Can You Really Tell If a Photo Is AI? We Tested 17 Images
Most people can no longer assume that a striking photograph came out of a camera. Synthetic images now circulate in news feeds, product listings and social posts, and the practical question has shifted from "is AI involved?" to "can I check this in a repeatable way without special training?" The answer is partly yes, but only if the check is structured as a workflow rather than a gut reaction. In a recent hands-on exercise, the AI photo check test put seventeen images in front of reviewers and showed how unreliable a single visual impression can be. This article builds on that experiment and on the parent piece by James Whitmore, which framed the problem as a question of origin rather than appearance. PromoPilot™ works with website owners and SEO specialists who need to verify visual assets before they publish them, and the procedure below reflects the kind of checks that such teams apply in practice.
The goal here is narrow: a verifiable sequence that a non-specialist can run, record and defend. It does not attempt a general survey of generative imagery. Instead, it separates provenance evidence, metadata evidence, optical evidence and factual evidence, because each category fails in different ways and each can be documented. By the end, the reader should have a concrete procedure, a criteria checklist and a method for stating how confident they are in the conclusion.
When they conflict, the honest outcome is a lower confidence and a clear note about the conflict
What You Need Before You Start
The first requirement is the original file. A screenshot, a messaging-app forward or a copy re-saved from a social platform has usually been recompressed, resized and sometimes stripped of metadata. Those transformations destroy the very signals you are trying to read. Request the highest-resolution version from the publisher, or locate it through the platform's original upload. If you cannot get the original, record that limitation explicitly, because every later conclusion becomes weaker.
The second requirement is a set of tools that cover three functions: reverse image search, metadata inspection and high-magnification viewing. Reverse search engines such as Google Lens or TinEye help locate earlier appearances, which is often more informative than the pixels themselves. A metadata viewer that reads EXIF, XMP and C2PA content credentials lets you examine camera fields and signing information. A zoom-capable viewer allows inspection at pixel level without the blur introduced by aggressive preview rendering.
The third requirement is context. Write down the claimed source, the photographer's name if given, the stated capture date, the platform where you found the image, and whether the image appears cropped or re-uploaded. Note also whether the caption makes a checkable factual claim, such as a place, a time or an event. Those claims become the basis of the final cross-check, so capturing them at the start prevents drift later.
Step-by-Step Verification Sequence
Start with provenance, not pixels. Run a reverse image search and trace the earliest occurrence you can find. Identify the first publisher, the account or outlet, and the date of first posting. An image that appears first on an anonymous account with no posting history, and then spreads through reposts that cite each other, warrants more scrutiny than one that originates from an established photographer's portfolio. Check whether the originating account is known, active and consistent in its past posting, since a sudden burst of unrelated dramatic imagery is a pattern worth noting. Missing or stripped metadata at this stage should be treated as a neutral signal. Most platforms remove EXIF data on upload, so absence alone proves nothing about how the image was made.
Next, read the metadata and content credentials. Check the camera make and model, lens data, exposure settings and timestamps for internal consistency. A file that claims a particular phone model but carries lens parameters that model cannot produce, or timestamps that postdate the stated event, has a contradiction you can document. Look also for C2PA content credentials or similar provenance manifests, and where they exist, verify the signing chain rather than accepting the presence of a credential as proof. Metadata is easy to remove and easy to forge, so its role is to corroborate or contradict other evidence. A consistent metadata record strengthens a conclusion; it rarely settles one by itself.
The third stage examines structural and optical coherence. Lighting is the most reliable place to start. Determine the dominant light direction from highlights and then check whether shadows fall consistently away from that source across the whole scene, including objects in the background. Examine reflections in glass, water and polished surfaces: a reflection should match the geometry of what is supposed to be reflected, and specular highlights should line up with the light sources you identified. Look at repeating textures such as brick, tiling or foliage, where generated images sometimes show unnatural periodicity or texture that changes character across a surface. Finally, assess background blur and depth of field. Blur should grade smoothly with distance from the focal plane, consistent with the lens implied by the metadata, and edges between sharp and soft regions should behave plausibly.
The fourth stage is fine detail at pixel level. Hands, teeth, ears, jewelry chains and background text are common failure zones, but the important word is "common," not "always." Count fingers and check joint placement, while remembering that hands are often partly occluded and that newer models handle them much better than earlier ones. Inspect signage and printed text for letter formation, consistent spacing and readable words, since text can look convincing at thumbnail size and collapse at full zoom. Study the edges of hair, foliage and fabric, where generation frequently blurs or fuses shapes that a camera would resolve as distinct. Record what you see and where, with coordinates or crop references, so another reviewer can reproduce your observations.
The final verification step is the factual claim. Confirm the location, weather, event and people against independent sources. If the caption places a flood on a specific street, compare the buildings, signage and vegetation against street-level imagery and against other photographs taken at the same time and place. If the image shows a named person at a named event, check whether that event has documented coverage and whether the person's presence is recorded. Where feasible, contact the claimed photographer or organization directly. A confirmed factual match from two or more independent sources is usually stronger than any visual argument, and a mismatch in the factual record can be decisive even when the pixels look clean. Before relying on any verdict, it helps to know what a tool like AI photo check test actually measures, since a result is only as meaningful as the input and the method behind it.
Criteria Checklist for Applied Use
A checklist keeps the procedure honest, because it forces each category of evidence to be considered and recorded rather than skipped when the picture looks obvious. Use the following criteria and mark each one as supporting, contradicting or not determinable:
- Earliest known source identified and traced to a first publisher
- Original high-resolution file obtained, or its absence noted as a limitation
- Metadata read and cross-checked for internal consistency, with absence recorded as neutral
- Content credentials, where present, verified through their signing chain
- Lighting direction and shadow geometry consistent across the scene
- Reflections and transparent surfaces physically plausible
- Hands, text and fine edges examined at full zoom with locations recorded
- Factual claim verified against at least two independent sources
- Confidence level recorded together with the specific signals that support it
Common Mistakes in Practice
The most frequent error is over-relying on a single visual tell. Practitioners often assume that extra fingers or garbled signage will always reveal a synthetic image. Those artifacts were common in earlier generations of models, but newer systems have reduced them substantially, so their absence proves little and their presence is only one data point. A similar trap is treating a smooth, polished look as evidence of generation. Professional retouching, heavy skin smoothing and studio lighting can produce the same visual effect, and a real photograph of a product or a portrait may look more artificial than a poorly lit snapshot.
A second mistake is misreading compression and editing as generation. Heavy JPEG compression, noise reduction, sharpening and platform processing all introduce artifacts that can resemble synthetic textures or halos. Legitimate photographs are also routinely cropped, filtered, composited and retouched, and each of those steps can produce irregularities. The safer approach is to ask whether a given irregularity is explained by the processing history you know about, and to record that explanation rather than attributing it to generation by default.
The third mistake is trusting detection tools as final verdicts. Automated classifiers can be useful for triage, but their accuracy depends on the training data and the type of image, and their false positive rates on heavily edited real photographs can be high. A tool's output should be recorded as one input among many, never as the conclusion. Without documenting the reasoning chain, including which signals were checked and what each showed, the conclusion cannot be verified or challenged by anyone else. Finally, many reviewers stop at visual inspection and skip provenance and factual verification. That is a costly omission, because tracing the earliest source or confirming the event often resolves the question faster and more reliably than any amount of pixel analysis.
Reporting Your Conclusion
A binary answer, "AI" or "real," overstates what most workflows can establish. A more defensible report states a confidence level, such as high, moderate or low, and lists the specific signals behind it, with the contradicting signals stated alongside them. For example, a report might say that the earliest source is an anonymous account posting for the first time, that the metadata is absent and therefore neutral, that lighting is consistent, that a sign in the background is unreadable at full zoom, and that the factual claim about the location could not be confirmed. That report tells the reader exactly where the uncertainty lies and what would change the assessment.
The strongest conclusions come from convergence. When provenance, metadata, optical evidence and the factual record all point in the same direction, confidence can be high even without a single decisive signal. Teams that publish or rely on visual assets benefit from adopting this structure as a standard, because it turns an intuitive judgment into a repeatable process that colleagues can audit. For website owners and SEO specialists working through this kind of verification, the same discipline applies to any image that carries a factual claim, and it is the most reliable way to keep unverified visuals from undermining the credibility of the pages they appear on.