AI Image Detectors, Compared

An AI image detector is one of two very different things, and mixing them up is why people get bad answers. A provenance checker reads a signal that was deliberately attached to the file. A statistical detector guesses from pixel patterns.

The provenance route uses tools like Content Credentials Verify from the Content Authenticity Initiative and Google's SynthID Detector. The statistical route uses classifiers like Hive Moderation, AI or Not, Illuminarty, and Sightengine.

This page compares both routes on what they actually read, what they output, and what each result is worth. The order matters: check provenance first, treat a classifier score as a weak second opinion, and never treat either as proof on its own.

Content Credentials Verify vs. Hive Moderation: Side-by-Side

The Two Things Called AI Image Detection

AI image detection splits into provenance verification and statistical detection, and the two answer different questions. Provenance verification reads a signal that a camera or a generator deliberately attached to the file. Statistical detection inspects the pixels and estimates odds.

Provenance signals come in two forms. C2PA Content Credentials are cryptographically signed metadata that travel with the file. SynthID is an invisible watermark that Google DeepMind embeds inside the image data itself.

Statistical detectors have no inside information. They compare an image against what they learned about generated pictures and human photographs, then report a confidence number.

• Provenance: a planted, verifiable signal. Strong evidence when present.
• Statistical: an inference from pixels. Always a probability, never proof.
• Absent provenance is not evidence of anything, because most images carry none.
• A classifier score of 90 percent is a guess with a number attached to it.

Content Credentials Verify vs. Hive Moderation

Content Credentials Verify and Hive Moderation are the clearest example of the two approaches, and picking between them is really picking a question. Verify reads attached provenance. Hive reads pixels.

Content Credentials Verify is the free checker run by the Content Authenticity Initiative. You drop in an image or paste a URL, and it reads the signed C2PA manifest, showing the issuer, the tools used, edit history, and whether AI was involved.

Hive Moderation sells AI-generated and deepfake detection as an API, with a free public demo and a browser extension. Hive returns confidence scores and, when available, the likely model behind the media.

Verify fails quietly and often: no manifest, no answer. Hive always answers, which feels more useful and is exactly why people over-trust it.

• Verify: free, reads signed metadata, tells you nothing when metadata is missing.
• Hive: free demo plus paid API, always returns a score, can be wrong in both directions.
• Verify going blank means it has nothing. Hive always produces a number anyway.
• Use Verify first. Use Hive only when Verify comes back empty.

SynthID Detector, AI or Not, and Others

Beyond those two, the tools worth knowing split along the same line: one more provenance route and a set of classifiers. SynthID Detector is Google's portal for checking whether content carries a SynthID watermark from Gemini, Imagen, Veo, or Lyria.

SynthID is designed to survive edits that destroy metadata, including cropping, filters, and lossy compression, according to Google DeepMind. Portal access has been limited to early testers, journalists, and researchers, but anyone can upload an image to Gemini and ask whether it carries a SynthID watermark.

In May 2026 Google said Content Credentials verification was rolling out in the Gemini app, with Search and Chrome to follow, and that Pixel 10 was the first phone to write Content Credentials from its native camera app.

On the classifier side, AI or Not covers images, video, and audio and offers a free check plus an API. Illuminarty adds a heatmap that shows which regions drove its verdict, which is useful when you need to explain a result. Sightengine targets developers who want detection inside an existing moderation pipeline.

• SynthID Detector: provenance, Google-model content only, robust to cropping and compression.
• Gemini: the practical SynthID route for most people; upload the image and ask.
• AI or Not: multi-format classifier with a free check and a paid API.
• Illuminarty: classifier with region heatmaps, better for showing your working.
• Sightengine: classifier built for pipelines and bulk moderation.

How to Check If an Image Is AI Generated

Check provenance first, run a classifier second, and inspect the picture yourself third. Doing it in that order stops you from treating a guess as an answer.

Step one costs nothing and takes seconds. Drop the original file into Content Credentials Verify, and if you suspect a Google model, upload it to Gemini and ask whether it carries a SynthID watermark.

Step two is the classifier. Run the image through one or two of Hive, AI or Not, or Illuminarty, and treat agreement between them as mildly informative rather than conclusive.

Step three is your own eyes and the surrounding context. Look for garbled text on signs, hands and teeth that do not resolve, lighting that disagrees with the shadows, and jewellery or patterns that change halfway across the frame.

Context often beats every tool. Reverse image search the picture, find the earliest posting, and check whether the account that shared it has a history of posting generated work.

• 1. Verify provenance: Content Credentials Verify, plus a SynthID check via Gemini.
• 2. Run a classifier: use it as a second signal, not a ruling.
• 3. Inspect manually: text, hands, reflections, repeated patterns, physics.
• 4. Check the trail: earliest source, uploader history, reverse image search.
• 5. Use the original file. A screenshot destroys the metadata in step one.

What Does an AI Watermark Look Like?

Most AI watermarks are invisible, so expect to see nothing at all. There are two visible kinds and two invisible kinds, and only the visible ones show up on screen.

Visible marks include a burned-in logo or badge in a corner, and platform labels such as an 'AI info' tag added by the site rather than the file. Those are the easiest to spot and the easiest to crop away.

Invisible marks include SynthID, embedded in the pixel data, and C2PA Content Credentials, stored as signed metadata. Neither changes how the picture looks, and neither can be seen without a tool that reads it.

Some viewers show a small 'cr' icon on an image that carries Content Credentials. That icon is a display convention, not the credential itself, so its absence proves nothing.

• Visible: corner logos, platform AI labels. Croppable and easy to fake.
• Invisible: SynthID in the pixels, C2PA in signed metadata.
• The 'cr' icon signals a credential exists in supported viewers.
• Seeing no mark tells you nothing about how the image was made.

False Positives and Why Nobody Should Be Accused

No AI image classifier output should ever be treated as proof that a person used AI, because these tools produce false positives on genuine photographs and real artwork. The consequences land on people, not files.

Classifiers are trained on pixel statistics, so ordinary processing can push a real photo toward an AI verdict. Heavy denoising, phone HDR stacking, AI upscaling, aggressive retouching, and re-compression all smooth an image in ways that resemble generated output.

Illustrators and photographers get hit hardest. Clean digital painting, airbrushed skin, and smooth gradients are exactly the textures classifiers associate with generation, which is why hand-drawn work gets flagged on art platforms and in competitions.

This matters most in education and employment. A student marked as a cheat or a freelancer dropped over a detector score has been judged by a probability engine that cannot show its evidence and has no appeal process.

If you run reviews, write the rule down before you need it: a classifier score can start a conversation, and it can never end one. Ask for source files, layered originals, camera raw files, or version history, because those are evidence and a score is not.

• Real photos can score as AI-generated; the error rate is not zero.
• Upscaling, denoising, and retouching all raise false-positive risk.
• Digital art is systematically over-flagged by pixel-based classifiers.
• Never discipline a student, employee, or contractor on a detector score.
• Ask for raw files and version history instead. Those can actually be checked.

How Accurate Are AI Image Detectors?

No public AI image detector has an accuracy figure you should rely on, because the published numbers come from vendors testing their own tools on their own data. AI or Not advertises 98.9 percent accuracy on its site, and that is a vendor claim measured on a vendor test set.

Independent results differ, and they differ a lot depending on what you feed the tool. Accuracy on a fresh, unedited output from a well-known generator is not the same as accuracy on a cropped, re-compressed screenshot from a social feed.

New image models appear constantly, and a classifier trained before a model existed has never seen its artefacts. Detection quality on the newest generators is usually the weakest part of any tool and the least advertised.

The number that matters for real decisions is the false-positive rate on genuine photographs, and no vendor publishes an independently audited one. Treat every marketing accuracy figure as unverified.

• Vendor accuracy claims are self-measured. Attribute them, do not repeat them as fact.
• Performance drops on edited, cropped, and re-compressed images.
• Brand-new generators are the hardest case for any classifier.
• No audited false-positive rate exists for any tool on this page.

Where Provenance Checks Fall Down

Provenance is the stronger signal, but it fails in one specific way: the signal often is not there. C2PA metadata can be removed by editing tools, file conversions, and platforms that re-encode uploads, and a screenshot drops it every time.

That is a documented limitation of the standard, not a scandal. OpenAI says the same thing about its own images: provenance signals help indicate origin, but metadata can sometimes be stripped by platforms, editing tools, or conversions.

Coverage is the other gap. A credential only exists if the camera, the generator, or the editor wrote one, and adoption is uneven across generators and phones as of August 2026.

SynthID closes part of that gap because it lives in the pixels rather than the metadata, so it survives cropping and compression. It only covers content from Google models, which leaves most of the internet uncovered.

So a present, valid credential is strong evidence. An absent one is the normal state of an ordinary image and carries no meaning at all.

• Metadata is strippable by design constraints, not by conspiracy.
• Screenshots and re-uploads routinely remove Content Credentials.
• Coverage depends on the generator, camera, or editor writing a credential.
• SynthID survives edits but only covers Google-model content.

Do AI Image Detectors Affect Your Rankings?

No, AI image detectors play no part in how Google ranks your pages, because Google does not run third-party image classifiers over your site. A detector score is not a ranking input.

Google's own provenance work points the other way. It is adding Content Credentials verification to Gemini, and to Search and Chrome, so people can inspect how an image was made. That is a transparency feature for readers, not a demotion signal for publishers.

Using AI images on your site is a quality and disclosure question rather than a search penalty question. An image that misleads readers hurts you because it misleads readers.

So do not build a workflow around scoring low on a classifier. Label generated visuals where it matters to your audience, keep the originals, and choose images that genuinely help the page.

• Google does not use third-party AI image detectors to rank pages.
• Provenance in Search and Chrome is reader-facing transparency, not a penalty.
• Disclosure and usefulness matter more than any detector score.
• Keep source files. They are worth more than a classifier result.

The Verdict

Start with provenance. Content Credentials Verify is free, takes seconds, and when it finds a signed manifest you have real evidence about where an image came from. Add a SynthID check through Gemini when a Google model is plausible.

Reach for a classifier only when provenance comes back empty, which will be most of the time. Hive Moderation is the strongest general-purpose option with a free demo and a paid API, AI or Not is a fast multi-format check, Illuminarty explains itself with heatmaps, and Sightengine fits into a pipeline. Verify current pricing on each vendor's own page.

Whatever the tool, the ceiling is the same. A classifier gives you a probability, never proof, and it does flag real photographs and real artwork as AI-generated. Never let a score decide something that affects a person, and remember that no detector output has any bearing on how Google ranks your pages.