The question people usually ask is simple: is this image real, or was it made by AI? The answer is usually not simple at all. A modern detector is not one switch. It is a stack of tests that look at different kinds of evidence and then combine them into a probability.
That matters because most public conversations about detection are still too binary. They assume there must be one hidden signature inside every synthetic image, or one model that can see what humans cannot. In practice, useful systems behave more like forensic checklists with math attached. They inspect pixel patterns, editing traces, metadata, provenance records, and learned model cues, then weigh how consistent those signals are.
The thesis for this article is straightforward: AI image detection works best as layered evidence, not as a single oracle. Once you understand that, detector outputs start to make more sense, including the cases where two tools disagree.
The first layer is still the image itself
The most familiar layer is the one people mean when they say “the detector looked at the image.” That covers a wide range of signals.
Some detectors look for visible or near-visible artifacts: anatomy glitches, inconsistent texture, odd edges, impossible reflections, or lighting that does not quite add up. The 2024 guide by Negar Kamali and coauthors organizes these cues into categories such as anatomical, stylistic, functional, physical, and sociocultural implausibilities. That paper is useful because it does not pretend every AI image fails the same way. Instead, it shows the kinds of mistakes that often surface when you inspect an image closely.
Kamali et al., 2024
Other detectors look at less intuitive image statistics. Research and product systems often inspect high-frequency residuals, compression behavior, or frequency-domain patterns rather than obvious “bad hands” artifacts. That approach exists because state-of-the-art generators have become much better at fixing the mistakes humans used to notice immediately.
In other words, a detector may be scoring the image at multiple levels at once:
- visible anomalies a human might also notice
- subtle distributional traces in textures and edges
- residual or frequency patterns that differ from ordinary camera capture
- model-based classification scores from a learned detector
If you only imagine “AI detection” as spot-the-sixth-finger, you miss most of the stack.
The second layer is metadata
Images are not just pixels. They also often carry metadata about capture, editing, rights, and workflow. Standards bodies such as IPTC describe how metadata fields can travel with an image, and that information can be meaningful context.
IPTC photo metadata overview
For a real camera photo, you may see EXIF-style fields that point to a device model, lens behavior, timestamps, or editing software. For a generated image, you may instead find software identifiers, workflow traces, or no useful camera data at all.
That said, metadata is not a magic answer either. It is easy to lose. Social platforms strip it. Screenshots remove it. Export steps can rewrite it. A missing EXIF block is not a verdict. A suspicious software tag is stronger than a blank field, but even then you still need context.
This is one reason serious detectors treat metadata as one input among many rather than the entire case.
The third layer is provenance
The most important conceptual shift in authenticity work over the last few years is the move from pure detection toward provenance. The C2PA explainer defines provenance as facts about the history of a digital asset, and says that a Content Credential is a cryptographically bound record of that history.
C2PA explainer
This changes the question. Instead of asking only, “does this image look fake?”, you can ask:
- where did this asset come from?
- what tool created or modified it?
- was the record tampered with?
- does the file still match the signed record attached to it?
That is a different kind of evidence than a detector score. A classifier estimates likelihood from patterns. A provenance record, if valid and intact, can show a documented history for the file.
But provenance is not a truth machine either. The C2PA explainer is explicit about that. It says Content Credentials do not tell you whether content is true, accurate, or factual. They tell you whether provenance information is well-formed, associated with the asset, and free from tampering. That distinction is easy to blur in marketing, but it is central in practice.
The fourth layer is learned model scoring
Many systems also include a learned classifier trained on large sets of human and generated images. This is the part most people imagine as “the AI detector model.”
These models can be useful because they compress a lot of subtle evidence into one score. They may pick up combinations of features that are hard to describe manually. They are also one reason detectors can still work on images that look perfectly plausible at a glance.
The catch is that learned detectors inherit the usual machine learning problem: they generalize unevenly. A system that does well on one family of generators or one benchmark may degrade on another, especially after cropping, recompression, screenshots, or model evolution.
NIST’s GenAI Image Challenge reflects this reality in how it frames the task. The Image-D discriminator task is not a one-time toy benchmark; it is an evaluation setup where systems must score whether an image was generated by AI or by a human, with metrics like AUC, EER, TPR at a chosen FPR, and Brier score. That is a strong signal that the field treats detection as an ongoing measurement problem, not a solved checkbox.
NIST GenAI Image Challenge
Why fusion matters
If each layer is imperfect, why do products combine them? Because the weaknesses are different.
- Pixels can be ambiguous.
- Metadata can be missing.
- Provenance may never have been attached.
- A learned classifier can overfit or drift.
Layering helps because these signals fail in different ways. A classifier may be unsure, but provenance may be present. Metadata may be blank, but the image may still contain residual patterns or implausible structure. A file may look ordinary, but a valid Content Credential may tell you it came from an AI workflow anyway.
This is also why detector outputs are often best expressed as confidence bands, explanations, or module-level signals rather than just “REAL” or “FAKE.” A careful system is effectively saying: here is the mix of evidence we saw, and here is how strong that mix looks under current models.
Humans are part of the loop, whether products admit it or not
One useful check on detector mythology comes from human studies. Kamali and colleagues ran a large experiment with 50,444 participants and found that scene complexity, artifact type, display time, and curation all affect how well people distinguish real from AI-generated images.
Kamali et al., 2025
That study cuts two ways.
First, humans are not great at this task in a stable way. Second, the cues are real enough that structured inspection still matters. So the most honest workflow is not “trust your eyes” or “trust the model.” It is usually “compare your eyes, provenance signals, and detector output, then see whether the evidence lines up.”
What this can and cannot tell you
What it can tell you
- why detectors are usually multi-signal systems
- why provenance and metadata belong in the conversation
- why classifier scores are probabilistic rather than absolute
- why disagreement between tools is not automatically a bug
What it cannot tell you
- that one detector output proves truth
- that an image without metadata must be AI-generated
- that provenance alone settles whether depicted events really happened
- that today’s best detector will stay best as generators change
The practical takeaway
If you remember one thing, make it this: detection is strongest when it behaves like evidence synthesis. It is not one hidden fingerprint, and it is not one giant model whispering the answer from beyond the pixels. It is a layered judgment call backed by different kinds of signals.
That view is less glamorous than the “100% accurate detector” headline, but it is much closer to how the field actually works.
If you want to test that layered approach yourself, try a scan on the Detectiks home page and compare the explanation with what you can independently verify in the file.
Last reviewed
May 11, 2026.
Sources
- C2PA and Content Credentials Explainer
- NIST Open Media Forensics Challenge briefing
- NIST GenAI Image Challenge
- Kamali et al., 2025: Characterizing Photorealism and Artifacts in Diffusion Model-Generated Images
- Kamali et al., 2024: How to Distinguish AI-Generated Images from Authentic Photographs
- IPTC photo metadata overview
