Guide · 9 min read · Updated September 2026

Signal one: generation artefacts, and why models betray themselves in texture

Every check VTYAI runs falls into one of four families of evidence: generation artefacts, physical consistency, editing signatures, and file metadata. This is the first of four guides that take each one apart — what it measures, what a strong result looks like, and where it fails. Generation artefacts are the family most people mean when they say a picture 'looks AI', but the useful version of that instinct is much more specific than a vibe.

By Ira Peoples

What an artefact actually is

An image generator does not draw objects. It starts from noise and repeatedly nudges it towards something that satisfies a statistical model of what images look like. Nothing in that process holds a concept of a hand with five fingers, a sentence with words in it, or a chair that could bear weight. It holds a very good sense of what regions of a photograph tend to look like next to one another.

An artefact is what is left over when that process produces something locally plausible and globally wrong. The model is not making mistakes in the human sense — every patch is exactly what its training says a patch like that should be. The failure appears when you ask whether all those patches could belong to one real scene.

Texture: the most reliable artefact family

Real surfaces have detail at every scale. Skin has pores inside blotches inside tonal shifts across a face. Fabric has fibre inside weave inside drape. Foliage has veins inside leaves inside branches. Photograph any of it and the fine scale keeps going until the sensor or the lens gives up.

Generated texture tends to be correct at one scale and hollow beneath it. Zoom into generated skin and the pores are often a uniform stipple rather than an uneven distribution; zoom into generated brickwork and the mortar lines are frequently too regular, with a repeating unit that a real wall would not have. Magnify a small patch to the point where you cannot tell what you are looking at: real detail stays chaotic, synthetic detail resolves into a pattern.

Boundaries: where two things meet

Generators are strongest in the middle of regions and weakest at edges. Hair against a background is the classic case: real hair produces thousands of individual strands that break the outline and pick up light from behind, while generated hair often merges into a soft mass with a suspiciously clean silhouette, or sprouts strands that fade into nothing rather than terminating.

The same principle applies to a collar against a neck, a glass against what is behind it, and fur against sky. Look at boundaries at high magnification and ask whether the two materials are actually interacting or merely adjacent.

Structure with rules: text, hands, teeth, mechanisms

Some things in the world are governed by rules a statistical model has no access to. Writing must spell words. A hand has a fixed topology of five digits with joints that bend one way. Teeth come in a symmetrical arrangement. A bicycle chain has to form a closed loop around two sprockets.

These are where artefacts are most visible to an untrained eye, and they are genuinely diagnostic — but they are also the failures model-makers have worked hardest to fix, and the ones a human can retouch in a minute. Text in particular has improved enormously: a headline may now be spelled correctly while a caption two inches away is still letter-shaped mush. Check the smallest text in the frame, not the largest.

Frequency: what the eye cannot see

Beyond what is visible, an image can be examined statistically. Camera sensors leave a characteristic noise fingerprint across the frame, and lens and demosaicing processes leave regular relationships between neighbouring pixels. Generated images are assembled differently, and in the high-frequency detail — the finest variation, invisible at normal viewing size — the distribution often does not match what any camera produces.

This is the part of artefact analysis that a person cannot do by eye, and it is one of the reasons an automated check can disagree with your intuition in both directions. It is also the part most easily destroyed, which brings us to the limits.

Why artefacts disappear

Compression is the great eraser. Every time an image passes through a messaging app or a social platform it is re-encoded, and re-encoding works by discarding exactly the fine high-frequency detail that artefact analysis reads. A generated image that would be caught instantly at full resolution can be genuinely unreadable after three shares.

Downscaling does the same thing more bluntly. So does printing and rephotographing, screenshotting, and applying a filter. When you have a choice, always check the largest, least-travelled copy you can obtain — ask the sender for the original rather than forwarding what you were shown.

Why real photos trip artefact checks

The confounders are everywhere in ordinary photography. Phone cameras apply aggressive noise reduction and skin smoothing by default, which flattens texture in precisely the way a generator does. Portrait mode synthesises background blur algorithmically, producing the too-clean boundaries described above. Beauty filters, upscaling software, and AI-assisted 'enhance' features all leave synthetic detail in an authentic photograph.

This is why a single artefact should never decide a verdict, and why VTYAI reports which specific checks fired rather than a bare percentage. One soft boundary in a portrait taken on a modern phone means very little. Flat texture, plus a clean boundary, plus letterform noise in the signage, plus a frequency profile unlike any camera, is a different matter — corroboration across independent artefact types is what makes the family worth trusting.

The short version

Artefacts arise because generators build locally plausible patches, not globally coherent scenes.
Texture that resolves into a pattern under magnification is the most reliable visible tell.
Boundaries between materials, and rule-governed structures like text and hands, fail first.
Compression, downscaling and filters erase artefacts — always check the least-travelled copy.
Phone processing and portrait mode mimic artefacts, so corroboration across types matters more than any single sign.