chat instead.
Your Own Evidence Matters More Than Any Tell
The four tells help you decide when to ask questions. What usually carries a case is your own record of what you shipped. A packing photo or short video showing the item intact, the label visible and the parcel sealed gives a platform reviewer something concrete to compare with the buyer's image. Shopee says it reviews the evidence submitted by both buyers and sellers before deciding a dispute (Shopee Help Centre), so what you can show matters. If you don't capture packing evidence yet, our packing station evidence SOP explains a simple setup.
When you submit a response, describe what you observed in neutral terms. For example (an illustrative wording): "The buyer's photo shows cracks that stop at the bezel; our packing video at 14:02 shows the screen intact and the unit wrapped in two layers of bubble wrap." Let the reviewer draw the conclusion. Our Shopee dispute evidence guide covers how to organise and upload evidence before the deadline.
Quick Triage Checklist
| Tell | What to check | Neutral next step |
|---|---|---|
| Lighting | Do shadows and highlights agree across the image? | Ask for one more photo with the packaging |
| Text and serials | Do labels and codes match your real product? | Ask for a close-up of a label you can verify |
| Geometry | Does the damage look physically plausible? | Ask for a short video from two angles |
| File details | Does the file look like a phone photo? | Treat as context only; ask for video in-platform |
| Your records | Do you have packing photos or video? | Attach them, tied to the order ID |
What Not to Do
- Don't accuse the buyer of using AI or of fraud, in chat or in your dispute response. Describe what you observed and let the platform decide.
- Don't make up reasons for your requests. Asking for more evidence to match the item to the order is honest and enough.
- Don't move the conversation off the platform or ask for files by email.
- Don't ignore a claim because you suspect the photo. Respond inside the deadline, with your own evidence.
Frequently Asked Questions
Can I reject a claim because I think the photo is AI-generated?
The decision on a disputed claim rests with the platform. What you can do is raise the specific inconsistencies you observed, attach your own packing evidence, and ask the buyer for additional photos or video. A suspicion on its own isn't evidence; a documented discrepancy is.
Is missing EXIF data proof that a photo is fake?
No. Metadata is often removed by messaging apps, screenshots and upload processes, so many genuine photos have none. Treat missing metadata as a minor signal at most.
What's the best evidence against a fake damage claim?
Your own record from before dispatch: a packing photo or short video that shows the item intact, the label visible and the parcel sealed, tied to the order ID. It speaks to what left your shop regardless of how the buyer's image was made.
Should I ask every buyer for a video?
Not as a blanket rule. Ask when the claim needs it, such as damage or a product not working, and use the same neutral wording for every buyer. Shopee's own guidance already expects clear photos or videos for many return reasons.
Respond Calmly, With Evidence
AI-generated damage photos are a new twist on an old problem, and the answer is the same as it has always been: calm, consistent requests for evidence and a good record of what you shipped. Hail Pilot helps marketplace sellers keep packing photos, buyer messages and dispute deadlines together in one case file, and can draft polite evidence requests for you to review before you send them. See how Hail Pilot works.
Written by Hail Pilot Editorial