Physics
Physical coherence
Does the image obey the laws of nature? Light, shadow, anatomy, proportions, materials. Such breaks are often small – but they reliably give an image away once you see them.
AI-generated mood image
How SemantIC works
SemantIC takes an AI-generated image and looks at it along three dimensions: whether it is physically plausible, whether it makes sense in substance, and which clichés it carries. On this page we show you what the tool does, where its criteria come from and – just as importantly – what it deliberately does not do.
01 · The idea behind it
AI image generators optimize for what reads as “good” at first glance: sharpness, colour, light, composition. They optimize far less for coherence in substance – the logic of the scene, proportions, clichés. Images thus become reliably beautiful, but not reliably correct.
The consequence is the real problem: a convincing image does not invite a closer look. The surface puts the eye at ease, and weaknesses slip through. We call this connection masking – the working hypothesis behind SemantIC: a convincing surface can make critical scrutiny harder. The study behind it supports this assumption; it cannot prove it.
Theoretical basis
This effect is not just an observation; it can be placed scientifically. Recent research on AI detection systems shows that artificial neural networks prioritize visual properties, whereas humans grasp images more through their meaning. That the same holds for image generators is the assumption of this thesis. SemantIC steps into exactly this gap: it looks where the eye is satisfied too quickly.
02 · The process
SemantIC separates content and aesthetics into two distinct assessment tracks – so that appeal does not gloss over the judgment on substance. Aesthetics are estimated twice: by a vision model and an independent reference model; if the two diverge strongly, the finding says so. Only afterwards are both results brought together.
Input
Your image
+ optional usage context & original prompt
Two separate checks · parallel
Check 1 · Content
Integrity analysis: 3 dimensions (physics · semantics · bias) plus research layer → integrity score.
Check 2 · Surface
Aesthetic assessment: the visual appeal only, reconciled with an independent reference model → aesthetic score.
Appeal and Substance are brought together only here – never before.
Merge
Appeal vs. substance
Reconciling aesthetics ↔ integrity → masking cues (no single metric).
Finding
Your result
Traffic light · 3 dimensions · reading mode · cues.
Model-agnostic – Gemini & Claude are the current standard models, interchangeable. The aesthetic score is the average of Claude and an independent reference model (LAION Aesthetic Predictor); if the two diverge by 20 points or more, the finding says so; if the reference fails, Claude counts alone. If you provide a prompt or context, an image-text comparison (CLIP) also runs – a pure diagnostic signal with no influence on scores and traffic lights. The symbols are stylized placeholders.
02 · b · The guardrails
An image model on its own tends to be led by visual perfection. SemantIC therefore applies guardrails to the assessment that push towards substantiated, traceable judgments.
01 Tendency
Gets won over by the beautiful surface.
01 Guardrail
Aesthetics are assessed separately – with no influence on the judgment on substance.
02 Tendency
Asserts vague impressions (“looks off”).
02 Guardrail
Every finding needs visible evidence on the image – otherwise it doesn't count.
03 Tendency
Answers freely and associatively.
03 Guardrail
Judges only within the fixed research vocabulary (reading modes, drivers, error types).
04 Tendency
Ticks off bias categories too hastily.
04 Guardrail
Gender and skin tone are only described; bias only on a real pattern (role, power dynamic).
05 Tendency
Sticks with the first assessment.
05 Guardrail
Deterministic rules re-check the result and correct contradictions.
That makes the assessment more disciplined and more traceable. Where the tool reaches its limits, SemantIC discloses that (section “We know the limits”).
“Masked → clear”: tendency dampened, guardrail sharp. Largely hard-enforced (fixed schema & check rules); the criteria come from 144 coded images.
03 · The three dimensions
Physics
Does the image obey the laws of nature? Light, shadow, anatomy, proportions, materials. Such breaks are often small – but they reliably give an image away once you see them.
Semantics
Does the scene make sense – and does it fit your context? A technically clean image can still be off in substance: a situation that would not happen that way, or that does not fit your piece.
Bias
Which roles, poses, clichés does the image show? A beautiful image of a cliché remains a cliché. Important: gender or skin tone are in themselves no finding – only a stereotyped casting of the role, power dynamic or a context break sets it off.
04 · The masking check
SemantIC assesses separately how good an image looks (its aesthetics) and how well it holds up in substance (its integrity). If appeal clearly exceeds substance, that is the breeding ground for masking – the perfect surface can cover flaws before you notice them. There is deliberately no single masking metric; the quadrant shows the principle, not the result of a check.
If an image sits far to the top left – lots of appeal, little substance – that is the breeding ground for masking. The further a point crosses the diagonal towards the top left, the more strongly aesthetics exceed integrity.
05 · Your input
Before SemantIC checks, we ask two things: your stance on the image and its intended use. Both only frame the recommendation – they shift neither the score nor the findings.
Stance
The same weakness weighs differently depending on whether an image shows a cliché deliberately or reproduces it unreflectively.
Intended use
From this, SemantIC derives how strict the standard should be.
Selection in the tool – shown here for overview only, not clickable.
06 · A cue, not proof
SemantIC gives no hard verdicts but reasoned cues: each finding shows where you should look more closely – not that an image is “wrong”. Whether it fits your context is not something any tool decides for you; the final judgment stays with you.
07 · Validated limits
Every limit here has been empirically tested – in documented test series against the human coding, and for anatomy additionally blinded with a second vision model. You should know when you can rely on a finding and when your own eye is needed.
Robust in testing
In the documented anatomy test run (16 images, 42 runs), no clean image triggered a red finding. If SemantIC shows red, take it seriously – yellow cues, by contrast, can sometimes be off.
Read it right
Green means “nothing found” – not “flawless”. It focuses your visual check; it doesn't replace it.
Known gap
The tool finds local hand and finger artefacts more readily than structural body flaws (head, limbs, leg placement) – but it depends on the image: in the tests there were both missed and falsely reported findings, checked across two vision models. An anatomy flag is therefore a check request to you, not a diagnosis.
Framing
Severe, moderate, minor are an estimate by the vision model, not a measurement. Whether a marked point is publication-critical is decided by your own look at the image.
The same tests also surfaced findings that a human eye had missed. The cues complement your visual check – every finding is built to be verifiable with a marking on the image.
08 · Where the criteria come from
The check logic rests on an empirical basis of its own: a qualitative content analysis of 144 AI-generated images. From it came the recurring error profiles, the reading modes and the visual drivers the tool works with – the practical part of the bachelor's thesis “Visual Bias in the AI-generated Image” (Multimedia Production, FH Graubünden).
The pipeline was developed over documented iterations and tested against the human coding – the validated limits above come from exactly these tests. It is deliberately model-agnostic: criteria, evidence requirements and check rules hang on the research basis, not on an AI model. The models used are interchangeable – the check grid stays.
09 · Privacy
Your image is transmitted to external services for analysis and is not stored at SemantIC itself – exactly what goes to whom and how long anything is kept is set out in the privacy policy.