How to Spot a Generated Image: The Signs and Their Reliability
Visual tells, technical checks, and an honest assessment of how far any of them can be relied on today.
Judging an image's origin has become a practical skill. No method is fully reliable, but a combination of signals distinguishes obvious cases and supports reasonable doubt in ambiguous ones.
Visual tells
The classic weak spots of generation, still present though rarer than before.
- Hands and fingers: count, joint directions, length relative to the palm.
- Text: lettering on clothing, signage, documents — pseudo-letters or distorted words.
- Paired objects: earrings, buttons, interior elements — different where they should match.
- Fine repeating structures: grilles, tiling, patterns, teeth.
- Reflections: in mirrors, windows, glasses — not matching the scene.
- Object boundaries: hair meeting a background in a physically impossible way.
Light-related tells
Often more reliable than anatomical ones, because they are harder to fix.
- Shadows running in different directions from different objects.
- No shadow where one is obligatory: under feet, under an object on a surface.
- Catchlights in the eyes that do not match the visible light source.
- Even illumination with no source — the commonest tell.
Realism-related tells
The least reliable but the most noticeable at first glance: flawless skin with no micro-texture, an implausibly clean scene, no incidental detail, composition without motivation, saturated pleasing colours.
These are characteristic of default settings and are easily removed deliberately, so their absence proves nothing.
Technical methods
- File metadata: sometimes contains provenance information, but it is easily lost and easily faked.
- Provenance signatures: a developing industry mechanism, reliable where applied.
- Automatic detectors: probabilistic, erring in both directions, with quality changing rapidly.
- Reverse image search: helps find an original or earlier publications.
Contextual checking
Often more effective than visual analysis. Questions worth asking:
- Are there other images of the same event from different angles?
- Who published first, and when?
- Does the image agree with known facts: weather, time of day, location?
- Is there independent corroboration of what is shown?
What this means in practice
The most realistic position today: visual tells help filter out weak fakes but do not guarantee detection of good ones.
Hence a practical shift: instead of asking "is this generated", it is more reliable to ask "where did this image come from and who is answerable for it". Provenance can be checked; appearance cannot.
How this affects your own publishing
The flip side: if the tells are unreliable, your genuine photograph can also be doubted. That already happens and will happen more often.
The practical answer is keeping sources and metadata, publishing from verifiable channels, and labelling generated material. Transparency works better than trying to prove authenticity after the fact.
Frequently asked
Which tells are most reliable?
Light-related ones: shadows running different ways, missing shadows where they are obligatory, catchlights not matching the source. They are harder to fix than anatomical defects.
Can automatic detectors be trusted?
They are probabilistic, err in both directions, and their quality changes rapidly. Useful as one signal, not as a final answer.
What if authenticity really matters?
Check provenance rather than appearance: who published first, whether other angles exist, whether the image agrees with known facts.
- #распознавание
- #проверка
- #практика