Is This AI?

Generators have learned to imitate the texture of phone photography: sensor-style grain, slight motion blur, imperfect framing, and even simulated low light can be requested in the prompt. The old shortcut of “too clean to be real” is weaker every year, which is exactly why the question “is this AI” now needs more than a glance.

Hybrid pipelines blur the line further. A common workflow starts from a real photo, applies generative fill to remove or add elements, upscales, and finishes with a filter. Only part of the file is synthetic, so scores often land in the middle range. Meanwhile every screenshot, repost, and platform recompression washes out the statistical cues detectors rely on.

The practical consequence: treat any single answer — human or machine — as provisional. A score plus provenance plus context gives you a defensible conclusion. A glance alone no longer does, and neither does a detector score on a twice-compressed thumbnail.

Zoom to full resolution and check the areas generators still find hard. Hands and fingers can merge or gain joints. Teeth and ears may blur into shapes that almost make sense. Glasses produce reflections that do not match the room. Text inside the scene — signs, labels, logos — may look right at a glance but collapse into gibberish when you read it letter by letter.

Then check the physics. Shadows falling in different directions, reflections that show the wrong scene, objects that float or intersect, liquids frozen mid-splash, fabric that behaves like plastic, and backgrounds that dissolve into soft mush are all worth pausing on. Repeated textures — the same brick, leaf, or pattern stamped across a frame — are another common tell.

One caveat: newer models fix many of these artifacts, especially at social-media sizes. The absence of visible artifacts is not proof of authenticity. Treat these signs as reasons to run a scored check, not as a substitute for one.

Camera captures carry a physical fingerprint: sensor noise that stays consistent across the frame, natural motion blur where things move, lens vignetting, and small optical imperfections. Real scenes also tend to be messy in believable ways — cables that tangle, dirt that collects, wear that matches use — and lighting that obeys a single physical source.

Provenance beats pixel inspection when it exists. An original camera file with intact EXIF, Content Credentials (C2PA), a burst of adjacent frames, or a second photo of the same scene from another angle all support a “real” conclusion far more strongly than any single-image score.

Context completes the picture. Does the poster’s history make sense? Do other accounts place the same event at the same time and place? Does the story survive a reverse image search? None of these alone is proof, but together they separate most real images from most fabrications.

Most images you will want to check arrive as screenshots, and screenshots are the hardest case: recompression and cropping erase part of the signal. The check is still worth running, but if the result feels ambiguous, ask the sender for the original file before drawing conclusions.

Synthetic avatars follow a pattern: a brand-new account, a polished face that flatters the profile’s purpose, and rapid engagement with strangers. If a profile picture matters to your decision — dating, hiring, marketplace trades — run the portrait check and read our deepfake detector guidance for face-focused next steps.

For viral images, find the earliest version. Reverse image search surfaces prior appearances, and AI-generated images often trace back to generator galleries, prompt-sharing communities, or AI stock sites. When the earliest copy is a video frame or a camera original from a credible account, you have your answer; when it is a fresh upload with no history, keep digging.

Step one: score the file with the checker above. Step two: verify the source — who posted it first, what else they post, and whether they answer questions about context. Step three: find one independent signal, such as a reverse image search result, a corroborating post, or provenance credentials. Step four: decide, and write down why if the decision affects anyone else.

Teams can turn the same workflow into policy. Auto-accept only when the score is low and the stakes are low. Route mid-range scores to human review, because hybrids and recompression live there. Hold high scores until the source checks out. The detector is the first gate in that pipeline, not the final word.

Re-run checks with current tools when a decision is challenged. Detection improves as generators change, and a score from an older engine is not a permanent label. Keeping the original file — not a screenshot — makes any later re-check meaningful.