How to Tell If an Image Is AI-Generated: A Practical Field Guide

Here's a question worth sitting with for a moment: if a photorealistic image of a place you've never visited appears in your news feed, what's your actual probability of correctly identifying it as AI-generated — before reading the caption? Most people who've thought about this guess somewhere around 70–80%. Controlled studies put the real number closer to 50–55%. That's coin-flip territory. This guide is designed to move you well above that baseline, methodically.

Why Our Intuition Fails Us (And Why That's Interesting)

The immediate instinct is to look for something obviously wrong — a melting hand, a blurred background that doesn't make physical sense, text that looks like it was composed by someone who learned the alphabet from a fever dream. And that instinct is almost right, except on one critical point.

Those catastrophic failures were characteristic of generative models from 2021–2022. The models used in production today — Flux, Midjourney v6, Stable Diffusion 3, Imagen 3 — fail in far more subtle ways. The error signature has changed. What used to be an obvious hand with eight fingers is now a slightly odd texture on a jacket collar that you'd probably scroll past.

This matters because if you're still using a mental checklist written for 2022-era generators, you're calibrated to the wrong problem. You'll miss the images that actually circulate today.

The general rule, before we get into specifics: the tell is almost never catastrophic anymore. It's structural — in the physics of light, the statistical distribution of fine detail, the coherence between foreground and background. Train your eye for those, not for melted faces.

The Visual Inspection Layer: What to Look For First

Start with what you can see without any tools. This sounds obvious, but the discipline here is to look at the right things in the right order — not just to have a vague impression.

Lighting Coherence

Cast shadows and specular highlights are the single most reliable visual tell in current AI imagery. Ask yourself: if there's a light source on the left side of the face (catchlight in the left eye, shadow falling right), does every other object in the scene follow that same geometry?

Think of it like a physics constraint. In a real photograph, all light obeys the same inverse-square law from the same sources. AI models synthesize lighting probabilistically — each region of the image is influenced by training data independently, and the global constraint is enforced only loosely. The result: a face lit from the upper left while the background wall behind it is lit from the front, with no apparent reason for the difference.

This isn't always present, but when it is, it's decisive.

Fine Detail Repetition and Texture Collapse

Zoom in — not to 100%, but to 200 or 300% — on regions with fine stochastic texture: hair, fabric weave, grass, gravel, brick. In a real photograph, every strand of hair is physically distinct. In a diffusion model output, you'll often find a kind of statistical averaging: the texture looks right from a distance but resolves into a repeating motif or a smooth blur at close inspection, rather than the genuine entropy of the real world.

This is especially visible at the boundary between two textures — where hair meets a shirt collar, or where a hand rests on a wooden surface. Real photos show a clean physical boundary. AI images often show a blending zone where the model was uncertain about which texture to commit to.

Ears, Teeth, and Jewelry

Hands are the famous failure point, and everyone checks them now. But ears and teeth are statistically more reliable in current models — they fail just as often and are checked far less. Earrings in particular: a pair of earrings in an AI image will frequently be geometrically asymmetric in a way that a real person wearing real earrings never would be. One stud will sit slightly higher, or have a different number of visible facets.

Teeth: real human teeth have individual variation in color, translucency, and alignment. AI teeth tend toward a homogeneous whiteness and an overly regular arrangement that looks more like a dental stock photo than an actual mouth.

Rule of thumb: focus your visual inspection on paired, symmetric features — the places where physical symmetry is near-perfect in reality but expensive for a model to enforce precisely.

Background Coherence

AI models are good at generating a plausible subject. They are significantly worse at generating a background that is consistent with the story the subject implies. If a person is depicted as standing in a kitchen, look at the counter behind them: are the objects on it physically plausible in their arrangement? Is the perspective of the cabinets consistent with the perspective of the person?

Backgrounds also tend to contain phantom objects — shapes that suggest something (a lamp, a window frame, a shelf) but don't fully resolve into that thing. They exist in a superposition of plausible objects. That's not how rooms work.

The Metadata Layer: What the File Itself Can Tell You

Visual inspection is necessary but not sufficient. A well-generated image from a professional workflow — say, an image from Midjourney processed through Lightroom and exported via a reputable stock platform — may have had its most obvious artifacts cleaned up. The metadata is a second independent signal.

EXIF Data: Presence and Plausibility

Real photographs taken with a camera contain EXIF metadata: camera make and model, lens focal length, shutter speed, aperture, ISO, GPS coordinates if location was enabled, and a timestamp. Images generated by AI generators typically have either no EXIF data at all, or EXIF data that was added artificially after generation.

Use a tool like ExifTool (free, command-line) or an online EXIF viewer. What you're looking for isn't just presence but plausibility. A photo allegedly taken outdoors in daylight with ISO 3200 and a shutter speed of 1/30s is suspicious. An image with a camera model field that says "NIKON Z9" but no lens data is suspicious. Absence of EXIF in a supposedly candid documentary photo is suspicious.

Important caveat: many social platforms strip EXIF data on upload. The absence of EXIF on an image downloaded from Twitter or Instagram proves nothing on its own. The presence of consistent, physically plausible EXIF is mildly reassuring — but it's fakeable, so weight it accordingly.

C2PA Content Credentials and SynthID

This is where the topic connects to something more sophisticated. The Coalition for Content Provenance and Authenticity (C2PA) standard allows an AI generator to embed a cryptographically signed manifest in the image file, declaring that the image was AI-generated, by what model, and when. Google's SynthID embeds an imperceptible watermark directly into the pixel distribution of the image, designed to survive compression and resizing.

If you encounter an image that carries C2PA metadata, a viewer like Content Credentials Verify will show you the provenance chain. This is the most reliable signal available — when it's present. The catch: it's only present when the generator or publisher chose to include it, which is far from universal.

For a deeper analysis of what SynthID actually does and what it doesn't do — including why 'removing' it is a category error — the article on SynthID bypass explained covers the technical landscape honestly. It's worth reading alongside this guide if you're trying to understand the full detection stack.

Rule of thumb: treat metadata as corroborating evidence, not as proof. A clean bill of health from EXIF means little. A C2PA signature from a known generator is close to conclusive.

Automated Detection Tools: What They Can and Can't Do

There's a whole category of tools that promise to tell you, with a confidence percentage, whether an image is AI-generated. Some of them are useful. Most of the marketing around them is not.

How Classifier-Based Detectors Work

The dominant approach is a binary classifier trained on a dataset of real photographs and AI-generated images. The model learns statistical patterns that distinguish the two distributions — patterns that are often invisible to the human eye but measurable in the frequency domain or in the distribution of pixel values at fine scales.

Think of it like the difference between a handmade Persian rug and a machine-made imitation. From across the room, they look identical. Under a microscope, the knot distribution in the handmade rug is genuinely random, while the machine-made one has a subtle periodicity. The classifier is essentially doing a microscopic inspection of the image's statistical texture.

We ran our own tests on this, documented in the piece where we tested our detection algorithm — including the failure modes that most tool providers don't advertise.

The Generalization Problem

Here's the core limitation of classifier-based detectors: they are trained on specific generators and tested on those same generators. When a new model architecture is released — or when an existing model is fine-tuned — the classifier's performance drops, sometimes dramatically.

A classifier trained primarily on Midjourney v5 outputs will underperform on Flux 1.1 outputs, because the statistical signature of the artifacts is different. A classifier that achieves 95% accuracy on a benchmark dataset may achieve 60% accuracy on images from a generator released six months after the benchmark was created.

This is not a flaw that careful engineering will fix. It's a structural property of the adversarial dynamic between generation and detection. The moment a detection method becomes widely known, the generators are fine-tuned against it — sometimes deliberately, sometimes as a side effect of other improvements.

Tools Worth Using (with Calibrated Expectations)

That said, some tools are genuinely useful as part of a multi-signal workflow:

• Hive Moderation AI Detector — updated frequently, covers most major current generators, gives per-generator probabilities rather than a single score.
• Illuminarty — strong on Midjourney and DALL-E 3, provides a visualization of which regions triggered the detection.
• AI or Not — simple interface, reasonable accuracy for general use cases.
• Google's SynthID detector — only works on images generated by Google's own models (Imagen, Gemini). If you're specifically investigating a Gemini-generated image, it's the most reliable option available.

Use them as one input among several, not as a verdict. A 73% AI probability from a classifier means: "this image has statistical properties more consistent with AI generation than with real photography in our training data." It does not mean the image is 73% AI-generated.

Rule of thumb: no single automated tool should be your final answer. Three independent signals pointing the same direction — visual artifacts, suspicious metadata, high classifier score — is a much stronger basis for a conclusion than any one of them alone.

Semantic and Contextual Signals: The Layer Most Guides Skip

There's a layer of analysis that most detection guides don't cover, because it requires thinking rather than running a tool. It's often the most decisive layer.

The Provenance Question

Where did this image come from? Who shared it first, and through what channel? A photorealistic image of a dramatic event, shared by an account created three weeks ago with no prior history, is a different proposition than the same image appearing in an established news outlet with a photo credit and a wire service timestamp.

This sounds obvious, but it's systematically underweighted in practice. People focus intensely on the image itself while spending almost no time on its claimed provenance. In most real-world detection cases, the provenance check resolves the question faster than any visual analysis.

The Specificity Trap

AI-generated images tend toward the archetypal. A generated image of "a protest" will show what a protest generically looks like — the crowd, the signs, the raised fists — rather than the specific chaos, specific typography, and specific people of a real event at a real time and place.

Look for specificity that would be expensive to fabricate: a license plate that matches the state's actual format, signage in the background that corresponds to a real business at the claimed location, clothing that reflects the specific fashion norms of the claimed time and place rather than a generic version of them.

Real photographs are full of incidental detail that makes no sense from a generative standpoint — a partially visible brand logo on a cup, a specific street sign in the background that places the image at a real intersection. AI images tend to fill that space with plausible-looking but non-specific placeholders.

Reverse Image Search

Run it. Always. Google Images, TinEye, and Yandex Images cover different indexes. If the image appears on a stock site under a generic description, or if an earlier version of it appears without the claimed context, that's material information.

If you're specifically dealing with watermarked images — from stock libraries like Getty, Shutterstock, or iStock — the watermark itself is useful provenance information. An image appearing to document a real event that turns out to carry a stock library's visual watermark pattern is not a document; it's a licensed image. Articles on Getty watermark removal and Shutterstock watermark removal discuss the watermark patterns these libraries use, which may help you recognize them when they appear in manipulated form.

Rule of thumb: context is data. An image doesn't exist in isolation. The story around it constrains what it can plausibly be.

Building a Repeatable Detection Workflow

All of this is more useful when it's organized into a repeatable sequence rather than a random checklist. Here's the workflow I'd recommend for a methodical assessment:

1. Provenance first. Where did this image originate? Who first published it, when, and with what context? Run a reverse image search before investing time in pixel analysis.
2. EXIF inspection. Download the image (not a screenshot — the file itself). Open it in ExifTool or an equivalent. Check for presence and plausibility of camera metadata. Look for C2PA signatures.
3. Visual inspection — macro level. Lighting coherence across the full scene. Background plausibility. Semantic consistency between the subject and environment.
4. Visual inspection — micro level. Zoom to 200%+ on ears, teeth, jewelry, and texture boundaries. Look for the statistical averaging signature in fine detail areas.
5. Automated classifier. Run through one or two classifiers, treating the output as a probability estimate, not a verdict.
6. Triangulate. Do the signals agree? Suspicious EXIF + high classifier score + texture artifacts pointing the same direction is strong. One signal in isolation is weak.

The entire sequence takes four to seven minutes for a thorough assessment. For most images, steps one and two will either resolve the question or substantially narrow it.

One practical note: this workflow assumes you have access to the original file. If you're working from a social media screenshot, you've lost the EXIF layer entirely and the visual inspection layer is compromised by compression artifacts. The base rate for AI-generated images in certain contexts is high enough that uncertainty should be your default, not confidence.

For a related perspective on how AI image processing actually works under the hood — relevant to understanding why these artifacts appear where they do — the article on how AI inpainting works gives a technically honest explanation of the underlying mechanisms.

What Changes When the Image Has Been Post-Processed

A significant fraction of AI images in circulation have been edited after generation — cropped, color-graded, compressed multiple times through social media pipelines, or selectively touched up. This is the hardest case for detection.

Post-processing degrades both visual artifacts and metadata signals. Multiple rounds of JPEG compression add noise that masks the statistical signature classifiers look for. Color grading can shift the lighting coherence just enough to make it look intentional. Cropping removes context.

Three signals survive post-processing better than others:

• Semantic inconsistency. No amount of color grading makes a background wall lit from two incompatible directions make physical sense. Visual coherence failures at the semantic level — objects that don't belong, perspectives that don't resolve — persist through most post-processing.
• Fine-detail texture at the micro level. Even after JPEG compression, the pattern of texture averaging at hair-to-skin boundaries tends to survive in a recognizable form, because it's a structural property of how the model synthesizes, not just a surface noise pattern.
• The provenance and context signals. Post-processing doesn't change where the image first appeared or the account that posted it.

When dealing with images that may have been deliberately processed to defeat detection — a real concern in adversarial contexts — lean even harder on provenance and semantic coherence, and weight classifier scores less.

It's also worth being honest about the ceiling here: a sufficiently skilled human editor, working deliberately on a high-quality AI image with the specific goal of defeating detection, can produce something that will fool most workflows most of the time. That's the current state of play. The goal of a detection workflow is not to be infallible; it's to be substantially better than chance and to make the cost of deception high enough to deter casual misuse.

FAQ

Can I tell if an image is AI-generated just by looking at it?

Sometimes, but less often than most people think. Research puts unaided human accuracy at roughly 50–55% for current high-quality AI generators — close to random chance. Visual inspection is most useful when you know what to look for: lighting coherence across the full scene, texture averaging in fine-detail areas like hair and fabric, and asymmetry in paired features like earrings or teeth. Even then, treat it as one signal among several rather than a verdict on its own.

What is SynthID and does it make AI images detectable?

SynthID is a steganographic watermarking system developed by Google DeepMind, embedded directly into the pixel distribution of images generated by Google's models (Imagen, Gemini). It's designed to survive compression and resizing and is imperceptible to the human eye. When present, it makes those specific images reliably detectable by Google's detector. The limitation: it only applies to Google-generated images, it requires the original generator to have applied it, and it's a separate system from the visual artifact analysis described in this guide. For a full technical breakdown, the article on SynthID bypass explained covers what the system does and doesn't guarantee.

Are AI detection tools reliable?

Useful, but not reliable as standalone verdicts. The best current classifiers achieve 85–95% accuracy on benchmark datasets, but performance drops significantly on generators released after the classifier was trained. The fundamental problem is that detection and generation are in an adversarial dynamic: as detection methods improve, generators improve in ways that defeat them. Use classifiers as one input in a multi-signal workflow — combined with provenance checks, EXIF inspection, and visual analysis — not as definitive answers.

Does EXIF data prove an image is real?

No. EXIF data can be added to any file after generation using widely available tools. Plausible, internally consistent EXIF data is weakly reassuring; implausible or absent EXIF is a mild signal worth noting. The stronger signal is the presence of C2PA Content Credentials from a known generator, which carries a cryptographic signature that's much harder to fake. But even that only tells you about images where the generator chose to embed those credentials — which is not universal.

What are the most common visual artifacts in current AI-generated images?

The most persistent in 2025–2026 generators: (1) lighting direction inconsistency between foreground subject and background, (2) texture averaging at boundaries between different materials (hair-to-skin, hand-to-object), (3) asymmetry in paired features like earrings or eyebrows, (4) teeth that are too uniformly white and regularly aligned, (5) background objects that suggest a shape without fully resolving into one. Obvious failures like extra fingers are increasingly rare in top-tier models.

Can an AI-generated image fool a reverse image search?

Yes, if the image was generated specifically for the use case and hasn't been published elsewhere before. Reverse image search is most useful for detecting AI images that were repurposed from stock libraries or used in a different context originally. For a genuinely novel generated image, it won't find a match — but the absence of a match doesn't confirm authenticity either. It's a filter, not a proof.

If an image has a stock watermark, does that mean it's real?

Not necessarily. Stock libraries like Getty, Shutterstock, and Adobe Stock now include AI-generated images in their catalogs — properly disclosed as such. A stock watermark tells you the image is licensed through that platform, not that it was photographed rather than generated. The platform's disclosure practices vary. The visual detection workflow described in this guide applies equally to watermarked stock images.