Illuminarty AI Image Detector: What It Checks and How to Verify Results

Illuminarty positions itself as a detector for AI-generated images, and separately offers a check aimed at AI-written text. On the image side, tools in this category generally work by feeding an uploaded image through a classifier trained to spot the statistical fingerprints that generative models leave behind — texture patterns that are too smooth or too regular, lighting that doesn't behave the way a camera sensor would render it, and frequency-domain artifacts that are invisible to the eye but detectable to a model trained on thousands of known AI and real images. The output is typically a probability or confidence score rather than a flat yes/no answer, which matters: a 72% AI-likelihood score is a signal to investigate further, not a verdict to act on by itself.

No independent, continuously updated public benchmark tracks Illuminarty's accuracy specifically, and any accuracy number a vendor publishes reflects a specific test set at a specific point in time — not how the tool performs on the image you just uploaded. What's well documented across the whole category of AI image detectors is that accuracy drops on images that have been resized, re-compressed, cropped, or run through a second editing pass after generation, because those steps disturb the exact artifacts the classifier was trained to look for. Detectors also tend to perform worse on newer generator models than on the ones they were trained against, since each new image model produces a slightly different artifact signature. Treat any single score as one data point produced under conditions you don't fully control, not a certified result.

Yes, often significantly. Re-saving an image as a JPEG at a lower quality setting, resizing it for a social platform, or taking a screenshot of it all strip or alter the fine-grained pixel-level information that AI image detectors rely on. A screenshot is a particularly common failure case: it captures what's rendered on screen but discards the original file's metadata entirely and adds a fresh layer of compression and color processing from the operating system's screenshot pipeline. If you're testing a detector's reliability, run the same source image through it twice — once as the original file and once after a screenshot or a save-as-JPEG round trip — and see how much the score moves. A tool that swings wildly between the two is telling you it's sensitive to file handling, not just image content.

Often, yes, when metadata is actually present. Camera photos typically carry EXIF data — camera make and model, lens, shutter speed, GPS coordinates if enabled — that AI-generated images simply don't have unless someone added it deliberately. Some AI generation tools and editing platforms now write C2PA Content Credentials into a file, a tamper-evident record of how an image was created or edited, which is a stronger signal than a classifier's guess when it's present and verifiable. The catch is that metadata is easy to strip — most social platforms remove EXIF data automatically on upload, and a simple screenshot discards it entirely — so its absence proves nothing on its own. Treat metadata as a strong positive signal when found intact, and a neutral result when missing, not as its own kind of negative proof.

Detectors automate a version of what a trained eye can often catch manually, and knowing what to look for makes you a better judge of any tool's output. Hands, teeth, and text within an image remain common trouble spots for many generators — extra or fused fingers, garbled signage, or asymmetric earrings and jewelry. Backgrounds are worth a close look too: repeating textures like brick or foliage sometimes tile in ways a real photo wouldn't, and reflections in glass or water occasionally don't match what should be reflected. Lighting consistency matters as well — shadows that fall in physically inconsistent directions across the same scene are a stronger tell than any single blurry detail, since a real camera captures one consistent light source.

The most reliable approach doesn't rely on any single detector, including Illuminarty. Start by building a small personal test set: a handful of images you know for certain are AI-generated (from a generator you used yourself) and a handful you know are real (a personal photo with intact metadata). Run all of them through whichever detector you're evaluating and note where it agrees with the known answer and where it doesn't — this tells you far more about a tool's real-world behavior than its marketing page does. From there, layer checks rather than picking one: a detector score, a metadata check, and a manual look at artifacts each catch different failure modes, and agreement across two or three methods is a far stronger basis for a conclusion than any one method alone.

Comparing tools is a reasonable habit precisely because no single detector catches everything, and different tools are trained on different data with different blind spots. If Illuminarty's result on a specific image seems uncertain or you want a second read, running the same file through another detector — NotGPT's image detector is one practical option for this kind of cross-check — can confirm whether multiple tools agree or whether the result is closer to a coin flip. Consistency across two or three independently built detectors is meaningfully more convincing than any single score, especially for an image where the stakes of getting the call wrong are real, like a dispute over authenticity in journalism, research, or a marketplace listing.