On this page
What ChatGPT puts in its images
ChatGPT makes pictures with OpenAI's GPT Image models. The same models run in OpenAI's API and in apps built on it, such as Canva and Adobe Firefly, so a "ChatGPT image" may never have been near ChatGPT itself. At the source, OpenAI adds three signals:
- Content Credentials (C2PA): a signed record in the file's metadata saying that an OpenAI tool made the image.
- OpenAI's own invisible watermark, hidden in the pixels.
- SynthID, Google's invisible watermark, which OpenAI now adds to ChatGPT images as well.
Don't expect a visible mark in a corner. The signals that matter sit in the file and in the pixels, and the first one is the easiest to lose.
Step 1: check for OpenAI's label
Get the best copy you can
Ask for the original file, or save the image itself rather than a screenshot of it. Every screenshot, crop and re-save can drop the label.
Read the Content Credentials
Drop the file into our Content Credentials checker. A valid record signed by OpenAI is strong evidence that ChatGPT or OpenAI's API made the image.
Try OpenAI's own checker
OpenAI offers a verification page for images from ChatGPT, its API and Codex. It looks only for OpenAI's own signals, so a "no" doesn't mean the image is real: another company's model may have made it.
If the label is there, you are mostly done. Usually it isn't. X and most social and messaging apps strip metadata on upload: in our sample of GPT Image 2 images collected from X, none still carried Content Credentials.
Step 2: look at the picture
GPT Image is good at photos and very good at text, so most of the old giveaways no longer apply. What still helps, and what doesn't:
Still worth checking
- Text that is perfect where real life is messy: every label on a shelf legible, a receipt with crisp, evenly aligned print, a handwritten note with no slips.
- A scene that is right in every detail but wrong as a whole: the perfect moment that no one else seems to have photographed.
- Light that flatters everything at once, with none of the harsh shadows, blown highlights or blur a phone usually produces.
- Small things that make no sense up close: jewellery that melts into skin, a zip that goes nowhere, cables that don't connect.
No longer reliable
- Hands and fingers: GPT Image gets them right most of the time.
- Garbled lettering: ChatGPT now writes long, correct text inside images.
- A warm, yellowish tint: a quirk people associated with earlier versions. Many ChatGPT images don't have it, and plenty of real photos do.
- Asking ChatGPT itself: it can describe a picture, but it isn't built to tell you whether a picture is AI, and it can be wrong either way.
Why a ChatGPT image can pass for a phone photo
People rarely ask ChatGPT for a glossy render any more. They ask for "a casual phone photo", "slightly blurry, taken with flash", or "like an amateur snapshot from 2015", and GPT Image is good at exactly that. The result has the noise, the slightly crooked framing and the cluttered background of a real phone picture.
That is why the useful questions are about context rather than polish. Who sent it? Was there a reason to make it up? Does it show something that should have left other traces, such as a receipt that should appear on a card statement, damage that should appear in a delivery record, or an event that other people should also have photographed?
Where ChatGPT images turn up most often, and the guide for each:
- Profile pictures and invented people: is this profile picture AI-generated?
- Receipts and invoices: fake receipts made with AI
- Damage photos for refunds and claims: AI photos in refund and insurance claims
- Viral pictures of events that never happened: check a photo before you share it
Step 3: trace where it came from
Most AI images that cause trouble are passed along by people who didn't make them. Finding the first post often answers the question faster than any close look:
- Run a reverse image search and sort the results by date.
- Look at the account that posted it first: does it post other AI images, or say that it uses AI?
- Read the replies and community notes: someone may already have found the prompt or the person who made it.
- If the picture shows a real event, look for other photos of the same moment taken by other people.
Step 4: check the pixels with Perspicax
Perspicax doesn't need a label. Proven computer vision techniques and TrueEye 2.3, our state-of-the-art detector, examine every region of the image, so it works on screenshots and re-saved copies as well as on originals.
Measured the way Perspicax answers: an inconclusive answer counts as a miss. See every number
In testing, it caught 97.4–98.8% of ChatGPT images across GPT Image 2 and GPT Image 2.5: 98.8% for GPT Image 2.5 Sunburst, 98.8% for GPT Image 2.5 Flare and 97.4% for GPT Image 2. On 6,811 GPT Image 2 images posted on X, where the platform had stripped their labels, it caught 89.7%.
Three possible answers
It tells you whether an image was fully generated by AI, not which model made it. For the model-by-model results, see our ChatGPT image detector.
Posters, receipts and screenshots
ChatGPT is often used for posters, menus, slides and receipts, not just photos. Perspicax is tuned for photo-like images. When you upload a screenshot, it crops to the photo inside it; posters, documents and artwork still get an answer, marked as a weaker signal.
For a document such as a receipt or an invoice, the paper trail matters more than the pixels. Our guide on fake receipts made with AI covers the checks finance teams use.
Check an image now
Drop, paste or link a photo. Perspicax checks every region of it with computer vision and TrueEye 2.3, and answers likely AI-generated, likely real or inconclusive.
Questions people ask
Can ChatGPT tell me if an image is AI?
It can describe a picture and guess, but it isn't built for detection and can be wrong either way. OpenAI's separate verification page checks only for OpenAI's own signals. A detector built for the job checks the pixels of any image.
Do ChatGPT images have a watermark?
Yes, invisible ones: OpenAI adds its own watermark and Google's SynthID, plus Content Credentials in the file. Don't count on a visible mark.
Does a screenshot remove ChatGPT's label?
It removes the Content Credentials, which live in the file's metadata. Invisible watermarks are designed to survive more, but you need the right checker to read them. Perspicax checks the pixels either way.
Can Perspicax tell if an image came from ChatGPT or from Gemini?
No. It tells you whether an image was fully generated by AI, not which model made it. When a label survives, it can name the tool; our guide on which AI made an image covers the other clues.
Is DALL·E the same as ChatGPT images?
DALL·E was OpenAI's earlier image model. ChatGPT now uses the GPT Image models. Older DALL·E pictures still circulate, and Perspicax checks any photo-like image, whichever model made it.
How accurate is Perspicax on ChatGPT images?
In testing it caught 97.4–98.8% of ChatGPT images across GPT Image 2 and GPT Image 2.5, and 89.7% of GPT Image 2 images as posted on X. Inconclusive answers count as misses in these figures.