— Learn
AI image detection glossary: forensics terms, plainly defined
Thirty-five terms from AI-image detection and digital forensics, defined in plain language: false positive rate, calibration, base rate, chain of custody, provenance, inpainting and the rest.
Published August 6, 2026
Detection and forensics both carry vocabulary that sounds precise and is used loosely, which is how a marketing claim and a measurement end up wearing the same words. These are the definitions this site uses. Where a term has a longer story, the guide that tells it is linked underneath.
How detection is measured
The vocabulary behind every accuracy claim, including ours. Most disagreements about whether a detector is any good turn out to be disagreements about these words.
- AI image detection
- Estimating whether an image came from a generative model rather than a camera, by reading statistical traces in the pixels that no eye can see. The output is a probability, not a proof. It can be wrong in both directions.See: How to tell if an image is AI-generated
- Accuracy
- Loosely, the share of images a detector judges correctly. As a single advertised number it means very little, because it blends two different errors and shifts with both the mix of the test set and where someone put the threshold. Ask which errors it counts, and on which images they were measured.See: How accurate are AI image detectors?
- False positive
- A real photograph the detector calls AI-generated. This is the error that damages people, because it accuses a genuine image and, by extension, whoever supplied it. Most vendors decline to quantify it.See: Why real photos get flagged
- False negative
- An AI-generated image the detector fails to flag. Every detector has them, and a calibrated one deliberately accepts more of them in exchange for fewer false alarms. So a clean result is never a certificate of authenticity.
- Detection rate (true positive rate)
- The share of genuinely AI-generated images a detector flags at a given threshold. Quote it without the false-positive rate from that same threshold and it means nothing, because either number can be improved by sacrificing the other.
- False positive rate (FPR)
- The share of real photographs wrongly flagged as AI at a given threshold. Of all the numbers published about a detector, this is the one worth having, and the one most often left out. Without it a detection rate is unusable.See: Why real photos get flagged
- FPR cap
- A promise made before the verdict: the threshold is picked so the false-positive rate stays at or under a stated ceiling. Vestige publishes two, a standard setting capped at 5% and a strict setting capped at 1%. The cap holds for every real image source in the benchmark, so the noisiest source is the one that sets the threshold.
- Threshold (operating point)
- The score above which an image is called AI. Nothing in the model fixes it; someone chooses it. Push it up and you get fewer false alarms and more misses, push it down and you get the reverse. A detector that will not tell you its threshold has not told you what its verdict means.
- Calibration
- Setting the threshold from measured data so a verdict carries a known error rate, instead of an arbitrary default like 0.5. It has to be done separately for each image condition, since compression and resizing move the whole score distribution. Calibration is what separates a number from a decision you can defend.See: Does compression affect detection?
- AUC
- Area under the ROC curve, summarizing how well a detector ranks AI images above real ones across every possible threshold. Because it averages over all of them it flatters detectors, and a ranking is not something you can act on. High AUC sits happily alongside a modest detection rate once a real threshold is applied.See: How accurate are AI image detectors?
- Base rate
- How common AI images actually are in the pile you are checking. When they are rare, even a small false-positive rate produces more false alarms than true catches, so a flag on an ordinary photo is more likely to be an error than a discovery. Forgetting this is the commonest way detection results get misread.See: Why real photos get flagged
Reading a detection report
What the fields on a Vestige report mean. Other tools use some of these words differently, so check their definitions rather than assuming ours.
- Verdict
- The report's plain-language conclusion at a stated operating point, rather than a bare number. Because a verdict only means anything in combination with the threshold and cap that produced it, the report prints both alongside.See: See a full sample report
- Confidence score
- The detector's raw output, from 0 to 1, where higher means more consistent with AI generation. Alone it says little: the same value can be unremarkable under one image condition and decisive under another. A calibrated threshold is what turns it into information.
- AI-flagged / Not flagged
- The two verdicts a report can carry. AI-flagged always appears with the cap that produced it, reading "AI-flagged, at most 5% false positives" or the strict 1% version, and means the score crossed a calibrated threshold rather than that generation has been established. Not flagged means nothing was found. That is not the same as the image being confirmed real.See: See a full sample report
- Degradation condition
- The processing history the analyzer estimates from the file: clean, recompressed, resized, screenshotted, sent through a messenger, or some combination. Each condition carries its own calibrated thresholds, because an identical score means different things depending on what the file has survived.See: Does compression affect detection?
- Benchmark
- A fixed set of images with known origins, scored under controlled conditions to measure how a detector performs. Whether it is worth anything depends on how closely it resembles the images you actually handle. Clean pictures from familiar generators produce numbers that do not survive contact with a forwarded screenshot.
- Test set
- The portion of a benchmark held back from training and used only for measurement. Let images or near-duplicates leak across that line and the reported numbers inflate into worthlessness. The separation has to be enforced and checked, never assumed.
- Pinned calibration run
- A named, versioned benchmark run that a report's thresholds come from, printed on the report itself. Models get retrained, and without a named run a number on last month's report can be neither reproduced nor challenged. Vestige reports carry the run and the pipeline version.See: See a full sample report
Forensic practice
Terms from established forensic work that carry over to disputed images. If an image may end up in a dispute, the handling matters as much as the analysis.
- Chain of custody
- The documented history of who held an item of evidence, when, and what they did with it. Gaps invite the argument that it changed along the way. For digital images the working equivalent is hashing the file on receipt and logging every transfer, since a re-saved copy is a different file.See: Can AI-generated images be used as evidence?
- Cryptographic hash (SHA-256)
- A fixed-length fingerprint computed from a file's exact bytes. Change anything, even something invisible, and the fingerprint changes completely, which is what makes it useful for showing a file was not altered between two points in time. It says nothing about whether the picture is truthful, only that the bytes are unchanged.
- Exhibit
- An item formally submitted as evidence in a proceeding, identified so everyone is arguing about the same thing. For images that means the specific file, named by hash, not a description of the picture or a copy pasted into a document.
- Provenance
- Where an image came from and what has happened to it since: who published it first, when, and through which platforms. Provenance often beats pixel analysis outright, because finding the original in a stock library or a 2019 news story ends the argument.See: How to tell if an image is AI-generated
- Metadata (EXIF)
- Fields stored alongside the image data recording camera, lens, timestamp and sometimes location. Useful when present, and trivially editable, so it corroborates rather than proves. Most platforms strip it on upload, which makes its absence close to meaningless.
- C2PA / Content Credentials
- An open standard for cryptographically signed provenance attached at capture or edit, letting an image carry a verifiable record of its own origin. Where one exists and validates, it beats any detector output. Adoption is still thin and the credential is easy to strip, so absence proves nothing.
- Reference sample (known sample)
- In forensic comparison work, material of established origin used as the baseline a disputed item is measured against. Image detection has no true equivalent. The comparison runs against statistical patterns learned from a benchmark population, not against a known sample from the same source, and that gap limits how far a detection result can honestly be pushed.See: Can AI-generated images be used as evidence?
- Questioned sample (disputed item)
- The item of unknown origin under examination, the thing the dispute is actually about. In image work that is one specific file. Preserve and hash it in the form it arrived, before anyone crops it, re-saves it, or forwards it through an app.
Generative AI and image degradation
How the images get made, and what happens to them on the way to you. The second half matters more than most people expect.
- Generative AI
- Models that produce new content rather than sort existing content. For images that covers text-to-image systems, editing tools that invent regions, and enhancement features that reconstruct detail. The breadth matters, because partial generation is now far more common than wholly synthetic pictures.
- Diffusion model
- The architecture behind most current image generators. It begins with noise and removes it step by step until an image matching the prompt emerges. Detection often keys on statistical residue that process leaves behind, which also explains why detectors weaken against generators they have never seen.
- GAN
- Generative adversarial network, an older architecture that trains a generator against a discriminator, and the thing that dominated synthetic faces before diffusion arrived. The distinction still matters, because a detector trained mostly on GAN output can do badly on diffusion output and the reverse.
- Deepfake
- Loosely, media manipulated with machine learning to show someone doing or saying something they did not. The word began as a term for face swaps in video and now stretches to almost any synthetic media. That looseness causes real trouble: a fully generated image and a swapped face are different technical problems with different detection profiles.
- Inpainting
- Generating new content inside part of an existing image, to remove a person or add an object. It is the hardest common case for detection: most of the file is a genuine photograph, only a fraction carries generative traces, and a whole-image score can come out looking unremarkable.
- AI upscaling
- Enlarging an image with a model that invents plausible detail instead of interpolating it. Phones, editing apps and platforms apply it routinely. Because it leaves generation-like statistics in an otherwise real photograph, it is one of the honest routes to a false positive.
- Re-encoding (recompression)
- Saving an image again in a lossy format, which happens on every upload, forward and screenshot. Each pass throws away fine detail, and fine detail is precisely what detectors read. Hence the gap between performance measured on clean files and performance you will actually see.See: Does compression affect detection?
- Computational photography
- The processing modern phones apply to every shot: stacked exposures, aggressive denoising, sharpening, night modes. What comes out is a real photograph that has been heavily reconstructed, and its statistics can resemble generated images closely enough to lift a detector's score. It is a known cause of false positives on genuine phone photos.
A term missing, or a definition you would argue with? Tell us at the contact page. This page gets corrected the same way the guides do.