An AI image detector can give you an answer without knowing who created the image, which camera captured it, or whether the scene happened. Understanding how it reaches that answer makes the result more useful—and puts a necessary limit on what it can prove.

Imagine receiving a convincing photograph with a detector result attached: “likely AI”. The temptation is to treat the result as a discovery about the file's history. But a visual classifier has usually done something narrower. It has compared patterns in the image with a distinction learned from other images.

That distinction can be informative. It can also fail. The useful question is what kind of evidence the detector has produced, and how that evidence fits with everything else you know.

Visual classification, metadata, signed provenance and watermark detection are four parallel checks.
Different checks answer different questions. The channels stay separate; this is a schematic, not a combined score.

A detector learns a distinction

In supervised image detection, training examples have labels: photographic, generated, or another category defined by the task. A learning procedure adjusts model parameters to reduce prediction error on those examples. At ordinary fixed-model inference time, the detector applies the learned rule to a new input; it does not receive that image's correct label.

The rule need not resemble a person's inspection. You might look for an impossible reflection or malformed lettering. A model can use statistical relationships that are difficult to see unaided. It might also learn distinctions associated with the datasets rather than with generation itself.

The research offers several implementations. Wang and colleagues trained a classifier on generated and photographic images. Ojha, Li and Lee investigated detection using a broadly pretrained visual representation. DIRE builds a representation from diffusion reconstruction error. These are examples of different approaches, rather than interchangeable descriptions of every detector. [1, 2, 3]

This matters when someone explains a result as “the detector found the AI fingerprint”. Without a validated explanation for that model and input, the phrase suggests more knowledge than the output supplies. A decision can be useful without exposing the particular evidence responsible for it.

The uploaded image is not always the image the model sees

A file must be decoded into image data before a visual classifier can process it. Its pipeline may resize it, crop it, normalize colour values, or examine selected patches. The exact sequence depends on the implementation.

Suppose a wide photograph contains a generated object near its edge. If the detector analyzes only a central crop, that object may be excluded. The score would then describe the information that reached the model, even though the interface displays the entire upload. That is a hypothetical pipeline example, not a claim about a particular uploaded photograph.

The original file also contains evidence that can be lost before classification. ImageSignals' technical principle is to inspect original bytes for metadata and provenance before conversion or resizing. That inspection and the model's visual preprocessing serve different purposes. Keeping the original makes it possible to ask both questions without confusing their answers. [8]

From patterns to a score, then a decision

Many classifiers output a number that increases as their learned rule favours the generated class. A threshold converts that score into a label. A result can therefore change because the score changes, because the threshold changes, or because a different model is used.

Consider an illustrative score of 0.64 and a threshold of 0.60. Under a “generated at or above the threshold” rule, it is flagged. With the same score and a threshold of 0.70, it is not. Neither decision reveals a new fact about the image's creation. The operating rule changed.

A decimal between zero and one is not automatically a reliable probability. Calibration asks whether predicted probabilities agree with observed frequencies across comparable examples whose origins are known. A single score cannot establish that agreement. Guo and colleagues studied this problem in neural network classification; their work is not an accuracy audit of ImageSignals. [4]

If an interface says “90%”, first establish whether it means a raw model output, a transformed score, a calibrated probability, or a summary of a benchmark. Those quantities answer different questions. None should silently become “90% proof”.

Four evidence channels answer different questions

Visual classification: what patterns does the model recognize?

A visual classifier estimates a category from image content or a representation derived from it. It can operate without an embedded origin declaration. Its reach depends on training, evaluation conditions and the input pipeline. [1, 2, 3]

Metadata: what does the file record?

EXIF, XMP and other fields may record a camera model, editing software or generation settings. Ordinary fields can be changed or removed. A camera name is a clue about a record in the file, not an unbroken account of every operation that produced the scene. Missing fields are similarly inconclusive. [8, 9]

Signed provenance: what recorded claims can be validated?

Content Credentials use the C2PA framework to associate assertions with assets and support checks on integrity, binding and trust. Validating those properties can strengthen the evidence for a recorded history. It does not independently establish that the event depicted is true. A genuine camera capture can show a staged scene. [5]

Watermark detection: is a supported embedded signal detectable?

Systems such as SynthID deliberately embed signals that compatible verification tools can look for. This differs from inferring generation from naturally learned statistical patterns. Coverage and robustness have limits; a negative result must be interpreted within the supported system's scope. [6]

These channels should remain visible separately. A metadata statement should not be added to a model score as if the result were a better calibrated probability. Nor should absent provenance lower the visual score. Evidence can complement other evidence without becoming arithmetic.

Why a negative result leaves questions open

“No supported signal detected” describes the outcome of a particular check. It does not establish why the signal is absent. Perhaps none was embedded. Perhaps the file lies outside the system's coverage. Perhaps a transformation affected detection. Without a controlled comparison, the result cannot identify the cause.

The original ImageSignals video includes a processed synthetic image and a verifier reporting no signals. The available record does not establish a matched, successful verification of the original before processing. The article therefore does not use that demonstration to claim that a particular edit removed a watermark.

The same restraint applies to an image classified as photographic. It may be photographic, or it may be a synthetic image the model missed. Conversely, a positive classification can be a false alarm. The labels describe predictions, not independently established ground truth.

A practical way to use the result

Begin with the question you need answered. “Was any part generated?” is broader than “does this model recognize signs of whole-image generation?” “Did this event happen?” is broader still.

Preserve the closest available original. Record which version was checked and, where available, the detector version and conditions. Read the visual result alongside any metadata or credential information. If they disagree, keep the disagreement in your notes; do not resolve it by choosing whichever label feels more decisive.

Then investigate context, including earlier appearances and alternative captions through independent source research. Within the file, ImageSignals' region search can help locate a named object; it does not search the web for earlier copies. Text visible in an image may provide another lead: extracting it can make a name or quotation searchable, but the extracted text still needs checking.

For the distinction between the evidence channels and the application's limits, consult the ImageSignals methodology. More detailed inspection belongs in advanced analysis, with each output interpreted according to what it actually measures.

Useful evidence is allowed to be incomplete

A detector does not need to be infallible to help. It needs a defined task, credible evaluation and an interpretation that respects its limits. The mistake is to turn a bounded observation into a universal verdict about a picture.

Use the result to decide what to inspect next. Preserve uncertainty when the evidence does not settle the question. That is how a classifier becomes part of image verification rather than a substitute for it.

Try the ImageSignals AI image detector.

Watch the video

Can You Still Tell What's Real? (AI Image Detection Explained) · ImageSignals · 5:06.

Watch on YouTube.

This article draws on research developed for the ImageSignals video series and expands it for readers.

Sources & further reading

  1. Wang et al., CNN-Generated Images Are Surprisingly Easy to Spot… for Now, CVPR 2020.
  2. Ojha, Li & Lee, Towards Universal Fake Image Detectors That Generalize Across Generative Models, CVPR 2023.
  3. Wang et al., DIRE for Diffusion-Generated Image Detection, ICCV 2023.
  4. Guo et al., On Calibration of Modern Neural Networks, ICML 2017.
  5. C2PA, Technical Specification v2.3, scope and validation. A pinned reference, not a claim about the latest version.
  6. Google DeepMind, Identifying AI-generated images with SynthID, 2023 technical introduction.
  7. ImageSignals, original video linked above. Demonstration limits are specified in the text.
  8. ImageSignals, Methodology and transparency: visual analysis, metadata and provenance are interpreted separately.
  9. NIST, Reducing Risks Posed by Synthetic Content: An Overview of Technical Approaches to Digital Content Transparency, 2024.