The first rule: there is rarely one decisive sign that says “this is AI”[1][2]
The question “is this image AI-generated?” sounds binary, but the file often is not. An image may be fully synthetic, partly generated, locally changed with inpainting, enhanced by generative retouching, or simply recompressed by a messaging app. Those scenarios leave different traces.
NIST treats synthetic-content detection as one family of techniques alongside provenance, watermarking and metadata. Its review also stresses that detector accuracy varies with compression, corruption, image style, depicted content and the relationship between the generator under test and the generators seen during training. A detector that performs well on one benchmark can degrade on a new model or a repeatedly exported image.
A credible analysis therefore does not ask a single service for a verdict. It assembles independent signals and asks whether they form a coherent technical story.
A detector score is a signal. Several independent signals are what turn a score into useful evidence.
1. Start with the original file, not a screenshot
Whenever possible, analyse the file obtained directly from the author, camera or source system. A screenshot destroys part of the technical context, changes compression and dimensions, and usually strips metadata. Many messaging and social platforms do something similar by creating their own derivative copy.
The original file gives you access to format details, colour profiles, EXIF/XMP, JPEG or PNG structure, encoding history and possibly provenance information. Even an absence of metadata is more meaningful in an original than in a file that has travelled through five platforms.
For disputed material, calculate a SHA-256 hash as well. It says nothing about authenticity by itself, but it proves that the file examined later is exactly the file that was preserved as evidence.
2. Check provenance and Content Credentials (C2PA)[1][3]
C2PA defines a standard for cryptographically bound information about the origin and history of digital media. Content Credentials can record how a file was created, what actions were performed and whether the source was, for example, a generative algorithm. That is qualitatively different from guessing based on pixels: a valid manifest describes recorded provenance.
Do not reverse that logic. Missing C2PA does not mean an image is genuine, just as missing EXIF does not mean it is fake. Adoption is voluntary, not every tool writes credentials, and some platforms transform or remove metadata during publishing.
When a manifest is present, validates correctly and declares a digital source such as trained algorithmic media, it can become a particularly strong piece of evidence about the provenance of that specific file.
3. Metadata is useful, but easy to overvalue[1]
EXIF can contain camera model, exposure, focal length, capture time, software and other fields. An unusual software tag, generator identifier or editing trace can be highly informative. So can inconsistency: a file may claim a certain camera while its structure or export characteristics do not resemble normal output from that device.
The limitation is obvious but important: metadata can be removed, rewritten or copied. It is one layer of the investigation, not the end of it. Plausible-looking EXIF is not a certificate of authenticity, and empty EXIF is routine after images pass through common platforms.
4. Pixel and compression forensics: look for inconsistencies, not ugly patches[1]
Image forensics can examine noise distribution, local compression differences, JPEG 8×8 grids, entropy, repeated textures and Error Level Analysis. These techniques can highlight regions that behave differently from the rest of the file.
They are not direct AI tests. A strong anomaly can come from legitimate retouching, repeated saving, sharpening or a human-made composite. Conversely, a fully generated image saved uniformly may look internally consistent at the compression level. Local forensics is strongest when combined with provenance, metadata and learned detectors.
Its practical value is often localisation. Instead of receiving only a number, an analyst can ask where the anomalies are and whether they overlap with the part of the image that matters to the case — for example, an alleged product defect.
5. AI detectors: use several and look for agreement[1][2]
Modern detectors learn statistical features from large collections of synthetic and natural images. Different services use different training data, architectures and score calibration. That is why the same image can receive materially different scores from different tools.
NIST runs dedicated evaluation programmes for image generators and discriminators precisely because detection quality must be measured against controlled benchmarks and new generators. The cross-generator case is especially difficult in practice: a detector has to recognise output from a model it may never have seen during training.
Agreement across independent layers is usually more informative than one extreme score. If four methods indicate low risk and one returns 92/100, the correct response is to investigate the disagreement — not automatically declare the image fake.
6. Visual inspection still helps, but only as triage[4]
The FBI lists signs such as distorted hands, teeth or eyes, irregular jewellery, inaccurate shadows, lag, voice mismatch and unrealistic movement as cues that can justify closer inspection. These remain useful for fast screening.
The problem is treating the absence of those errors as proof of authenticity. Generators have improved anatomy, typography and geometry, while genuine photographs processed with aggressive denoising, HDR or compression can look more synthetic than an AI image.
Visual review should therefore answer “what deserves closer inspection?” rather than “what is the final verdict?”.
A practical workflow for checking an image
When the result affects a refund, business decision, moderation action or dispute, a repeatable procedure is safer than an employee's visual intuition.
- Preserve the most original file available and do not resave it before analysis.
- Calculate SHA-256 and record basic technical facts: format, dimensions and size.
- Check C2PA / Content Credentials, EXIF, XMP and software traces.
- Inspect file structure and local pixel or compression anomalies.
- Run more than one independent AI/deepfake detector.
- Compare results with context: source, date, related images and the subject of the dispute.
- Express the conclusion as a risk level and a set of findings, not an infallible claim about the author's intent.
Good forensics does not try to prove AI at any cost. It separates strong evidence from what the file simply cannot establish.
What does 20/100, 60/100 or 90/100 actually mean?
A risk score is useful only when its composition is visible. A 20/100 score supported by six consistent layers is not the same as 20/100 from a single classifier. Likewise, 90/100 driven by direct provenance or a generator signature has a different evidential weight from 90/100 produced by one opaque detector with no supporting traces.
A useful report therefore shows the aggregate score and the source results, layer availability, metadata, technical limitations and strongest findings. The reader should be able to understand why the system reached that level of risk.
In formal disputes, technical analysis should be treated carefully. It can support a decision or identify material for further examination, but it does not by itself establish authorship, intent or legal responsibility.