How to Check if an Image Is AI-Generated (2026)

Why detecting AI images is harder than it sounds

A common assumption is that AI images have a recognizable look. Early generative models did produce distinctive artifacts: blurry faces, impossible anatomy, dreamlike textures. Tools trained on those outputs achieved reasonable accuracy at classifying them.

But those classifiers were chasing a moving target. As image generators improved, the artifacts they relied on became rarer. By 2025, benchmark studies from teams at MIT, Stanford, and Google showed that state-of-the-art classifiers misidentified AI images at rates high enough to make them unreliable for any consequential decision.

The deeper problem is that detection approaches the wrong problem from the wrong angle. A classifier tries to infer provenance from pixel patterns. Provenance systems like C2PA record and verify provenance at the moment of creation. The latter is inherently more reliable, which is why the industry has shifted toward embedding rather than post-hoc detection.

That said, provenance standards are still being adopted. Not every image carries a credential. So in practice you need a layered approach: start with the most reliable signals and treat the less reliable ones as supporting context, not conclusions.

Method 1: Content Credentials (C2PA)

Content Credentials are the closest thing the industry has to a verifiable answer. They are a cryptographically signed provenance manifest, embedded into the image file itself, that records how the image was created and edited. The standard is maintained by the Coalition for Content Provenance and Authenticity (C2PA), whose members include Adobe, Microsoft, Google, OpenAI, BBC, Associated Press, Nikon, Canon, and many others.

What a Content Credentials manifest contains

• The originating software or hardware (e.g. 'Adobe Firefly', 'ChatGPT Image Generator', or a specific camera model)
• A creation timestamp
• A record of edits applied after creation, including which tools performed them
• A cryptographic signature that makes the manifest tamper-evident
• Optionally, the creator's identity if they chose to bind it

How to verify Content Credentials

The simplest path is the public verification tool at contentcredentials.org/verify. Drag an image into the tool and it will display the full manifest if one exists. No account is required and no image data is sent to third parties beyond what the verification step requires.

Google announced at Google I/O 2026 that Chrome will add native C2PA credential display, showing a small indicator directly in the browser when a signed credential is detected. Google Search is also gaining the ability to surface credential information in image results. These integrations mean Content Credentials will increasingly show up without any deliberate action from the viewer.

Which AI tools embed Content Credentials

What Content Credentials cannot do

A credential tells you what it claims. It cannot tell you whether an image that lacks a credential was made by AI. Absence of credentials is not evidence of a real photograph. Anyone can take a screenshot, strip metadata, or simply download from a tool that does not embed credentials. The standard solves the disclosure problem for compliant tools; it does not solve the adversarial detection problem.

Additionally, stripping EXIF data from an image also removes the C2PA manifest if the manifest is stored in the file's metadata block rather than as a cryptographic hash bound to the pixel data. Newer C2PA versions address this by binding the manifest more tightly to the image content, but older implementations can be stripped.

Method 2: SynthID (Google DeepMind invisible watermark)

SynthID is Google DeepMind's approach to watermarking AI-generated images. Unlike a visible watermark, SynthID modifies pixel values in a pattern that is imperceptible to the human eye but detectable by a trained neural network. The modification is distributed across the entire image rather than concentrated in a corner, which makes it harder to remove by cropping.

What SynthID can withstand

• JPEG re-compression at typical quality settings
• Cropping (including aggressive crops that remove large portions)
• Resizing and aspect ratio changes
• Minor color grading adjustments
• Screenshot captures at common screen resolutions

• Heavy AI-based inpainting that rewrites large image regions
• Extreme resampling or resolution upscaling that reconstructs pixels
• Adversarial attacks specifically designed to remove the watermark

SynthID is embedded in images generated by Google's Imagen model (accessible through Gemini and the Vertex AI API). OpenAI announced it would also embed SynthID in ChatGPT and API-generated images starting in May 2026, alongside their existing C2PA manifest. This makes SynthID increasingly relevant even outside the Google ecosystem.

How to check for a SynthID watermark

Google provides a public SynthID Detector where you can upload an image and receive a result indicating whether a SynthID watermark is detected, not detected, or inconclusive. The tool is free and does not require a Google account for basic checks.

A detected watermark is a strong indicator that the image originated from a Google-affiliated AI tool. A not-detected result is not evidence of a real photograph; it only means no SynthID watermark was found. Many AI tools do not embed SynthID.

Method 3: EXIF metadata clues

Before Content Credentials became widespread, EXIF metadata offered the most accessible window into an image's history. It still does, when the metadata is present and unmodified.

The key field to examine is Software. Many AI generation tools write their name here by default. Common values that indicate AI origin include strings like "Adobe Firefly", "Midjourney", "Stable Diffusion", "DALL-E", or model-specific identifiers. If the Software field names a recognized AI tool, that is meaningful evidence.

Other useful fields include Camera Make and Model (absent in AI images unless spoofed), GPS coordinates (AI images almost never carry real GPS), and Lens information (also absent in AI outputs).

How to inspect EXIF data

Use the free SammaPix EXIF viewer. Drop any image into the tool and it displays every metadata field in a structured view. The tool runs entirely in your browser, so the image is never uploaded anywhere. This makes it suitable even for sensitive images you would not want to send to an external server.

The critical limitation of EXIF analysis

Metadata is trivially editable. Stripping all EXIF from an image takes a single command. Adding plausible-looking camera metadata takes only slightly more effort. A sophisticated actor who wanted to pass off an AI image as a real photograph would simply strip the metadata or inject fake camera data before publishing.

Absent metadata tells you nothing useful. An image with no EXIF data could be AI-generated, could be a photo from a privacy-conscious photographer who always strips metadata, or could have come through any social platform that strips it automatically. Treat EXIF findings as corroborating evidence rather than conclusions in isolation.

Method 4: Visual tells (and their limits)

Visual analysis was the first widely used method for spotting AI images. It still works as a rough filter, particularly for images from older models or quickly generated outputs. But it is becoming less reliable with each generation of AI tools, and placing too much confidence in it leads to both false positives (accusing real photos of being AI) and false negatives (missing well-rendered AI images).

Artifacts that still appear in many AI images

• Hands and fingers: Extra digits, fused fingers, or anatomically impossible joint angles. This artifact has become less common in frontier models but persists in mid-tier and locally run models.
• Text inside the image: Signs, labels, book spines, and shirt graphics often contain garbled or nonsensical characters that look plausible at a glance but dissolve into noise on close inspection.
• Reflections: Eyes and glasses in portraits often reflect a scene that does not match the environment the subject appears to occupy. Mirrors and reflective surfaces in the background frequently show physically inconsistent content.
• Skin and fabric texture: AI models sometimes produce unnaturally smooth or perfectly uniform textures, particularly on skin, fabric weaves, and hair. Real photographs have natural variation at the pixel level.
• Background incoherence: Objects in the distance sometimes merge into each other or fade in ways that do not follow natural perspective or depth of field. Architectural elements like windows and tiles can be irregular when examined closely.
• Lighting inconsistency: Shadows that fall in contradictory directions, or catch-lights in eyes that do not match the apparent light source, are common in AI portraits.

Why visual analysis is unreliable as a primary method

The artifacts listed above are traits of current models, not of AI image generation as a fundamental technology. Frontier models from OpenAI, Midjourney, Flux, and others render hands correctly most of the time in 2026. Text rendering, while still imperfect, has improved substantially. The error rate in visual detection is no longer low enough to confidently label an image based on a cursory inspection.

There is also a strong confirmation bias risk. If you expect an image to be AI-generated, you will find artifacts in it. Real photographs have noise, lens aberrations, compression artifacts, and unusual lighting too. Treating these as evidence of AI authorship leads to false accusations. Use visual cues as a starting question, not a verdict.

The honest bottom line

No single method gives you a reliable answer in every case. What you can do is apply the methods in order of trustworthiness and accumulate evidence:

1. 1Check for Content Credentials first. Go to contentcredentials.org/verify. If a signed manifest is present from a known AI tool, you have your answer with high confidence. If the manifest shows camera hardware from a real device and an unbroken editing history, that is meaningful evidence the image is real.
2. 2Run the SynthID Detector. A detected SynthID watermark confirms an AI origin. A negative result narrows the field but confirms nothing.
3. 3Inspect EXIF metadata. Use the SammaPix EXIF tool. Look for AI software names, absent camera information, or implausible metadata. Treat this as supporting evidence.
4. 4Apply visual analysis with skepticism. Look for the artifacts described above, particularly in hands, text, and reflections. Note observations but do not treat them as conclusions.
5. 5Acknowledge uncertainty when it exists. If no credentials are present and the image is visually convincing, the honest answer is that you cannot determine its origin from the image alone. Additional context (platform, source, reverse image search) may help.

The real fix is provenance at scale. C2PA is gaining rapid adoption across major AI platforms, camera manufacturers, and browser vendors. As embedding becomes the default and verification becomes native in browsers and search engines, the question of "is this AI?" will increasingly answer itself without any deliberate investigation. We are in a transition period where the infrastructure is being built but is not yet universally available.

If you publish AI content: how to disclose it

Detection is one side of the provenance problem. Disclosure is the other. If you create and publish AI-generated images or video, the regulatory and ethical landscape in 2026 is moving toward mandatory transparency.

The EU AI Act, which became fully applicable in August 2026, requires that AI-generated content intended for the public be clearly labeled as such. The law applies to content distributed in the European Union regardless of where the creator is located. Violations can result in significant fines under the Act's enforcement framework.

Beyond regulatory compliance, disclosure builds audience trust. Platforms including LinkedIn, YouTube, and TikTok have introduced their own AI-generated content labeling requirements. Voluntarily labeling AI content before platforms do it for you positions you as a transparent creator rather than someone caught out by an automated system.

Practical disclosure options

• Visible label on the image: Use the SammaPix AI Label tool to add a standardized disclosure badge to still images. The tool runs in your browser, requires no upload, and lets you choose the badge position and style.
• Video disclosure label: Use the SammaPix AI Label for Video to add a persistent disclosure overlay to video files without re-encoding the entire clip.
• Embedded credentials: If your creation workflow uses Adobe tools, you can attach Content Credentials directly from Photoshop or Lightroom. This embeds a signed manifest that verification tools like Chrome and contentcredentials.org/verify will surface automatically.
• Caption or alt text disclosure: For platforms that do not support metadata, a clear text disclosure in the caption or image description is the minimum viable approach. Something like "Image created with [tool name]" satisfies the spirit of the requirement.

SammaPix is also building a Content Credentials checker as a native tool, which will let you verify C2PA manifests directly in the same browser-based environment as the rest of the tool suite. This page will link to it once it launches.