The signals you can stack
Filename hints (mj_, sdxl_, dalle_), embedded C2PA manifests, unusual dimensions (1024×1024, 1536×1536), and known vendor strings in the file bytes all push the needle. None of them alone is proof.
Visual cues that still help
Garbled text, extra fingers, asymmetric jewelry, and impossible reflections remain common giveaways in 2026, though top-tier models clean these up most of the time.
Why model-based detectors are unreliable alone
Classifier models that judge 'AI vs real' are easily fooled by light editing, compression, and out-of-distribution photos. Use them as one signal among many, never as the final word.
A pragmatic workflow
Check AI Watermark gives you the metadata layer in seconds. For high-stakes decisions, combine it with a reverse image search, a model-based detector, and a careful human look at the image itself.
Check AI Watermark
Upload an image to scan for AI watermarks, provenance signatures, and metadata clues.
Open Check AI WatermarkFrequently asked questions
More from this guide
What Is an AI Watermark?
Invisible signals embedded in AI-generated images, and what they actually tell you.
Read articleHow AI Watermarking Works
From pixel-domain perturbations to C2PA manifests — how invisible signals get embedded and read.
Read articleWhy AI Watermarks Exist
Provenance, accountability, and the policy pressure pushing every AI vendor to mark its output.
Read articleAI Watermarks vs Metadata
Embedded pixel signals and EXIF/C2PA manifests solve overlapping problems in different ways.
Read article