AI detector photo analysis — how detection actually works
Detectors do not recognise “AI”. They measure a handful of physical and documentary signals, then weigh them against each other.
The five signals a detector can look at
Every credible detector builds its answer from a small set of measurements, and each measurement has a different failure mode. Metadata and provenance records are documentary evidence: they say what the file claims about its own origin, and a signed credential is among the strongest evidence you can get — but most cameras and most generators still attach nothing at all. Watermarks work the same way in reverse: a hit is meaningful because the mark had to be written at generation time, while a miss tells you only that this particular provider did not mark this particular file.
The remaining signals are statistical. Frequency-domain analysis looks for the regular patterns that upsampling leaves behind, because generators typically build an image at low resolution and scale it up, which introduces periodic structure a camera never produces. Noise-residual analysis asks a different question: does the fine-grained texture behave like sensor noise, which is stronger in the shadows and differs from frame to frame? A generated image often has noise that is too uniform, or noise that sits on top of the image instead of inside it. Finally, semantic checks look for physical impossibilities — a reflection that does not match its subject, a shadow pointing the wrong way, text that dissolves into shapes.
| Signal | What it measures | Reliability |
|---|---|---|
| Metadata and provenance | EXIF blocks, XMP packets and signed C2PA content credentials — who or what claims to have made the file. | Strong when present and signed; meaningless when absent |
| Watermarks | Invisible marks embedded at generation time by a specific provider, readable only by that provider's verifier. | Decisive on a positive hit; a negative result proves nothing |
| Frequency-domain fingerprints | Periodic patterns left by upsampling and deconvolution, which show up as regular peaks in the frequency spectrum. | Moderate — degrades with resizing, cropping and re-compression |
| Noise residuals | Whether the fine-grained noise in the image behaves like sensor noise, which is stronger in shadows and varies per frame. | Moderate — denoising and HDR stacking erase the same signal |
| Semantic and physical consistency | Whether shadows, reflections, hands, text and geometry obey the physics of the scene they appear in. | Weakest and most subjective — needs a human reviewer |
Why real photographs get flagged
False positives are not a bug in one tool; they follow from what the tools measure. A photograph that has been denoised loses exactly the noise residual a detector looks for. A long exposure smooths water and clouds until the texture statistics resemble a generated render. Focus stacking, HDR merging and heavy noise reduction all push a genuine photograph toward the statistical profile of a synthetic one. Add a screenshot or a re-export that strips metadata, and a real photo can arrive looking as bare as a generator output.
The opposite error is just as common. A generated image that has been resized, cropped and re-compressed a few times loses the frequency fingerprints that gave it away, and if its metadata was cleared along the way, a metadata-based check finds nothing to report. This is why a single score is such a poor basis for a decision, and why the honest answer to “is this AI?” is usually a list of observations with a confidence level attached, not a verdict.
A detector is not a verdict. It reports what its signals found. Use it to decide where to look next — the file's provenance, the person who sent it, the original capture — not to close the question.
What this means when you receive a file
Work from the strongest evidence down. Ask for the original file rather than a re-export, because the original still carries the metadata and credentials that an export has lost. If the file claims a camera, check that the claim is internally consistent — make, model, lens and exposure should describe a real combination. Then look at the weaker statistical signals as supporting context. If you want to see what a specific file carries before you form a view, you can inspect its metadata in your browser without sending it to anyone, or compare the available detector options and their trade-offs first.
How image source affects detector accuracy
A photo shot directly on a modern smartphone and exported without editing arrives with full EXIF — GPS, shutter speed, lens model — and unmodified frequency structure. Detectors reading statistical texture have the cleanest possible signal, and metadata validators can confirm hardware consistency. The result is the most reliable reading you will get. The moment that image passes through a social platform, however, it is re-encoded, often stripped of EXIF, and resized to fit delivery constraints. Instagram, WhatsApp, and similar services apply their own compression passes, which flatten some of the very frequency anomalies detectors rely on. An AI-generated image that survived one compression round may score closer to “real” after a second, and a genuine photo may pick up artefacts that push its score in the wrong direction.
Screenshots introduce a third layer. A screenshot is a pixel capture of a rendered surface — it has lost all camera metadata by definition and carries the compression signature of the screen-capture codec, not any original source. Running a screenshot through an AI detector tells you almost nothing reliable about the underlying content; the tool is measuring the screenshot's own encoding, not the image that was displayed. If the original file is accessible, always test that instead. If only a screenshot exists, treat any detector score as low-confidence and weight metadata signals at zero. Understanding which generation of compression you are working with is the single most important context you can bring to an in-browser metadata inspection.
Building a reviewable detection record
A single score is not evidence — a documented process is. Before running any test, save the exact file you intend to submit: note its filename, file size, and SHA-256 hash so you can prove later that the file was not modified between collection and analysis. Record the tool name, its version or URL, the date and time of the run, and the full output — not just the headline percentage. If the tool provides a breakdown by signal type (texture, metadata, frequency), capture that too. Store everything together in a timestamped folder. This discipline costs two minutes and converts a disposable result into something a colleague, editor, or adjudicator can independently review.
Running a second tool against the same file is not optional when stakes are high. Different detectors weight signals differently — one may emphasise EXIF consistency while another focuses on pixel-level frequency. Where both tools agree, confidence rises. Where they diverge, the disagreement itself is informative: examine which signal each tool is responding to and whether the source-compression history you documented explains the gap. Cross-referencing two or more detector tools and logging their individual rationale is the standard practice in any context where the finding will be used to support a decision or dispute.
What a detector can and cannot actually see
A detector reads what is in the file — and only what is in the file. It sees the pixel array: a grid of colour values that it can subject to statistical analysis, frequency decomposition, and noise modelling. It sees the metadata headers that the software or camera attached when the file was written. It sees, if they exist, any embedded watermarks or cryptographic content credentials. Nothing else is accessible to it.
That leaves out a great deal. A detector cannot see whether the image was generated with a human prompt or as part of an automated pipeline. It cannot see whether a genuine photograph was later composited with AI-generated elements — a real sky swapped for a generated one, for instance — unless the compositing left a measurable seam in the pixel statistics or the metadata now contradicts what the pixel content implies. It cannot tell you who made the editorial decision to publish the image, or why. It cannot distinguish between a photo that was run through an AI upscaler after capture (which alters frequency fingerprints) and one that was generated from scratch, because both operations leave similar artifacts at the statistical level.
The boundary matters most when people use detector scores as a proxy for questions the tool was not built to answer. A high AI-probability score does not mean the image was generated with intent to deceive. A low score does not mean the image is authentic in any editorial sense. A detector answers one question — does this file show statistical properties more consistent with a generator or a camera — and that is all it answers. Any broader interpretation is an inference the human reviewer supplies, not something the tool produces.
Concrete false-positive and false-negative patterns
False positives — flagging real photographs as AI-generated — cluster around a predictable set of post-capture operations. Heavy noise reduction is the most common trigger: modern denoising algorithms, including the AI-assisted ones built into Lightroom and Capture One, replace organic sensor noise with smoother texture that reads statistically similar to generator output. Astrophotography and low-light stacking are especially prone to this because they deliberately average away the per-frame noise variation that detectors rely on. Long exposures of water and clouds flatten texture across the frame in the same way. Focus stacking, which merges multiple focus planes, and HDR merging, which blends exposures, both produce images whose texture statistics sit between the profiles the detector was trained on.
Platform re-encoding adds a second layer of false positives. A real photograph extracted from a WhatsApp thread or downloaded from an Instagram post has already been through at least one aggressive compression pass. The frequency structure is altered, the EXIF is stripped, and the file size is a fraction of the original. A detector reading that file sees exactly the profile it was trained to associate with synthetic output. It is technically correct — the file does not look like a raw camera file — but the cause is platform processing, not generation.
False negatives — AI-generated images that pass as real — follow the inverse logic. Any post-generation processing that degrades the generator's fingerprints reduces detectability. Resizing the image to a smaller resolution and re-saving it as JPEG at medium quality removes most of the frequency artifacts that frequency-domain analysis depends on. A second resize in the other direction blurs them further. Adding grain in Photoshop or Lightroom partially restores the noise residual that the detector looks for, pushing the score toward real. Some generators, particularly those trained on large datasets of real photography, already produce noise profiles that are harder to distinguish; their outputs start closer to the boundary before any post-processing. A generated image that has also had plausible EXIF injected — a practice that requires only a free command-line tool — will beat metadata checks entirely and rely entirely on the pixel-level analysis to catch it.
Metadata signals versus pixel-layer signals, and how this site reads your file
The distinction between metadata analysis and pixel analysis matters practically because they answer different questions with very different confidence levels. Metadata is documentary: EXIF fields, XMP packets, and C2PA content credentials are structured text records attached to the file at write time. Reading them requires no image decoding at all — a parser extracts the fields and checks whether the values are internally consistent. This is fast, deterministic, and, when the data is present and signed, quite reliable. The weakness is availability: most files that have passed through any editing or sharing step arrive with metadata stripped. When the metadata is absent, this entire class of evidence is simply gone.
Pixel-layer analysis is the opposite in almost every respect. It requires decoding the full image, computing statistical summaries over the colour values, applying transforms like the discrete cosine transform to examine frequency content, and comparing the resulting signatures against trained reference distributions. It is slower, noisier, and more sensitive to file history — every resize, re-compress, and filter operation shifts the statistics. The output is a probability estimate, not a match, and the error bars widen as the file's processing history grows longer.
When you use the metadata inspection tool on this site, the analysis runs entirely inside your browser. The file is read by your device's JavaScript engine using the Web File API; nothing is uploaded to a server. The tool decodes the EXIF and XMP blocks from the file binary directly in memory, checks field consistency, and looks for C2PA manifest records if any are present. What you see in the results is drawn from the file's own headers. Because the computation stays local, files you would not share externally — internal documents, pre-publication images, anything subject to a confidentiality obligation — can be checked without leaving your machine. The trade-off is that the browser-side tool focuses on the documentary layer: the metadata fields it can parse quickly and without model inference. For the deeper pixel-level statistical analysis, you would need to use one of the dedicated tools listed on the detector comparison page, which do require the file to be uploaded. Knowing which layer you need to examine — and which layer a given tool actually examines — is the first decision to make before you start.
What detector results can and cannot support
In news verification, a detector score is useful as a triage filter — it can flag images that warrant deeper reverse-image search, source tracing, or expert forensic review. It is not, on its own, a publishable finding. Responsible fact-checkers treat a high AI-probability score as a reason to escalate investigation, not as proof of manipulation. In e-commerce return and fraud disputes, a metadata inconsistency — a file dated before the claimed purchase, or camera data that does not match the stated device — can support a case when combined with other evidence, but a detector percentage alone will not satisfy a platform's dispute resolution process. Document the inconsistency precisely and pair it with purchase records or communications.
Copyright and licensing disputes are the most demanding context. Courts and arbitration panels expect traceable chain-of-custody evidence, reproducible methodology, and, in most jurisdictions, expert testimony to interpret technical findings. A detector result produced months after the fact, without a contemporaneous hash record, will face credibility challenges. Use detector output as an early indicator to decide whether deeper forensic investment is warranted — not as the forensic report itself. If a dispute is heading toward formal proceedings, engage a qualified digital-forensics professional who can produce a defensible methodology report. The value of tools like the ones compared on AI image detector comparisons is precisely in helping you make that early triage call quickly and without specialist software.