How to Detect AI-Generated Images in 2026

In 2023, spotting an AI image meant finding the extra finger. In 2026, that heuristic is dead. Modern diffusion models handle hands, faces and lighting with an accuracy that would have seemed impossible eighteen months ago. The visual tells that journalists and fact-checkers relied on have largely been patched by the models themselves.

That is not a reason to give up. It is a reason to use better methods.

The current state of AI image generation

To understand why detection is hard, you need to understand what you are up against. Four generators dominate the 2026 landscape, and they do not fail in the same ways.

What these models still cannot do

Even the best generators have systematic weaknesses, though they are shrinking with each model release. The remaining limitations in 2026 are mostly about contextual coherence rather than object rendering: a room where windows cast shadows in two incompatible directions; a crowd where no two people's clothing interacts plausibly with the lighting; a "newspaper" headline that reads correctly but has fonts that shift mid-word at high zoom.

These are not always visible at a glance. They require deliberate scrutiny. And when images are compressed, resized or screenshot-cycled through social media, even these artefacts often disappear.

Visual tells that still work (and when they fail)

Visual inspection is not useless — it is just unreliable as a standalone method. Here is what experienced fact-checkers still check, and an honest assessment of when each tell fails.

Hands and fingers

The classic tell. AI models historically generated hands with six fingers, fused knuckles or anatomically impossible joint angles. Flux.1 and Midjourney v6 have largely solved this for common poses. Where it still holds: complex hand gestures, hands holding objects, two hands interacting, or hands at unusual angles — extended fingers viewed from above, for instance. If the subject's hands are not visible at all, that is also worth noting; a great many AI images are composed specifically to avoid showing hands.

Text and lettering

AI-generated text inside images is frequently wrong in ways that are hard to describe but easy to spot: letters that look correct at a glance but do not form real words; fonts that drift across a single line; text that wraps nonsensically. DALL·E 3 can render short, common words correctly. Longer strings, unusual names and non-English scripts remain unreliable. Check any readable text in the image carefully, and zoom in — compression frequently hides exactly this class of error.

Background symmetry and repetition

Diffusion models fill backgrounds by sampling from texture distributions. When the model runs out of unique training signal, it tiles. Look for crowds where multiple people share the same face at different scales, brick walls with identical mortar lines repeating on an unnatural grid, or foliage that repeats in a fractal pattern. These artefacts are most visible in high-resolution versions of an image, not in previews.

Lighting and shadow consistency

Single-light-source coherence is something humans are extremely sensitive to but often do not consciously notice. In AI images, shadows frequently do not match: a subject lit from the left has a shadow falling to the right; reflections in the eyes show a window that is not present in the scene; skin highlights imply a light source that does not match the room. This is one of the most reliable tells in photorealistic images, but it requires slowing down and explicitly asking "where is the light coming from?" for every shadow in the frame.

Texture artefacts in fabric, fur and hair

AI textures look statistically correct at medium zoom but break down at high zoom. Fabric weave has an inconsistent thread count. Hair at the edges of the frame becomes soft in a way that does not match real optical blur. Animal fur often tiles or shows discontinuities where the model lost track of direction. This is most visible when the image has been preserved at original resolution — which is precisely what social media compression destroys.

Why visual inspection alone fails at scale

The practical limitation of visual inspection is not knowledge — it is time. A journalist or content moderator reviewing 200 images a day cannot spend four minutes on each one. At real-world moderation scale — election misinformation campaigns, financial fraud schemes, academic submission pipelines — the sheer volume of images makes human review the bottleneck.

The adversarial problem

Bad actors specifically optimise against visual inspection. They know moderators check hands, check text, check lighting. The response is to generate images that avoid those failure modes — subjects with hands hidden, text removed from the composition, controlled studio lighting — or to run generated images through post-processing pipelines that add film grain, JPEG artefacts and colour noise to obscure generator fingerprints.

This is not theoretical. Research published in early 2026 documented coordinated disinformation campaigns using AI-generated headshots that had been processed through multiple compression and resampling cycles specifically to defeat both visual inspection and first-generation detection tools. In the blind tests reported in that work, the images passed human review at a 94% rate — including review by professional fact-checkers.

Scale requirements versus human capacity

A platform receiving 10 million images a day cannot employ enough reviewers to inspect a meaningful fraction of them manually. Even sampling strategies fail when the base rate of AI-generated content is rising: a 2% base rate means 200,000 AI images per day at that scale. At three minutes per image, you would need thousands of full-time reviewers just to cover the fraction that sampling surfaces.

The gap between what a human expert can detect with unlimited time and what is practically detectable at production scale is the entire problem. Automated detection is what closes that gap.

How automated image detection works

Automated AI image detection works by analysing statistical properties of image data that are invisible to the human eye but measurable computationally. There are three main approaches, and understanding them helps you evaluate any tool that claims to do this — including ours.

Frequency domain analysis

Real photographs have a specific distribution of high-frequency detail — noise, grain, fine texture — that differs from synthetically generated images. Camera sensors introduce noise in characteristic patterns, lenses introduce characteristic optical aberrations, and JPEG compression creates characteristic artefacts. Diffusion models produce a different statistical fingerprint in the frequency domain: they are often "too clean", or they show frequency artefacts corresponding to the upsampling steps in the generation process. Several commercial detectors, SightEngine among them, use frequency analysis as one signal in a larger classifier stack.

Neural artefact detection

Large detection models are trained on datasets of known AI-generated and real images, learning the subtle patterns left by particular generator architectures. Midjourney v6 leaves different traces than Stable Diffusion XL; DALL·E 3 images have distinctive tonal and edge characteristics. The challenge is that these fingerprints shift with every model update, which means continuous retraining. Detectors that have not been updated against Flux.1, released late in 2025, perform measurably worse on that generator than on older ones.

It is worth being precise about what this class of model outputs. A neural detector returns a probability that an image is synthetic. It does not return an author. Architecture-specific fingerprints inform the training, but a probability score is not evidence that a particular product made a particular picture.

Metadata and provenance signals

The C2PA standard — Coalition for Content Provenance and Authenticity — allows images to be signed with cryptographic provenance data at the point of generation. DALL·E 3 and Adobe Firefly already embed C2PA content credentials. Where present, this is the strongest signal available: it is not probabilistic, it is cryptographically verifiable, and it can state outright that a file was machine-generated. The limitation is coverage. Adversarial actors strip metadata, plenty of generators never implement the standard at all, and many platforms discard credentials on re-encode. Provenance is a best-case tool, not a general solution.

What published benchmarks measure

The figures below come from published evaluations of the named tools rather than vendor marketing or user anecdotes. They are for unmodified, non-adversarially processed images.

Add post-processing — resampling, added grain, repeated compression cycles — and every one of those detection rates drops, typically by 15 to 25 percentage points. Note also that no single detector performs consistently across all generator types. That inconsistency is the strongest practical argument against relying on any one of them.

How our image check works

Our image check runs an uploaded or linked image through several independent detectors and, in parallel, reads any content-provenance credentials embedded in the file. You get every detector's raw score, the provenance result if one exists, and a combined verdict that explains itself.

Why one detector is not enough

Every detector has a training distribution — the set of generators and image types it has learned to recognise. When a new generator such as Flux.1 arrives, a single-detector tool is effectively blind to a meaningful share of AI imagery until it has been retrained, which can take weeks or months. Running several detectors with different training distributions gives you coverage during those gaps, because they rarely all go stale at once.

What we do instead of averaging

Combining detectors only helps if you combine them well. Suppose one detector returns 91% and another returns 34%. Averaging gives 62.5%, a number that reflects neither model's actual finding and sits awkwardly near most thresholds. Plain averaging dilutes the signal from whichever detector actually found something.

The tempting fix is a maximum rule: flag the image if any detector crosses the threshold. We deliberately do not do that. It would let one over-eager model veto every other signal, and the cost lands on people submitting genuine photographs. Instead we use a reliability- and decisiveness-weighted consensus, which avoids plain averaging without adopting a maximum rule.

Detectors close to 50 are treated as abstaining. Anything within ±8 points of 50 is excluded from the calculation entirely. A model returning 48% is not casting a vote — it is telling you it does not know, and a non-answer should not be allowed to drag a confident result toward the middle. The score is still displayed to you; it simply does not enter the maths.

Remaining votes are weighted by decisiveness and reliability. Among detectors that did commit, each vote is weighted by how far it sits from 50 — its decisiveness — multiplied by a reliability factor specific to that detector.

The reliability weights are deliberate, not incidental. A declared content-provenance credential outweighs any statistical guess, because a signed declaration that a file was machine-generated is closer to ground truth than a pixel-level inference. And detectors we know to be false-positive prone are weighted so that they cannot single-handedly override a confident ensemble.

Genuine disagreement is surfaced, not smoothed. When confident detectors point in genuinely opposite directions, we say so rather than emitting a midpoint and calling it a verdict. A split result is a real finding about an image, and flattening it into 62.5% throws that information away.

Applied to the example above, the low 34% score might mean that detector has not been updated for the generator in question — or it might mean the image genuinely carries mixed signals and the 91% is a false positive. You see both numbers and the weighting, so you can make that judgement instead of inheriting ours.

What we do not claim

We do not tell you which generator made an image. A statistical detector produces a probability that an image is synthetic; it does not produce attribution. If you see us name a generator, it is because a provenance credential in the file declares it — never because a classifier guessed. Any tool that reports "this was made by Midjourney" off the back of a probability score is overstating what the underlying maths supports.

We also do not claim complete detection on adversarially processed images. If a picture has been through several rounds of compression and resampling designed specifically to defeat detectors, no current tool reliably catches it. The distance between a clean AI-generated image and an adversarially laundered one is the frontier of this problem, and anyone claiming to have closed it is selling something.