Is This Image AI Generated?

The reason is not that people are careless. It is that the artifacts are not in the parts of an image anyone looks at. Generation leaves its trace in fine pixel statistics, in invisible watermarks, and in content credentials written into the file. None of that is visible at any zoom level on any screen.

The visual tells that do exist are the ones being actively fixed. Midjourney's hands got better between v5 and v6. DALL-E 3 renders legible text. Flux does both. Every checklist of "look for six fingers" is a checklist of problems that model authors prioritised, which means it describes the previous generation by definition.

The checks run in order. FauxLens reads the C2PA content credentials and EXIF or XMP metadata, looks for the invisible watermark OpenAI and Google embed in images their AI tools make, and then has a specialist detection model score the pixels. If that model is unavailable, a multimodal vision model takes over.

The result is one verdict with a confidence score and plain-language evidence lines. The confidence is not a calibrated probability, and a heavily compressed image from a messaging app usually gets a less certain answer, because that is the honest result for that input.

Most scans take a few seconds, and a slow one can take up to about 45. When a content credential or a generator signature in the metadata settles the question, the scan stops early, and the evidence says so.

Then the file is discarded. Not queued for deletion, not retained for a window. There is nothing left to leak.

Step 2 - Wait for the analysis. The checks run in order and most scans finish within a few seconds. You will see a progress indicator while the analysis runs.

Step 3 - Read the overall verdict at the top of the report. The verdict is "AI-Generated," "Likely AI-Generated," or "No AI Detected." The confidence score tells you how strongly the evidence supports the verdict. A confidence of 90% or higher is a strong result. Below 70% means the signals are mixed and the image may have been heavily post-processed.

Step 4 - Review the evidence. Scroll down in the report to see what each check found. A result that says an AI tool altered part of the image usually means a real photograph that was AI-upscaled or partly repainted, not something generated from nothing. Treat that as an edited photo rather than a fake one, and go looking for which region was touched. A watermark or a content credential naming an AI tool is the strongest evidence there is.

Step 5 - Check who made it, if the file says. A content credential or the metadata can name the AI tool, and a watermark names the vendor, OpenAI or Google. FauxLens does not guess a specific model such as Midjourney, DALL-E 3 or Flux from the pixels alone.

Step 6 - Decide on next steps. If you are satisfied with the verdict, the report is shareable via a unique link. If the result is inconclusive and the stakes are high, see the limitations section below and consider additional verification steps.

Hands and fingers: Count fingers carefully. AI generators produce hands with six fingers, merged digits, or rubber-like joints more often than any other body part. Midjourney v6 has improved significantly but fails on complex hand poses.

Background text: Look for any text in the background: signs, labels, books, anything with writing. AI generators hallucinate plausible-looking but phonetically meaningless letter sequences. The text will look like writing but read as nonsense.

Ear structure: Ears are geometrically complex and AI models often generate ears that are asymmetric or structurally incorrect - lobes that merge into the jawline, missing antihelix curvature, or ears at slightly impossible angles relative to the head.

Eye reflections: In a real photograph, both eyes reflect the same light source from the same angle. AI-generated faces often have inconsistent reflections between the two eyes, or reflections that do not correspond to the background lighting in the image.

Skin texture: Real skin has pores, fine hairs, slight color variations, and texture that changes across the face. AI-generated skin is often too smooth and too uniform - it looks airbrushed even at high resolution. Zoom in to 200% on the forehead or cheek.

Background coherence: Look at the background elements behind the main subject. AI generators often produce background objects with impossible geometry: chairs that do not connect properly, buildings that violate perspective, grass or foliage with unnatural repetition.

These visual checks are useful but not sufficient on their own. The latest generators produce images that pass all visual checks. A forensic analysis is necessary for any decision that carries real consequences.

"Likely AI-Generated" means that AI-generation signals are present but the confidence is in the 55-85% range. This result typically occurs when: the image has been re-compressed by a social media platform (which degrades some forensic signals), the image is a crop of a larger AI-generated image (reducing the amount of analyzable pixel data), or the image was AI-generated but then edited with Photoshop or other tools afterward (which can partially mask AI fingerprints). A "Likely AI-Generated" verdict is not a weak result. It means the evidence leans strongly toward AI origin with some degree of uncertainty. Treat it as a strong indicator requiring additional verification before acting on the assumption the image is authentic.

"No AI Detected" means none of the checks found reliable signs of AI generation in the image. This does not guarantee the image is authentic. It means the analysis found no evidence of AI origin at the time of scanning. An adversarially optimized image designed to evade detection, or an AI image that has been extremely heavily post-processed, could potentially return a "No AI Detected" result. If you have strong independent reasons to suspect AI generation despite this result, contact us with the image for a deeper manual review.

If you disagree with the result: consider whether the image was re-compressed before you received it (which degrades signals), request the original uncompressed file from the source, and run the original through FauxLens. If the stakes are significant (legal, financial, or journalistic), engage a certified digital forensic examiner for chain-of-custody analysis and sworn expert testimony.