There is no reliable single test for whether an image is AI-generated. There are several partial tests, and they differ in how much a result means. A valid Content Credential from a known generator is close to proof. A detector reporting "87% AI" is closer to an opinion. The procedure below runs the tests roughly from strongest to weakest, and at each step it matters which direction the evidence runs. Most of these checks can show that an image is synthetic. Almost none can show that it is not.
Step 1: Check for Content Credentials
Many large generators attach a C2PA manifest to their output. These include Adobe Firefly, OpenAI's image tools, and Google's image models. The manifest can record that the content was AI-generated. Upload the original file to Content Credentials Verify or Adobe's Inspect tool, or run c2patool image.jpg locally (see verification tools). On LinkedIn, a small "cr" pin on an image opens the same information.
How to read the result:
- Valid manifest with a generative-AI action or source type. Strong evidence the image was generated or edited with AI, as declared by the signer. Check who signed it and whether the certificate is on the C2PA trust list.
- Valid manifest from a camera, no AI actions. Evidence that a device signed the capture. This is meaningful but not absolute; see what provenance cannot prove.
- Invalid manifest. The file changed after signing. That is often harmless, but it means the manifest's claims no longer describe these exact bytes.
- No manifest. This tells you almost nothing. Most images online have none, because most social platforms remove the manifest on upload and a screenshot drops it entirely.
Adobe's Inspect tool also looks for durable Content Credentials: a watermark in the pixels that points to a manifest stored in Adobe's cloud. That lets it recover credentials from some stripped copies.
Step 2: Check for a SynthID watermark
Google embeds SynthID, an invisible watermark, in the pixels of images made or edited with its AI tools. Because it sits in the pixels, it usually survives metadata stripping, compression, and screenshots. Google itself says it is not foolproof against extreme manipulation.
The documented way to check is the Gemini app. Upload the image at gemini.google.com (while signed in) or in the Android app, and ask "Was this created or edited by Google AI?" Gemini checks for SynthID and for Content Credentials. Google's help page lists limits: one file at a time, about ten image checks in a rolling 24-hour window, and, importantly, "Gemini can currently only recognize content created by Google AI tools." In May 2026 Google announced that SynthID checks were rolling out to Search and Chrome, with Content Credentials checks to follow. The same announcement said OpenAI, Kakao, and ElevenLabs would adopt SynthID. Availability varies by country and product, so if you do not see a check in Search or Chrome, use the Gemini app.
A SynthID detection is strong evidence that all or part of the image came from a participating AI system. A negative result rules out only watermarked output from participating generators. It says nothing about open-source models, services that do not use SynthID, or images whose watermark was destroyed by heavy editing.
Step 3: Read the metadata
Even without a signed manifest, many files contain plain metadata. The key field is the IPTC Digital Source Type. The value trainedAlgorithmicMedia means "created using Generative AI," and compositeWithTrainedAlgorithmicMedia means "edited using Generative AI," for example by inpainting. Meta's platforms read these fields when applying their AI labels. With ExifTool:
# Digital source type (XMP IPTC Extension)
exiftool -XMP-iptcExt:DigitalSourceType image.jpg
# Everything, grouped, to look for generator names in Software,
# Creator Tool, or free-text fields
exiftool -a -G1 -s image.jpgGenerator tools sometimes also write their name or the prompt into XMP or PNG text chunks. Presence is informative. Absence is not, because plain metadata is unsigned, trivially removed, and just as easily added. Treat an AI source type as a strong hint and a camera make and model as a weak one. The metadata analysis page covers the inconsistencies worth looking for.
Step 4: Reverse image search
This is often the most useful step in practice, because it answers a different question: where the image came from. Run the image through Google Lens, TinEye (sort by oldest), Bing Visual Search, and Yandex. See reverse image search workflows. Look for:
- An earlier copy posted by an AI artist, a generator's community gallery, or a stock library's AI section.
- An earlier copy with more context, such as a watermark, caption, or higher resolution that still carries its credentials.
- Near-duplicates with small differences, which can indicate several generations from one prompt.
- The real photo that a synthetic or edited version was derived from.
No matches is common for new AI images and for new genuine photos alike, so it does not point either way.
Step 5: Look for visual tells
Close inspection still finds errors in some generated images, but the list gets shorter with every model release. The visual indicators page keeps a current field guide. Useful areas to check in 2026:
- Text and symbols: signage, labels, licence plates, and logos that are almost right.
- Physical consistency: reflections, shadows that disagree with the light source, and objects that merge into each other.
- Background coherence: crowds, railings, windows, and repeating structures that lose their logic away from the subject.
- Context: uniforms, insignia, weather, and architecture that do not match the claimed time and place.
Treat tells as a reason to look harder, not as proof. Current models avoid many of the classic errors, such as extra fingers. Real photos also contain oddities: compression artifacts, motion blur, lens distortion, and phone computational processing can all look "wrong." Checking context against independent sources is often more reliable than judging pixels.
Step 6: Use a detector, and weigh it carefully
Classifier-based detectors, from commercial services such as Hive to open-source research models, output a probability that an image is synthetic. They are useful as one input, and they are the only option when every other check is empty. Their error rates on real-world content are much worse than their marketing suggests.
The clearest public evidence comes from Deepfake-Eval-2024, a benchmark built from deepfakes actually circulating on social media in 2024. Compared with earlier academic benchmarks, the AUC of state-of-the-art open-source image detectors dropped by about 45%, and many off-the-shelf models scored close to 0.5, which is chance level. Fine-tuned commercial systems did better but still lagged behind human forensic analysts. The AI image detection page explains why benchmark accuracy fails to carry over to real-world images: new generators, re-compression, resizing, and screenshots all move images away from the training data.
Rules for using detectors:
- Run the best-quality copy you can find. Compression and screenshots degrade results in both directions.
- Use more than one detector, and treat disagreement as "unknown."
- Never publish or act on a detector score alone. False positives on real photographs are well documented and can cause real harm to real people.
- Remember the base rate. If most images in your stream are genuine, even a detector with a low false-positive rate will produce many false alarms.
Putting it together
| Check | Positive result means | Negative result means |
|---|---|---|
| Content Credentials | Strong: signer declared AI use | Almost nothing |
| SynthID | Strong: participating generator | Not from a participating generator, or watermark destroyed |
| IPTC source type | Moderate: unsigned and easily edited | Almost nothing |
| Reverse search | Often decisive: origin found | Nothing |
| Visual tells | Weak to moderate | Weak: modern models avoid many tells |
| Detector score | Weak to moderate, depending on corroboration | Weak |
A sound conclusion combines independent signals. Examples: "SynthID detected and the earliest copy is from a generator gallery," or "signed camera capture from a trusted certificate, consistent with independent reporting from the scene." When the signals are weak or conflict, the correct answer is "undetermined." That is a legitimate result, not a failure. The full verification workflow puts these steps into a broader procedure, and newsroom verification workflows shows how professional desks record and defend the result.