Image generators have made spectacular progress, yet they still leave traces. These AI image artifacts — small recurring flaws in anatomy, text, reflections or textures — remain one of the most accessible ways to spot a synthetic image with the naked eye. You just need to know where to look and understand which ones are vanishing with recent models.
This guide lays out a detailed catalog of about a dozen characteristic AI image flaws, with a summary table and tips for turning these visual cues into a genuine verification reflex.
Why AI Images Produce Artifacts
Understanding where flaws come from helps you know where to look. A generator does not "understand" a scene: it predicts statistically plausible pixels, with no model of the physical world or the human body.
No World Model
The AI has no notion of a skeleton, of gravity, or of conservation of matter. It reproduces learned correlations. As a result, whatever is rare in the training data (close-up hands, precise text, complex reflections) becomes a high-risk zone for errors. This is the foundation for learning to spot AI visually.
Flaws That Regress — But Not All of Them
Each model generation fixes part of the crude artifacts. Six-fingered hands become rarer; global semantic inconsistencies (shadows, reflections, scene logic) persist longer. The most durable AI image cues are those that require an understanding of the world, not just a believable texture. Our general principles are in the guide on how to detect an AI-generated image.
Two Families of Artifacts: Local and Global
It helps to distinguish two families. Local artifacts concern an isolated detail: an extra finger, a melted tooth, an absurd letter. These correct themselves relatively well as resolution and model size increase. Global artifacts concern the coherence of the whole: a shadow that contradicts the light across the image, a reflection incompatible with the scene, a perspective that doesn't close. These require spatial reasoning the generators lack, and remain the strongest cues. When you're trying to conclude, always favor the global over the local.
Human Anatomy: The Richest Terrain
The human body concentrates the largest number of detectable errors, because it combines rigid structure with subtle variation.
Hands, Fingers and Nails
AI hands remain emblematic: extra or missing fingers, inconsistent lengths, impossible knuckles, fingers fused together. Even when they look fine at first glance, they often betray a detail on zoom: a melted nail, a missing joint.
Eyes, Teeth and Ears
Eyes display catchlights that should be identical between left and right — AI often desynchronizes them. Irises can take irregular shapes. Teeth appear in variable numbers or merge into a single mass. Ears, rarely scrutinized, are frequently asymmetric or poorly structured.
Hair, Skin and Symmetry
Hair turns into strands that start from nowhere or fuse with the background. Skin is too smooth, without realistic pores or grain. Elements that should be symmetric (earrings, glasses frames, shoulder pads) differ subtly from left to right.
Limbs, Postures and Contact Points
An often-overlooked signal: the contact points between the body and its environment. A hand that "holds" an object without actually gripping it, fingers passing through a glass, feet floating slightly above the ground, a seated person whose weight doesn't deform the cushion. Count the limbs too: an extra arm behind a character, a leg merging with the neighbor's in a group. Multi-person scenes multiply these errors exponentially, because AI handles occlusions and overlaps poorly.
Text, Patterns and Manufactured Objects
Anything relying on precise logical structure trips the AI up.
Pseudo-Text
This is one of the most reliable AI image cues: signs, books, labels, plates display characters that mimic the alphabet without forming words. Even recent models often fail as soon as the text is long or in the background.
Repeated Patterns and Architecture
Repetitive elements (tiles, bricks, windows, railings, rows of seats) drift: alignments that shift, patterns that change their pitch, impossible perspectives. Architecture shows stairs that lead nowhere or beams that cross illogically.
Jewelry, Watches and Accessories
Jewelry and watches are a classic weak point: breaking links, inconsistent clasps, dials with aberrant hands. These objects demand a mechanical regularity the AI reproduces poorly.
Light, Reflections and Scene Physics
The most durable artifacts stem from global physical coherence, which is hard to fix.
Inconsistent Shadows
Look for multiple contradictory light sources: cast shadows in divergent directions, objects lit differently within the same scene, shadows that are missing or floating.
Impossible Reflections
Mirrors, windows, glasses, metallic surfaces or water should reflect the scene coherently. AI often produces reflections decorrelated from the actual content, or a face in a mirror that doesn't match the subject.
Depth and Blur
Background blur (bokeh) is sometimes applied inconsistently: a sharp object in a zone that should be blurred, or impossible depth transitions. This complements Error Level Analysis (ELA), which reveals recomposited regions.
Color Temperature and White Balance
A real scene has a coherent color temperature, dictated by the light source (cool daylight, warm bulb). Generators sometimes mix incompatible moods within a single image: a face lit in warm light against a cool-lit background, with no logical transition. This white-balance inconsistency, subtle but revealing, escapes most observers even though it betrays the absence of a real physical scene.
Artifact Summary Table
| Artifact | Zone | Reliability today | Difficulty for AI to fix |
|---|---|---|---|
| Hands / fingers | Anatomy | Medium (declining) | Medium |
| Eye catchlights | Anatomy | High | High |
| Fused teeth | Anatomy | Medium | Medium |
| Pseudo-text | Objects | High | High |
| Drifting repeated patterns | Background | High | High |
| Inconsistent jewelry / watches | Objects | Medium | Medium |
| Divergent shadows | Light | Very high | Very high |
| Impossible reflections | Light | Very high | Very high |
| Imperfect symmetry | Anatomy/objects | Medium | Low |
| Over-smooth skin | Texture | Low | Low |
The "light" and "pseudo-text" artifacts are the most robust over time: favor them when in doubt.
Limits of Visual Inspection
Spotting an artifact is sometimes enough to conclude; its absence never proves authenticity.
The False-Negative Trap
An image with no visible flaw can perfectly well be generated by a recent model or retouched to hide the cues. Conversely, a real photo may contain oddities (compression, bad framing) wrongly taken for artifacts. That is why the eye alone is not enough.
Cropping, Upscaling, Recompression
Post-processing erases part of the artifacts: an upscaler smooths hands, recompression masks textures. At that stage, you need signals deeper than visual inspection.
Confirmation Bias
One last trap, a human one: once convinced an image is fake, you "find" artifacts everywhere, including in real photos. Conversely, an image that confirms our expectations escapes critical examination. Rigor means examining every image with the same checklist, regardless of what it depicts, and weighting cues rather than trusting an impression. This is exactly the logic an automated forensic analysis brings: it doesn't "want" to prove anything.
Generator-Specific Artifacts
While the families of flaws are universal, each generator imprints its own signature, useful for refining the diagnosis.
"Style" as a Cue
Some models lean toward a very polished, high-contrast, cinematic aesthetic, to the point where perfection itself becomes suspect: lighting too controlled, flawless skin, impeccable composition. Others produce softer textures or a "painterly" rendering. Recognizing these tendencies helps direct the analysis, without ever constituting proof on its own. To go further on each model's signatures, see our guides on recognizing a Midjourney image and on detecting open-source generations.
How Flaws Evolve Over Time
Today's artifacts are not yesterday's. Early generations multiplied crude aberrations; recent models have almost eliminated them, shifting the battle toward finer cues. This dynamic demands constant monitoring: a frozen catalog of artifacts goes stale fast. That's why serious tools continuously update their detection models, rather than relying on a static list of visual flaws.
The Risk of Over-Reliance on a Single Cue
Beginners often fixate on one famous artifact — hands — and conclude on that basis alone. This is a mistake on two fronts: recent models render hands correctly, producing false negatives, while a genuine photo with an awkward hand pose can trigger a false positive. The robust method always combines several converging cues across different families (anatomy, light, text, texture) before forming a judgment, and treats any single anomaly as a prompt to look further, not as a verdict.
From the Naked Eye to Forensic Analysis
When doubt persists, visual artifacts become a starting point, not a conclusion. A multi-layer analysis objectifies the verdict.
TruthLens's Multi-Signal Approach
TruthLens combines metadata inspection (EXIF, C2PA), Error Level Analysis, AI vision trained on generation signatures, and noise/watermark analysis. Each layer weights a global confidence score, far more robust than a single visual cue.
Crucially, these layers catch what the eye cannot. ELA highlights regions recompressed at a different rate — the telltale sign of a paste-in or an inpainted edit, even when the seam is visually invisible. Noise analysis exposes the absence of a coherent sensor fingerprint, a dead giveaway of synthesis. AI vision recognizes statistical patterns in the pixel distribution that no human inspection could surface. Together, they turn a subjective impression into a measurable, reproducible verdict.
A Defensible Report
Beyond the verdict, TruthLens generates a certified PDF report (SHA-256 hash + timestamp), useful for journalism, insurance or legal use. To analyze a doubtful image, drop it onto the forensic image analysis page. For a no-signup first triage, also see our free verification methods.
Training Your Eye Day to Day
Spotting artifacts is a skill that develops through regular practice.
Build an Observation Routine
Adopt a fixed inspection order: first global coherence (light, perspective), then faces and hands, then text and repeated patterns, finally uniform surfaces. A systematic order avoids skipping a zone and limits confirmation bias. Zooming to 200% or more remains essential for fine details.
Compare Generators and Eras
Practicing on images from different generators sharpens your sensitivity to each model's signatures. Older generations are full of instructive artifacts; the most recent ones force you to hunt for subtler global cues. This exercise prepares you to recognize future developments, where only the deepest signals will remain — and where the forensic tool becomes indispensable.
When to Stop Trusting Your Eyes
There is a clear threshold beyond which visual inspection should hand off to instrumentation: when the image is destined to serve as evidence, when it has obviously been post-processed (recompressed, upscaled, filtered), or when two trained observers disagree. At that point, continuing to argue over a fingernail is counterproductive. The mature move is to run a layered analysis, obtain a weighted score, and — if the stakes warrant — a certified report. The naked eye is a fast, free first filter; it was never meant to be the final word.
FAQ
What is the most reliable artifact for spotting an AI image?
Lighting inconsistencies (shadows and reflections) are the most durable, because they require a physical understanding of the scene that models struggle to reproduce. Pseudo-text comes right behind. Hands, long the go-to, are increasingly unreliable with recent models.
Have recent models eliminated all flaws?
No. They have sharply reduced crude artifacts (hands, anatomy), but global semantic inconsistencies — reflections, shadows, scene logic, background pseudo-text — persist. No generator today is free of every detectable cue.
Can a real photo look like an AI image?
Yes. Aggressive compression, motion blur, bad lighting or clumsy retouching can mimic artifacts. That's why a single cue is never enough: you must cross several signals and ideally resort to multi-layer forensic analysis.
Do you need a tool to detect artifacts?
For a quick scan, a trained eye is often enough. But to reach a verdict in case of doubt, mask false positives and produce evidence, a tool like TruthLens brings signals invisible to the eye and a reasoned verdict.