AI Image Forensics Glossary: Essential Terms

How to use this glossary

The terms below are grouped by the role they play in an investigation: generation, metadata, provenance, pixel-level forensics, detection, evaluation, and practical verification. Each definition is written for readers who need technically accurate language without assuming a machine-learning background.

A useful investigation moves between these groups. For example, a detector score may suggest synthetic origin, metadata may describe an export tool, and provenance may show whether a signed creation record is present. None of those signals should be interpreted in isolation.

AI generation and editing terms

Generative AI: software that creates new content from learned patterns rather than retrieving a fixed stored image.

Diffusion model: a generative model that learns to reverse a noise process and gradually construct an image. Many modern image generators use diffusion-based architectures.

Latent diffusion: diffusion performed in a compressed representation instead of directly on full-resolution pixels, reducing computation while preserving semantic structure.

GAN: a generative adversarial network in which a generator and discriminator are trained in competition. GAN fingerprints remain relevant to older synthetic-image research even though diffusion models now dominate many consumer tools.

Prompt: text or other instructions supplied to a generative system. A prompt describes intent, but it does not reliably reveal the exact pixels or edits that produced the final file.

Seed: a value used to initialize a generation process. Reusing a seed may help reproduce similar outputs when the model, settings, and software version are also controlled.

Inpainting: replacing or generating content inside a selected region of an existing image.

Outpainting: extending an image beyond its original boundaries with generated content.

AI upscaling: increasing resolution with a model that may invent texture and detail rather than only interpolating existing pixels.

Synthetic media: image, audio, video, or text content created or substantially altered by computational generation systems.

Metadata and file-structure terms

EXIF: a common metadata format that can store camera model, capture settings, timestamps, orientation, and sometimes location. EXIF can be removed, rewritten, or absent for legitimate reasons.

IPTC: metadata commonly used by newsrooms and asset-management systems for captions, rights, creators, and editorial information.

XMP: an extensible metadata framework used by many creative applications to store editing, rights, and workflow information.

Metadata stripping: removal of metadata during export, upload, messaging, screenshotting, or privacy processing.

Software tag: a metadata field that may identify an application used to save or export a file. It can provide context but is not proof of how every pixel was created.

MIME type: a declared file type such as image/jpeg or image/png. Reliable systems verify the underlying bytes instead of trusting only the filename or client declaration.

Container structure: the internal organization of an image file, including segments, chunks, metadata blocks, and encoded image data.

Provenance and authenticity terms

Provenance: information about where an asset came from and what happened to it over time.

C2PA: an open technical standard for attaching cryptographically verifiable provenance information to media assets.

Content Credentials: a user-facing implementation of provenance information that can describe creation tools, edits, and signing entities when credentials are present and valid.

Manifest: a structured provenance record associated with an asset. A manifest may contain claims, assertions, signatures, and relationships to earlier versions.

Claim: a signed statement within a provenance system that binds assertions to an asset.

Assertion: a specific piece of provenance information, such as an action, ingredient, or metadata statement.

Digital signature: cryptographic evidence used to verify that signed data has not changed and that it was signed by a particular key. Trust in the signer still requires context.

Trust chain: the sequence of certificates or authorities used to evaluate whether a signing identity should be trusted.

Ingredient: a source asset referenced as an input to a later asset in a provenance record.

Watermark: a visible or invisible signal embedded in media. Watermarks can support attribution or detection, but their robustness and coverage vary by implementation.

Pixel and compression forensics terms

JPEG artifact: a block, ringing pattern, or loss of detail introduced by lossy JPEG compression. These artifacts are common in ordinary photos and are not automatically evidence of manipulation.

Double compression: evidence that an image may have been JPEG-encoded more than once. Recompression is common after editing and social-media processing.

Error level analysis: a visualization of JPEG recompression differences. It can help generate questions, but it is not a reliable standalone detector of editing or AI generation.

Noise residual: the high-frequency component left after estimating and removing image content. Researchers analyze residuals for camera, processing, and generation patterns.

Sensor pattern noise: weak camera-sensor variation sometimes used for source-camera analysis. Resizing, denoising, recompression, and screenshots can greatly weaken it.

Frequency analysis: examination of image information in frequency space to identify periodic, spectral, or processing-related patterns.

Resampling trace: evidence associated with resizing, rotation, warping, or other geometric interpolation.

Color filter array artifact: traces associated with how a camera reconstructs color from a sensor mosaic. Generated or heavily processed images may not preserve normal camera-pipeline traces.

Screenshot recapture: creation of a new image by capturing a displayed version of another image. This changes metadata, dimensions, compression, and many low-level signals.

AI detection and scoring terms

Classifier: a model that assigns an input to one or more categories, such as camera-origin, synthetic, or inconclusive.

Feature: a measurable pattern used by a model or forensic method. Features may be visual, statistical, semantic, metadata-based, or provenance-based.

Model fingerprint: a recurring statistical pattern associated with a generator or processing pipeline. Fingerprints may weaken when models or post-processing workflows change.

Ensemble: a system that combines multiple models or evidence sources instead of relying on a single detector.

Confidence score: a numeric expression of model certainty under its calibration assumptions. It is not the probability that a legal or historical claim is true.

Authenticity score: a product-specific summary of evidence related to apparent real-world or camera origin. Its meaning depends on the methodology used to compute it.

Threshold: the score boundary used to convert a continuous output into a category or action.

Abstention: a deliberate inconclusive result when available evidence is insufficient or conflicting.

Calibration: the degree to which score levels correspond to observed outcome frequencies on representative test data.

Distribution shift: a difference between real-world inputs and the data used to train or validate a detector.

Evaluation and benchmark terms

True positive: an AI-generated image correctly identified as AI-generated under the benchmark label definition.

False positive: a real or camera-origin image incorrectly classified as AI-generated.

False negative: an AI-generated image incorrectly classified as real or not detected.

Precision: among items predicted positive, the proportion that are actually positive in the labeled test set.

Recall: among all positive items in the labeled test set, the proportion the system identifies.

F1 score: the harmonic mean of precision and recall. It summarizes two error dimensions but can hide class-specific and threshold-specific behavior.

ROC curve: a plot of true-positive rate against false-positive rate across thresholds.

Area under the curve: a threshold-independent ranking summary. It does not by itself describe performance at the operating point used in production.

Holdout set: data kept separate from training and model selection for final evaluation.

Unseen generator: a model family or version not represented in training data. Performance on unseen generators is important for measuring generalization.

Practical verification terms

Reverse image search: searching for visually similar or earlier versions of an image to establish context, reuse, or provenance clues.

Source verification: confirming where an image was first published, who supplied it, and whether the source can provide an original file or supporting context.

Chain of custody: documentation of how evidence was acquired, stored, transferred, and analyzed.

Corroboration: support from independent evidence such as additional photos, eyewitness records, trusted publications, provenance data, or source files.

Mixed-origin image: an asset containing both camera-captured and generated or heavily edited regions.

Human review: interpretation by a person who considers technical output together with context, stakes, and alternative explanations.

• Use multiple independent signals.
• Preserve the original file whenever possible.
• Treat absence of metadata or credentials as unknown, not proof of fakery.
• Escalate high-stakes decisions to qualified human review.

A responsible interpretation workflow

Begin with the question you are trying to answer: whether the file is likely synthetic, whether it was edited, whether its source is trustworthy, or whether a specific provenance claim can be verified. These are related but different questions.

Next, preserve the best available file, review metadata and provenance, inspect visual and compression evidence, compare detector outputs, and seek contextual corroboration. Finish by recording uncertainty and plausible alternative explanations rather than converting every signal into a definitive label.

Is image forensics the same as AI image detection?

No. AI image detection is one part of image forensics. Forensics also examines editing, compression, metadata, source-camera traces, provenance, and contextual authenticity.

Can EXIF metadata prove a photo is real?

No. EXIF can support an investigation, but it can be removed, rewritten, copied, or generated by software.

Does missing metadata mean an image is AI-generated?

No. Websites, messaging apps, screenshots, and privacy tools commonly remove metadata from genuine photographs.

What does a C2PA manifest prove?

A valid manifest can verify signed provenance statements and whether they remain cryptographically bound to the asset. The meaning still depends on the signer, assertions, and trust context.

Does the absence of Content Credentials mean an image is fake?

No. Most images do not yet carry Content Credentials, and credentials can be removed by unsupported workflows.

Can JPEG artifacts reveal AI generation?

Compression artifacts can contribute context, but they are common in both real and generated images. They should not be used as standalone proof.

What is the difference between confidence and accuracy?

Confidence describes the strength of a model output for one item. Accuracy summarizes performance across a labeled dataset. A confident result can still be wrong.

Why do AI detectors disagree?

They may use different training data, features, thresholds, generator coverage, and calibration methods. Post-processing can also affect each detector differently.

Can screenshots fool forensic tools?

Screenshots replace many original file signals and can reduce detector performance, but they do not guarantee that an image becomes undetectable.

What is a false positive in AI image detection?

It is a genuine or camera-origin image incorrectly labeled as AI-generated. False positives are especially important in high-stakes moderation, journalism, and identity workflows.

Should a detector ever return inconclusive?

Yes. Abstaining is often more responsible than forcing a binary answer when evidence is weak, conflicting, or outside the model's validated coverage.

Can image forensics provide legal proof?

A consumer analysis can support an investigation, but legal conclusions typically require validated methods, qualified experts, preserved evidence, and documented chain of custody.