Glossary
The vocabulary of AI-creative QA
Plain definitions for the failure modes, standards and scores this field runs on — each linked to the check or research that goes deeper.
- AI slop
- Catch-all for the visual carelessness that marks low-effort generated imagery: melted edges, smeared textures, ghost objects, nonsense geometry. Individually minor, collectively the reason audiences distrust an image without being able to say why. The artifacts check hunts it region by region.
- Garbled text
- Rendered text that almost spells the word — “SPEICAL OFFRE”, invented letterforms, melted digits. It is an architectural failure, not a maturity phase: models never see characters and generate at a resolution where small glyphs cannot survive. Why AI still can’t spell covers the research; the text check catches it.
- Anatomy error
- Extra fingers, thumbs on both sides, joints bending against themselves, merged teeth — the most screenshotted class of AI failure. Hands fail most because they are small, articulated and photographed in thousands of poses. The anatomy check inspects every rendered person limb by limb.
- Melted edges
- The wax-like boundary where a generated object loses its outline — jewellery dissolving into skin, a logo whose corners round off into the background. A signature diffusion artifact, flagged by the artifacts check.
- Ghost object
- A half-formed thing with no purpose: almost-a-chair in the background, a handle attached to nothing, a limb belonging to no one. Models fill space with plausible texture, and plausible is not the same as real. Caught by the artifacts and object-physics checks.
- Texture repetition
- The same patch of fabric, foliage or crowd repeating across a region — a tiling artifact from how models synthesize large uniform areas. Reads as fake at a glance in print sizes. The texture-coherence check measures it.
- Impossible physics
- Structures and interactions the physical world would refuse: stairs to nowhere, shadows pointing at the light source, reflections showing a different scene, liquid pouring uphill. The object-physics and lighting checks flag them with pinned regions.
- Training-data regurgitation
- When a model outputs a near-copy of an image it was trained on. Measured, not hypothetical: research found ~1.88% of sampled Stable Diffusion generations closely matched a training image, with heavily duplicated images most at risk. The Filippa K case shows the commercial blast radius; the IP-similarity check is the pre-publication defence.
- IP proximity
- How close a generated image sits to existing protected work — measured by reverse-image search and similarity scoring rather than argued about after publication. Proximity is evidence, not a verdict, but it is the evidence a rights holder starts from. See the IP-similarity check.
- Content Credentials (C2PA)
- Cryptographically signed provenance metadata attached at creation — who made the image, with what tool, edited how. The open standard is C2PA. In practice credentials frequently do not survive the publishing pipeline; we measured where they get stripped. The forensics check reads what the file still carries.
- AI disclosure
- Labeling requirements for synthetic content — most concretely the EU AI Act’s transparency obligations applying from August 2026, plus platform-level rules that already exist. Our guide to Article 50 covers who must disclose what, when.
- Digital twin (model)
- An AI replica of a real person — in fashion, a model’s licensed likeness generated into new imagery. The consent, compensation and contract questions are live; likeness disputes without consent are already in court. Anatomy and provenance checks apply to twins like any rendered person.
- Inpainting fix
- Regenerating only the defective region — the hand, the garbled word — and compositing it back over the original so the rest of the image stays untouched. Chekr’s one-click fixes work this way, with a before/after judgement that rejects fixes that drift.
- Integrity score
- A 0–100 roll-up of everything a scan found, weighted by severity and confidence, that gates whether a creative passes, needs work or is blocked. It turns “looks fine to me” into a number a pipeline can act on. Every scan returns one.