How to Check If an Image Is AI Generated (2026 Guide)

OpenAI published a post called “Advancing content provenance for a safer, more transparent AI ecosystem”. It announced something the whole industry had been circling for years: OpenAI joined the C2PA Steering Committee, started adding Google’s SynthID watermark to its generated images, previewed a public tool that tells you whether an image came from OpenAI, and opened access to a DALL-E 3 image-detection classifier through its Researcher Access Program. That single post is a useful marker. The question “is this image real?” is no longer a niche worry for fact-checkers. It is now a default check that journalists, recruiters, dating-app users, insurance adjusters, and ordinary people run on a regular basis.

TL;DR

To check if an image is AI-generated, work down a reliability ladder: first inspect its C2PA Content Credentials (cryptographic provenance metadata), then scan it for invisible watermarks like Google SynthID, then run it through a machine-learning detection tool, then look for visual tells by eye, and finally try a reverse image search. No single method is conclusive; provenance metadata is the strongest signal when present, but it is easily stripped, so you combine methods and weigh the evidence.

The fastest practical workflow:

1. Get the original file. Screenshots and re-saved copies destroy the most useful evidence. Ask for the source file whenever you can.
2. Check C2PA Content Credentials. Upload the file to a Content Credentials verifier and read the provenance manifest, if one exists.
3. Scan for invisible watermarks. Run the image through Google’s SynthID Detector to see if it carries a Google or OpenAI watermark.
4. Run an ML detection tool. Use a classifier-based detector or API for a probability score when no provenance or watermark is found.
5. Inspect the image manually. Look at hands, text, jewelry, teeth, reflections, and lighting, knowing these tells fail on top-tier models.
6. Reverse image search. Trace the image’s history online to see where and when it first appeared.
7. Weigh the evidence. Treat the result as a confidence level, not a verdict.

Why this is hard, and why you need a method stack

Two years ago, you could spot most AI images in a second. Hands had six fingers. Text in the background was alien gibberish. Eyeglasses melted into faces. That era is over for the leading models. Current image generators produce photographs that pass casual inspection and often survive expert inspection. A widely cited statistic illustrates the gap: in research on high-quality synthetic media, human subjects correctly identified deepfake videos only about a quarter of the time. Your eyes are not a reliable detector anymore.

So the field moved toward two better ideas. The first is provenance: instead of trying to detect fakery after the fact, attach a signed record of origin to the file at creation time. The second is watermarking: hide a statistical signal inside the pixels that a matching decoder can later read. Both are proactive. Both depend on the generator cooperating. And both can be defeated by someone determined to strip the evidence.

That is the core reason you need a stack rather than a single tool. Provenance is the strongest signal, but it is opt-in and fragile. Watermarks are durable but model-specific. Classifiers work on any image but are probabilistic and produce false positives. Visual inspection is universal but unreliable on good models. Each method covers a weakness in the others. Run them in order, and stop when you have a clear answer; if you never get one, that uncertainty is itself the finding.

Method 1: Check C2PA Content Credentials (most reliable)

C2PA stands for the Coalition for Content Provenance and Authenticity. It is an open technical standard, backed by Adobe, Microsoft, Google, OpenAI, the BBC, and major camera makers, that attaches a tamper-evident, cryptographically signed block of metadata to a media file. That block is called a manifest. The user-facing brand name for this is Content Credentials.

When a tool that supports C2PA creates or edits an image, it writes a manifest describing what happened: which software produced the file, when, and whether AI was involved. The manifest is cryptographically signed. If anyone alters the image afterward without re-signing, verification fails. OpenAI has attached C2PA Content Credentials to DALL-E 3 images since 2024, and its May 2026 announcement confirmed it is now a C2PA Conforming Generator, meaning other platforms can read, preserve, and pass along that provenance data.

How to check it

You do not need to be technical. Use a free, in-browser verifier:

1. Get the original image file. This matters more here than anywhere else.
2. Open a Content Credentials inspector. The official one lives at contentcredentials.org; other free C2PA viewers exist.
3. Upload or drag the file in. Processing happens in your browser; the file is not sent to a server.
4. Read the result. You will see one of three outcomes: a valid manifest with origin details, no Content Credentials data at all, or an invalid or tampered manifest.

A valid manifest can tell you the image was generated by a specific AI tool, or that it came from a particular camera and was edited in named software. That is about as close to ground truth as image checking gets.

The big caveat

C2PA only helps when the credential is present and intact. Here is what destroys it:

• Screenshots. Screenshotting an image creates a brand-new file with no manifest.
• Re-encoding. Many “save as” operations and format conversions drop the metadata.
• Social platforms. Several platforms strip or do not preserve C2PA metadata on upload, though this is slowly improving.
• Deliberate removal. Anyone who wants to hide an image’s origin can strip the manifest in seconds.

There is a subtler limitation worth stating plainly: C2PA verifies the integrity of the manifest, not the truth of the content. A staged or misleading photo can carry a perfectly valid signature if nobody altered it after signing. So treat a valid credential as strong evidence of origin and edit history, not as proof that the image depicts reality. If you find no Content Credentials, that is not evidence the image is fake or real; it just means this method gave you nothing, and you move to the next one.

Method 2: Detect invisible watermarks like SynthID

If provenance metadata is the receipt stapled to the outside of the package, an invisible watermark is a pattern woven into the fabric itself. SynthID, developed by Google DeepMind, embeds an imperceptible signal directly into the pixels of images generated by Google’s Gemini and Imagen models. The change is invisible to you but detectable by a matching decoder.

The advantage over metadata is durability. Because the watermark lives in the pixels, it survives many transformations that wipe C2PA metadata: screenshots, cropping, compression, color adjustments, and re-saving. According to Google, SynthID has been applied to billions of pieces of AI content across images, audio, video, and text. And as of the May 2026 announcement, OpenAI is adding SynthID watermarking to its own generated images, which means a single detector now covers content from two of the largest model providers.

How to check it

Google runs a public SynthID Detector portal. Upload an image, and it scans for the watermark and reports whether one is present, often highlighting the regions most likely to carry it. You can read the background on the technology at Google DeepMind’s SynthID page.

The caveats

Watermark detection is powerful but narrow:

• It is model-specific. SynthID detection tells you whether an image came from a SynthID-watermarking model. An image from Midjourney, Stable Diffusion, or a model that does not watermark will simply come back clean. Clean does not mean human-made.
• Coverage is partial. Watermarking is opt-in per provider. Open-source models can be run without it entirely.
• Heavy edits can still degrade it. SynthID is built to survive common transformations, but aggressive manipulation, regeneration, or extreme compression can weaken the signal.

A positive SynthID result is a strong, hard-to-fake “yes, this is AI.” A negative result is close to meaningless on its own. That asymmetry is the key thing to remember.

Method 3: Use an ML detection classifier or API

When an image carries no provenance metadata and no detectable watermark, which describes most images in the wild, you fall back to detection classifiers. These are machine-learning models trained on large sets of real and synthetic images. They learn statistical fingerprints that AI generators leave behind, like the absence of natural sensor noise, frequency-domain artifacts, or texture patterns invisible to the eye. You upload an image and get a probability: “87% likely AI-generated.”

Several commercial and free tools work this way, and many also try to name the model that produced an image. Independent 2026 benchmarking found leading detectors landing in roughly the 90s for accuracy on standard test sets, with real-world performance varying by source model, compression, and editing. OpenAI’s own DALL-E 3 detection classifier, opened to researchers in May 2026, is a specialized example of this category.

For developers, the practical move is to call a detection API from your own application. If you are evaluating providers, our roundup of the best AI image detection APIs for developers compares accuracy, pricing, and supported models, and if you would rather own the pipeline, this walkthrough on how to build your own AI image detector API covers training and serving a classifier yourself. Either way, you will want to send real test images through the endpoint and confirm the responses before relying on them. That request-and-verify loop is the everyday job of an API client; you can script and replay those calls in Apidog with saved environments so your detection check behaves the same in development and production. If your detection logic is wired into an AI agent, the Apidog AI agent debugger helps you trace exactly what the model sent and received.

The caveats

Classifiers are the most broadly applicable method, and the least certain:

• False positives are real. Detectors sometimes flag genuine photographs as AI, especially heavily edited photos, screenshots, low-light shots, or images with smooth, low-texture surfaces. This has caused real harm to artists and students wrongly accused.
• They lag new models. A classifier trained before a new generator launches has not learned that model’s fingerprint. Detection accuracy on brand-new models is often much lower than the headline number.
• Adversarial edits defeat them. Light noise, recompression, or filters can push an image past a classifier’s decision boundary.
• The score is a probability, not a verdict. “73% AI” means the model is uncertain. Treat anything in the middle of the range as inconclusive.

We go deeper into these failure modes in a dedicated piece on why AI image detection fails. The short version: never make a high-stakes decision, like accusing someone of fraud or cheating, on a single classifier score.

Method 4: Manual visual inspection (use with caution)

This is the method everyone reaches for first, and it should be one of the last. On older or weaker models, and on outputs from people who did not cherry-pick the result, visual tells still work. Here is what to look at:

• Hands and fingers. Still the classic weak spot. Count fingers. Check for extra knuckles, fused digits, or hands that bend the wrong way.
• Text. Background signs, book spines, posters, and clothing logos often render as warped, almost-but-not-quite letters.
• Teeth and jewelry. Too many teeth, uneven sizes, earrings that do not match, necklace chains that merge into skin or change thickness.
• Joints and accessories. Watch straps, glasses arms, and belt buckles that connect illogically or vanish behind the body.
• Reflections and shadows. Mirrors, sunglasses, and water that do not reflect the scene correctly. Shadows falling in inconsistent directions.
• Lighting and physics. Multiple light sources that contradict each other; a perfectly lit subject against an impossibly blurred background.
• Backgrounds. Repeating textures, objects that blend into each other, architecture with impossible geometry.
• Output resolution. Many models emit images at a small set of fixed dimensions. A photo whose dimensions exactly match a known generator default is a soft hint, never proof.
• Skin and texture. An airbrushed, plasticky uniformity. Real skin has pores, asymmetry, and fine imperfection.

The honest caveat

Read this twice: manual inspection fails on current top-tier models. The leading 2026 generators produce correct hands, legible text, and consistent lighting most of the time. Anyone presenting a fake deliberately will discard the bad outputs and keep the flawless one. So visual inspection has two correct uses. First, it can quickly confirm AI generation when you do spot a clear error; an obvious six-fingered hand is a strong “yes.” Second, it builds intuition. But the absence of visible tells tells you almost nothing. A flawless image is exactly what a good generator, or a good camera, produces. Do not let “I can’t see anything wrong” become “therefore it is real.”

Method 5: Reverse image search

Reverse image search does not analyze pixels for AI fingerprints. It analyzes history. You submit the image to a search engine and see where else it appears online. That context can answer the AI question indirectly and sometimes decisively.

Use Google Images, TinEye, or a similar service. What you are looking for:

• A clear origin. If the image traces back to a stock library, a named photographer’s portfolio, or a news report from years ago, it is probably a real photograph that predates the current AI wave.
• An AI origin. If it traces to an AI-art community, a “made with Midjourney” gallery, or a prompt-sharing site, you have your answer.
• A suspicious profile. A “person” whose photo appears on dozens of unrelated profiles, or nowhere else at all, is a classic synthetic-identity signal. Many fake LinkedIn and dating profiles use AI faces precisely because they leave no search footprint.
• Context mismatch. An image presented as breaking news that actually surfaced years ago, or in a different country, is being misrepresented regardless of whether it is AI.

Reverse image search is most useful as a cross-check. It will not catch a freshly generated image that has never been posted before, since there is nothing to find. But for any image that has circulated, it adds a layer of evidence the pixel-based methods cannot.

Comparison: the five methods at a glance

No method stands alone. This table summarizes the trade-offs so you can pick the right starting point for your situation.

The pattern: methods 1 and 2 give you near-certainty when they fire, but they often return nothing. Methods 3 through 5 always give you something, but never certainty. Strong practice is to run 1 and 2 first for a possible quick win, then use 3, 4, and 5 together to build a weighted judgment.

How to combine the methods into one verdict

Put it together as a simple decision flow:

1. Original file in hand? If yes, check C2PA Content Credentials. A valid AI manifest is a near-definitive yes. A valid camera manifest with no AI assertion is strong evidence of a real photo. No credentials means continue.
2. Scan for SynthID. A positive hit is a strong yes. A negative result means continue; it rules nothing out.
3. Run a classifier. A very high score (above roughly 90%) plus other context is a confident yes. A very low score is a lean toward real. A middling score is inconclusive, so weight it lightly.
4. Inspect visually. A clear anatomical or text error is a confident yes. No visible errors changes nothing.
5. Reverse image search. Use it to confirm origin and catch misrepresented images.
6. Write down a confidence level, not a binary. “High confidence AI-generated, based on a positive SynthID hit and a 94% classifier score” is a defensible statement. “It’s fake” is not.

The mindset that keeps you out of trouble: you are assembling evidence, not flipping a switch. When the signals agree, you can be confident. When they conflict or all come back empty, the correct output is “undetermined,” and you should say so rather than guess.

Conclusion

Checking whether an image is AI-generated in 2026 is a process of weighing evidence, not running one test. Key takeaways:

• Work the stack in reliability order: C2PA Content Credentials, then invisible watermarks like SynthID, then ML classifiers, then visual inspection, then reverse image search.
• Provenance and watermarks give near-certainty when they fire but frequently return nothing, because metadata is stripped and watermarks are model-specific.
• Classifiers always return a score but never certainty; false positives are common, so never accuse anyone on a single number.
• Manual inspection fails on current top models. Use it to confirm obvious errors, never to declare an image real.
• The OpenAI May 2026 announcement matters because adding SynthID and C2PA conformance makes provenance more durable across two major providers, the direction the whole industry is heading.
• Always report a confidence level, and say “undetermined” when the evidence is empty or conflicting.

If you are a developer building detection into a product, the natural next step is wiring up a detection API and confirming it behaves under real inputs. Download Apidog to design, debug, and test that integration in one workspace, with saved requests and environments so your image-check endpoint behaves the same everywhere.