How to tell if an image is GPT Image 2 (OpenAI)
GPT Image 2 is OpenAI's image model, released on 21 April 2026 and available in ChatGPT, the OpenAI API and Codex. It is harder to spot by eye than most generators because it renders long passages of text, including non-Latin scripts, and produces photographic scenes with no sensor noise and no camera EXIF. The most dependable signal is provenance rather than pixels: OpenAI signs its image output with C2PA Content Credentials and, since 19 May 2026, also embeds a SynthID invisible watermark. Both can be removed, so this page covers what to look for when they are gone.
Visual tells
Fine speckle or grain laid over the whole frame. Users on OpenAI's own developer forum have reported this consistently since launch, describing it variously as pointillism, JPEG-style compression noise, or coloured specks scattered across the image. It is easiest to see in flat regions such as skies, plain walls and studio backgrounds, where a real photograph at low ISO would be smooth.
Diagonal grid or tiling patterns in textured areas. Reports on the OpenAI developer forum describe faint diagonal grid artefacts appearing across surfaces, and light and dark spot patterns that follow the shape of an object rather than the lighting. Fine repeating microstructure (foliage, gravel, fabric weave, glass) is where it shows up most.
Text that is almost entirely correct, with one localised failure. GPT Image 2 renders short signs, labels and multilingual captions cleanly, so the tell is not garbled text everywhere. It is a poster or shopfront where most of the type is sharp and correct but one long string, one multi-line block, or one piece of stylised lettering breaks down.
Hands and articulated poses still fail on the hard cases. Simple hands are usually fine. Errors concentrate in two hands interacting with an object, or a specific grip such as holding a pen, where finger count and joint geometry go wrong while the rest of the image stays photographic.
Neutral colour, which is the opposite of the older OpenAI tell. GPT Image 1 and 1.5 had a well-known warm yellow cast. GPT Image 2 renders neutral whites and keeps cool light sources cool, so the absence of a yellow tint does not rule out an OpenAI image and should not be treated as evidence the image is real.
Near-identical images across a set. Users report that within a single ChatGPT conversation the model carries texture and noise over between generations, so repeated prompts return images sharing the same speckle pattern or looking nearly the same. A batch of images that share an identical background texture is worth a second look.
Metadata & provenance
C2PA Content Credentials: yes. OpenAI is a C2PA conforming generator and attaches a cryptographically signed manifest to images from ChatGPT, the OpenAI API and Codex. The manifest records the generating tool and a timestamp and is signed by OpenAI, so tampering breaks the signature. You can read it with exiftool, the open-source c2patool, contentcredentials.org/verify, or OpenAI's own tool at openai.com/verify. SynthID invisible watermark: yes, from 19 May 2026. On that date OpenAI announced it had integrated Google DeepMind's SynthID watermark into images produced by ChatGPT, the OpenAI API and Codex, alongside the existing C2PA metadata. Because SynthID is embedded in the pixels at generation time rather than attached as metadata, it survives re-encoding, resizing, cropping and screenshots that destroy C2PA. OpenAI's public Verify tool checks for both signals. Output resolutions and formats: the API accepts any size where both edges are multiples of 16, the longest edge is 3840px or less, the aspect ratio is no wider than 3:1, and the total pixel count falls between 655,360 and 8,294,400. OpenAI's documented examples are 1024x1024, 1536x1024, 1024x1536, 2048x2048, 2048x1152, 3840x2160 and 2160x3840, plus an `auto` default. Output format is PNG by default, with JPEG and WebP also available. A study of 10,217 GPT Image 2 images collected from X in the six days after launch found that where native dimensions survived, they were 1024x1024, 1536x1024 or 1024x1536. Metadata gets stripped in normal use, and this is not a hypothetical. The same study found that X's CDN removed all embedded metadata on upload, leaving every image as a bare JPEG with no EXIF, no XMP and no C2PA markers. Screenshots, re-saves, format conversion and most social platforms have the same effect. An image with no Content Credentials is not evidence that it is real.
How well WeDetect handles GPT Image 2 (OpenAI)
We have not measured wedetect's recall on GPT Image 2 specifically, and this page does not claim a figure for it. The figures we have published are roughly 89% overall on modern generators and roughly 85% recall on Midjourney, DALL-E, SDXL and Flux. On Stable Diffusion 3.5 our measured recall is roughly 62%, which is well below the rest and we would rather say so than bury it. Every visual verdict is probabilistic and can be wrong in both directions. Independent work suggests GPT Image 2 is a hard case for pixel-only detectors generally. A 2026 benchmark of five published detectors on text-rich GPT Image 2 images (posters, infographics, receipts, tables, UI screenshots) found the best of them reached 80.45% accuracy, with per-category F1 ranging from 13.58% on tables to 99.24% on commercial posters, and F1 collapsing from 0.83 to 0.34 once the images were saved as JPEG at quality 95. wedetect was not one of the systems tested, so those numbers are context rather than a measurement of us. Practical advice for the page: run the original file, not a screenshot or a re-download. If the C2PA manifest or SynthID watermark is intact, that is a far stronger answer than any visual model can give. If provenance has been stripped, treat the visual result as an indication and say so rather than presenting it as proof.
Check an image now
Drop the image in below — free, no sign-up, deleted after analysis.
Drag & drop, paste, or click to upload
PNG · JPG · WEBP · HEIC — your image is analyzed privately and deleted
This is an estimate, not proof — the detector can be wrong, producing both false positives and false negatives. Don't rely on it as your only source; verify anything important. wedetect.ai is not liable for decisions made from these results. See our Terms.
Frequently asked questions
Does ChatGPT put a watermark on images?
Yes, two kinds. Every image includes C2PA Content Credentials, a signed metadata block naming the tool that made it. Since 19 May 2026 images from ChatGPT, the OpenAI API and Codex also carry a SynthID watermark embedded in the pixels themselves. There is no visible logo or corner mark. The SynthID layer is designed to survive screenshots and re-compression, while the metadata is not.
How can I tell if an image was made with GPT Image 2?
Check provenance first. Upload the original file to a Content Credentials reader or to OpenAI's verify tool and see whether a signed OpenAI manifest is present. If the file has been screenshotted or passed through a social platform, that manifest is usually gone, and you are left with visual judgement: look for fine speckle over flat areas, faint diagonal grid patterns in textures, and a single garbled block of text in an image where the rest of the type is clean. A visual verdict is probabilistic, not proof.
Why does my downloaded ChatGPT image have no metadata?
Most likely something in the chain removed it. A study of GPT Image 2 images posted to X found the platform's CDN stripped all EXIF, XMP and C2PA data on upload, leaving plain JPEGs. Screenshots, format conversion, resizing and many messaging apps do the same. Missing Content Credentials tells you the file has been processed, not that it is a real photograph.
Can GPT Image 2 images be detected after a screenshot?
Partly. A screenshot removes the C2PA manifest, so that route is closed. The SynthID watermark is meant to survive screenshots and can still be checked through OpenAI's verification tool. Our own visual model still runs, but it is working without provenance at that point, and we report the result as inconclusive rather than guessing when the evidence is thin.
Other generators: Midjourney · DALL·E 3 · GPT-4o · Stable Diffusion · Flux · Grok · Firefly · Gemini · SDXL · Nano Banana · Flux 2 · SD 3.5 · Sora · all generators.