Jawline and hairline seams
Swapped faces are composited onto a host head, so the boundary at the jaw, ears and hairline often carries a soft edge, a mismatched noise floor, or hair strands that terminate unnaturally.
Face forensics
AI face detector
Upload a portrait, profile picture or video still and DetectFake assesses whether the face was generated by a model or swapped onto someone else. You get a verdict, a confidence score, and the specific facial signals behind it — not an unexplained percentage.
Check a face photo 3 free scans every day
Facial signals we report
Skin texture and pores
Real skin has irregular pores, fine hair and uneven specular highlights. Generated faces frequently show a plasticky, evenly distributed texture, especially across the cheeks and forehead.
Eyes and catchlights
Reflections in both eyes should agree with each other and with the visible light sources. Generated faces commonly show mismatched catchlights, irregular pupils, or irises with repeating detail.
Mouth interior
Teeth and tongue remain a weakness for generative models: merged or miscounted teeth, blurred gum lines and inconsistent inner-lip shadowing are strong indicators.
Accessories and symmetry
Earrings that differ between ears, glasses whose frames break behind the temple, and collars or necklaces that lose continuity are typical of synthesized or inpainted faces.
Face versus scene lighting
The key light on the face should match the shoulders and background, with a consistent contact shadow under the chin. A face lit differently from its scene suggests the head was composited in.
How to check a face in four steps
- 1
Get the best copy available: save the original upload rather than a screenshot or forwarded copy, since re-encoding removes forensic detail.
- 2
Run the scan: upload the image (or a video, which is sampled into frames locally) and read the per-signal breakdown.
- 3
Zoom in yourself on the signals the report flags — jawline, hairline, eyes, teeth and accessories are where synthesis shows first.
- 4
Weigh the context: who posted it, whether the same face appears elsewhere, and whether a mid-range confidence means the result is simply inconclusive.
What this tool is not
DetectFake does not perform facial recognition, reverse face search or identity matching, and it never tries to name the person in an image. It also cannot certify a face as genuine: heavy retouching, beauty filters, low resolution and aggressive compression all reduce reliability in both directions. Read our methodology and limitations before relying on a verdict for a consequential decision.
AI face detector FAQs
What does an AI face detector actually check?
DetectFake examines the face region for the artifacts generative models and face-swap pipelines leave behind: a soft or shifting seam along the jawline and hairline, skin texture that is too uniform for real pores, eye reflections that disagree with the scene, teeth that merge or change shape, and accessories such as earrings or glasses that are geometrically inconsistent. Each signal is reported separately alongside the confidence score.
Can it tell a generated person from a real one?
Often, yes. Fully synthetic portraits from models like the StyleGAN family and current diffusion models tend to show tightly centered eye placement, smooth background collapse, symmetric lighting and dissolving hair strands. DetectFake reports these as signals rather than proof, because an authentic studio portrait with heavy retouching can share some of the same characteristics.
Does it identify who is in the photo?
No. DetectFake is not a facial-recognition or face-search tool. It does not match faces against any identity database and does not attempt to name the person in the image. It only assesses whether the pixels look generated or manipulated.
Is the AI face detector free?
Three scans per day are free and include the complete report — verdict, confidence and every individual signal. Silver raises the limit to 30 scans a day with deeper video frame sampling, and Gold is unlimited.
Why does a screenshot give a weaker result?
Screenshots and images forwarded through messaging apps are re-encoded, which erases the fine sensor and compression detail facial forensics depend on. Always analyze the highest-quality original file you can obtain, and treat mid-range confidence on a compressed screenshot as inconclusive.