Whether an AI-generated image can be identified can be judged in three steps: first zoom in and look for flaws in hands, text, lighting, and details, then find the original and check its file information, and finally use official verification tools to check Content Credentials or watermarks; if you still cannot confirm after all three steps, do not treat it as a real photograph.
Step 1: Zoom in on hands, text, and details
Current generative models rarely show obvious six fingers or extra fingers, but text remains a weak point: letters on signs, posters, and book spines are easily misspelled, distorted, or turned into meaningless symbols. Also check whether the lighting is consistent, and whether the edges of people, reflections in glasses, passers-by in the background, and the lines of tables and chairs suddenly break or melt together. Thumbnails easily hide these problems; checking at original size, section by section, is what makes the inspection meaningful. Normal-looking details do not mean a photo is real, however: the better the generation quality, the fewer clues this step provides.
Step 2: Check the original, not just a screenshot
Screenshots, repeated forwarding, and recompression strip away the information carried in a file, so try to obtain the original whenever possible. When checking file information, look at whether shooting settings, device model, and creation time are complete and whether they contradict each other; a picture claimed to be a live scene that carries no shooting information at all, or information pointing to a generative tool, should raise suspicion. At the same time, use reverse image search to find where the image first appeared, who published it, and whether there is an original source or shooting record. An image with no identifiable origin that keeps being reposted should be treated as unverified content, however convincing it looks.
Step 3: Verify credentials and watermarks with official checking tools
Content Credentials are provenance records carried with a file that can show which tool generated an image or what edits it went through, and they can be viewed on the official Content Credentials verification page by uploading the original file. Google SynthID is an invisible watermark hidden in the pixels that survives screenshots and compression, and related tools can confirm whether an image carries such a watermark. OpenAI also provides a content verification tool for checking whether an image carries provenance signals it supports. Finding a credential or watermark confirms the generation source; finding nothing only means that tool found no signal it supports, not that the image is a real photograph.
No result does not mean the photo is real
No method can determine with one hundred percent certainty whether an image was made by AI. Credentials and watermarks can be lost during editing or resaving, not every generative tool writes such signals, and visual details become harder to tell apart as models improve. The safer approach is to weigh several pieces of evidence together: details, original-file information, verification results, and the publishing source must corroborate each other before the image deserves trust. If key evidence is missing or contradictory, treat the image as unverified, do not forward it or use it as a news picture, and ask the publisher for the original or proof of shooting if needed.
