An AI image detector returns a probability, not a verdict. In NewsGuard’s May 2026 audit of five leading tools, real news photos were wrongly labeled AI-generated 13.33% of the time, and the worst tool got 40% wrong. Watermarks (SynthID) and Content Credentials (C2PA) are stronger evidence when present, but say nothing when absent. Check provenance first, treat detectors as a hint, then verify the source.
Picture a small shop owner who gets a “damaged parcel” photo from a customer asking for a refund. Someone on the team drops it into a free detector and the tool comes back with a high AI score. (Hypothetical, but easy to imagine.) Deny the refund? Almost certainly not on that score alone, and the reasons are more specific than “detectors aren’t perfect.”
The AI image detector scoreboard: vendor claims vs. what NewsGuard measured
NewsGuard, a news-reliability rating company, published an audit on May 8, 2026. It took 15 authentic news photos from Reuters, the Associated Press, The New York Times, The Guardian and Google Earth, all tied to one conflict. It then had AI tools edit each photo twice: once lightly (“enhance the lighting, blur the background”) and once heavily, in a way that changed the meaning of the scene. That made 45 images. Five detectors saw all of them, using each tool’s free or cheapest tier, and anything scored 50% or higher counted as AI.
| Tool | Headline claim on its own site | Real photos wrongly flagged as AI (of 15) | Heavily altered images caught (of 15) |
|---|---|---|---|
| ScamAI | 95.3% “detection accuracy” (site notes it varies by media type) | 6 (40%) | 12 (80%) |
| ZeroGPT | No accuracy figure published for images | 3 (20%) | 14 (93.33%) |
| AI or Not | 98.9% “AI detection accuracy,” from its own evaluation on a public academic dataset | 1 (6.67%) | 15 (100%) |
| Hive | Says a 2024 independent study found it beats competing models | 0 | 9 (73.33%) |
| Sightengine | Markets “highest accuracy,” citing an academic study | 0 | 5 (33%) |
Three things stand out. First, the two tools with a perfect record on real photos, Hive and Sightengine, were also the two weakest at catching the heavily altered images. Second, the tools disagreed constantly: on 35 of the 45 images, at least one tool gave a different verdict from the rest. Third, a vendor’s accuracy percentage and NewsGuard’s result aren’t measuring the same thing. One comes from a dataset the vendor chose; the other comes from 15 newsroom photos. Neither is lying. They just answer different questions, the same trap we found when the same model scored differently on the same benchmark depending on the yardstick.
Read the small print before you repeat any of these numbers
With 15 real photos per tool, one wrong call moves a score by almost seven points. The gap between AI or Not (1 miss) and Hive (0 misses) is within noise. NewsGuard also used the cheapest tier of each product, so paid plans could behave differently, and every image came from conflict reporting, which tends to be compressed, blurry and oddly lit.
That last point matters because the vendors themselves say it. ZeroGPT’s CEO told NewsGuard that resizing, compression, blur, high contrast and unusual lighting can push a real image toward an AI verdict. AI or Not’s CEO said low image quality can do it too. Any AI image detector reads statistical patterns in pixels, and ordinary photo processing leaves patterns of its own.
There’s also a definition problem hiding in the lightly edited group. Those photos were touched up by AI tools, and the detectors flagged them anywhere from 27% (Hive, Sightengine) to 93% (ScamAI). ScamAI’s co-founder argued that even a filter leaves a trace, so it counts. That’s a defensible position, but it answers “did AI touch this?” rather than “is this photo fake?” Nobody has agreed where the line sits, and the tools don’t tell you which question they’re answering.
A much larger test points the same way, with caveats. A September 2026 arXiv preprint (LAION-Mobile) audited twelve academic detectors, including on a 9,115-image set of smartphone photos. Its authors report that only five of the twelve beat chance on modern AI content, and that once those five were calibrated for it, they still flagged 39% to 63% of real smartphone photos. Their conclusion is that a detector’s false-alarm rate depends heavily on how its threshold was set. It’s a preprint, not peer-reviewed, and these are research models rather than the commercial tools above. Still, it explains why the same product can look great on a vendor’s page and useless on your photos.
Watermarks and Content Credentials read something else entirely
A classifier guesses from the pixels. SynthID and C2PA don’t guess; they look for a signal that a participating tool put there on purpose.
SynthID is Google’s invisible watermark. In a May 19, 2026 announcement, Google said it has watermarked over 100 billion images and videos plus 60,000 years of audio, and that OpenAI, Kakao and ElevenLabs are bringing SynthID to more of their AI-generated content. It also named a partnership with NVIDIA for video from its Cosmos models. Checking works through the Gemini app (Google says it has been used 50 million times), with Search at launch and Chrome to follow. Google’s own model names, including the Nano Banana family, are sorted out in our Google AI models guide, and the SynthID angle on image generators is in our Nano Banana Pro vs. Midjourney comparison.
C2PA Content Credentials are signed metadata describing how a file was made or edited, with or without AI. Because they can record a real camera capture, they can support authentic photos as well as flag AI ones. Google says Pixel 10 was the first phone to add them in its native camera app, and Meta plans to label camera-captured media with them on Instagram. Verification began rolling out in the Gemini app on May 19, with Search and Chrome promised in the following months. Availability moves fast, so check what’s live before you build a process around it. The standard itself is described at contentcredentials.org.
Both come with a catch that vendors are upfront about. OpenAI’s provenance documentation says its check only looks for OpenAI’s own signals, so a “not detected” result doesn’t rule out that OpenAI made the image, and it can’t identify images from other companies’ models. It also warns that editing, converting or sharing a file can strip the metadata, while SynthID “may” survive some transformations, which is not a promise it survives all of them. A hit is strong evidence. A blank is close to meaningless.
Businesses get a third route: Google announced an AI Content Detection API on Google Cloud, launching with trusted partners, and it lists uses like sorting feeds and preventing insurance fraud. The announcement includes no accuracy figures, so treat it as unproven until someone publishes an audit.
Side by side: what each method can and can’t tell you
| Method | What it reads | Best at | Fails when | Where you can use it |
|---|---|---|---|---|
| AI image detectors (classifiers) | Statistical patterns in the pixels | A quick hint on an image with no history, even with metadata stripped | Real photos are resized, compressed, blurred or AI-enhanced; new generators appear; tools and thresholds disagree | Web tools, many with free tiers |
| SynthID watermark | An invisible signal embedded by participating generators | Strong evidence of AI origin when found; may survive some edits | The generator never adopted it; heavy edits; it can’t say who made the image or why | Gemini app; Search and Chrome per Google’s May 2026 rollout plan |
| C2PA Content Credentials | Signed metadata on origin and edit history, including camera capture | Backing up an original, unedited file with a valid credential | Screenshots, re-exports and uploads drop it; absence proves nothing either way | Gemini app; other C2PA readers |
| Context checks | Who posted it first, other angles, timestamps, the original file | The only method that catches a real photo used in a false context | Slow, needs a person, and the original may not exist | Free: reverse image search, ask the sender for the original |
Check an image in this order
- Get the original file, not a screenshot. Screenshots and re-exports are the fastest way to lose Content Credentials.
- Look for provenance. Run it through the Gemini app’s SynthID and Content Credentials check. If ChatGPT might have made it, OpenAI documents a check for its own signals.
- Then use a detector, and use two. Note the scores and expect them to disagree, since they did on 35 of 45 images in NewsGuard’s audit.
- Check the story around the image. Find the first poster, look for other angles, run a reverse image search. This is the same source-first habit we describe in our guide to fact-checking AI-generated content.
- Price the mistake. If a wrong call means accusing someone, denying a refund or publishing a claim, stop at “unverified” and bring in a person with the right expertise.
When not to use a detector at all
Don’t use one to accuse a person. That covers a freelancer’s submitted work, a customer’s evidence and an employee’s report. NewsGuard warned that detector results can be cited to dismiss real images and that they unfairly malign the people who posted them. A score is not proof of anything you’d say out loud.
If you take or edit photos yourself, keep the untouched original next to any edited version. AI-assisted lighting and background blur are exactly what got real photos flagged in the audit, which matters if you run product shots through the tools in our AI ecommerce roundup. If you publish AI-generated images yourself, don’t strip their labels; the disclosure side is covered in our EU AI Act guide. And if customers send you photos as evidence, add one line about handling them to your AI usage policy.
This is general information, not forensic or legal advice. For disputes or anything that may end up as evidence, use a qualified examiner.
Frequently Asked Questions
Are AI image detectors accurate?
It depends on the test. Vendors publish accuracy figures from datasets they choose. In NewsGuard’s May 2026 audit of five leading tools on 45 images, real photos were wrongly flagged 13.33% of the time overall, ranging from 0% to 40% by tool. Treat any single score as a hint, not a verdict.
Can a detector say my real photo is AI-generated?
Yes. Vendors themselves say resizing, compression, blur, high contrast, unusual lighting and low image quality can push a real image toward an AI verdict, and light AI-based touch-ups get flagged too. Keep your untouched original and any camera metadata so you can show where the photo came from.
Does a missing SynthID watermark mean an image is real?
No. SynthID only appears in content from generators that adopted it, and OpenAI’s documentation says a “not detected” result doesn’t rule out that its tools made the image. A blank result means only that no signal was found.
How can I check whether an image came from ChatGPT or Gemini?
The Gemini app can check images for SynthID and Content Credentials, and it’s the option most people can use directly. OpenAI documents a provenance check for its own signals, which is aimed at developers. Neither can identify images from generators that don’t use these signals.
Are C2PA Content Credentials and SynthID the same thing?
No. Content Credentials are signed metadata that can be lost when a file is edited, converted or shared. SynthID is an invisible watermark inside the image itself that may survive some edits. They cover each other’s weaknesses, which is why Google and OpenAI are pairing them.
Should a business run customer photos through an AI image detector?
Only as one signal among several. Ask for the original file, look for provenance signals, check the surrounding story, and escalate to a person before denying a refund or making an accusation based on a detector score.
Shurah is the founder of AI Tools Daily, tracking pricing, licensing and policy changes across AI tools so readers can make decisions without wading through marketing claims themselves.
