AI Image Detector

An AI image detector is three different tools wearing one name. A provenance check reads a watermark or signed metadata and returns yes or no — that is evidence [2][3]. A pattern detector scores the pixels and returns a probability — that is an estimate [9]. Your own eyes return observations you can explain to someone else, with no software involved. Pick by what you need, not by which landing page looks best, because none of the three can do another's job.

• Need evidence — something you can state as fact? Start with the watermark route: SynthID Detector: how to use Google's AI watermark check.
• Need a probability on an image with no watermark? Compare the scoring tools: Best AI image detector in 2026.
• Want to judge it yourself, with no detector in the loop? How to tell if an image is AI generated.

Decide what you need first

The rows are split by what you have in your hand, not by which route you end up on: two rows can point to the same route and still be different problems.

Read it as a branch, not a ranking: choosing the wrong row is the most common way people waste a credit or a working day.

The table also shows where the routes meet. Metadata is not a fourth route — it is a provenance check you run yourself, the same class of evidence as row one, read out of the file instead of out of a platform.

If you need evidence: watermarks and provenance

Provenance tools read a signal embedded when the file was created [1]. Google's SynthID portal checks images, video and audio for a watermark, and states plainly that it "is not a general AI detector" — it only speaks for content from companies that adopted SynthID, which now include OpenAI, NVIDIA and Kakao, with Apple announced as coming [2][4]. OpenAI's own verifier reads SynthID watermarks and C2PA Content Credentials in images and audio from ChatGPT, its API and Codex, and does not currently support text [3].

What that buys you: a positive result you can quote. What it does not: any opinion about an image from a generator outside the programme. A negative is a dead end, not a verdict.

In practice it takes one file at a time and answers with one of two words. The format list is longer than most people expect — HEIC and HEIF are both accepted [2], which is what an iPhone produces and what most people assume will be rejected. The full walkthrough, including the daily quota, is on SynthID Detector: how to use Google's AI watermark check.

If you need a probability: pattern detectors

These score the pixels. DeepAI, QuillBot, Sightengine and Hive all work this way [8][9][11], and the good ones say out loud that they are estimating: DeepAI's own page warns that "detection is probabilistic" and that digital art, airbrush work and AI-based editing can sway the result [9]. At volume you are buying different features — TruthScan sells bulk analysis and a 500ms response time on enterprise plans [7], Sightengine sells an API [8].

What that buys you: coverage of generators that never watermark anything. What it does not: a number you can compare across tools, because we could not find a published test set for any of the detectors we compared — so there is no basis, in anything we read, for putting their scores on one scale.

One reading habit worth keeping: a score is the model's confidence, not its accuracy. A tool that returns 94% is not right 94% of the time; it is 94% sure about this file. Coverage has the same problem in reverse — a detector trained on 2023 generators can read a 2026 model's output as clean, and that failure looks exactly like a confident low score. The comparison, including why vendor accuracy claims do not travel between tools, is on Best AI image detector in 2026.

If you want to judge it yourself: eyes and files

No tool needed. You look at hands and text at 100%, then at whether the light agrees with itself, then at depth, noise and the file's metadata. Each check has a known failure mode — cropping, denoising, re-compression, a genuine two-light setup — so the skill is not spotting an anomaly, it is knowing which anomalies survive scrutiny.

That is a full method rather than a paragraph, so it lives on its own page: How to tell if an image is AI generated.

Why the order matters — and why this is our inference

The ordering below is our inference, not a documented rule: no vendor and no standards body that we read publishes a recommended sequence. The reasoning is about what a number does to the person holding it.

If a pattern detector runs first, you get a percentage before you have any other fact. A percentage reads like an answer, so the investigation usually stops there — and if the watermark check afterwards comes back negative, the 87% is still sitting in your head as if it were evidence.

The more damaging mistake is not the order, though. It is what people do with a negative. A provenance check that returns "no signal found" feels like a clean bill of health: the file came back fine, so it must be real. It did not come back fine. It came back saying that the one thing the tool can see is absent — and outside the participating companies, the tool was never looking at all [2]. Silence and cleanliness look identical if you read the result quickly. In our reading, that misreading is more common than getting the order wrong, and it is the one that ends up in published copy.

The one case where you can skip provenance is when the file's own history already rules it out: a decade-old print, a scan with scanner noise, or a camera original with intact EXIF and a plausible chain of custody.

Three ways people mix them up

A probability treated as evidence. "94% AI" is a model's confidence, not a finding. It cannot be quoted as proof, and a different tool will give a different number for the same file [9].

A roundup treated as a conclusion. The high-ranking comparison pages grade vendors on their own claims rather than testing them — one such page lists its criteria as "accuracy claims" and discloses that the product it ranks first is its own [6]. That is a menu, not a measurement.

A human accuracy figure treated as a threshold. In one large study, people identified AI-generated images correctly 63% of the time [5]. That is a reference point for human judgement on that study's images — not a bar a detector inherits. The reason is not pedantry: the same study found participants most accurate on human portraits and significantly worse on natural and urban landscapes [5]. Change the mix of images and you change the score, which is exactly why a figure from one test cannot certify a tool measured in another.

FAQ

I have one image and no context. What do I do first? Open the file's properties. If it carries camera data or a provenance credential, you may not need a detector at all. If it carries nothing, assume the metadata was stripped and move to a pattern detector as a second opinion.

When do all three routes fail? On a small crop with no metadata, after several rounds of re-compression, on a printed-and-photographed image, and on any question about whether the content is true rather than whether a tool made it. In those cases the honest answer is "cannot determine from this file."

The watermark check found nothing. Does that mean it is not AI? No. It means no participating model's signal was found. The portal covers only companies that adopted SynthID [2], and we have not seen a published figure for how much of the generator market that leaves out; heavy editing can also degrade the signal [2]. Absence of a watermark is the weakest negative in this whole field.

Three tools gave me three different scores. Which one is right? None of them is "right" in the sense you want. They were trained differently, and we could not find a published test set for any of them, so their numbers are not on one scale in any way we can verify [9]. Use them as a second opinion, or better, score them on your own images.

Is free enough, or do I have to pay? Free covers single images and the watermark route, subject to quotas [2]. You pay for throughput, an API, or file-level reporting — that is, when the job is volume rather than curiosity [7][8].

I need a sentence I can put in a report. What do I use? Only a provenance hit supports a statement of fact [3]. Otherwise write what you did: which checks you ran, what you found, and that the result is an assessment rather than a determination.

Sources

1. Google DeepMind — Google SynthID overview: deepmind.google/models/synthid/
2. SynthID Detector portal, supported formats and terms: synthid.com
3. OpenAI — Verify OpenAI-generated content: openai.com/research/verify/
4. Google blog, October 7, 2026 — Google expands SynthID Detector for AI content: blog.google/innovation-and-ai/models-and-research/google-deepmind/synth-id-ai-content/
5. Roca et al., "How good are humans at detecting AI-generated images? Learnings from an experiment", arXiv:2507.18640: arxiv.org/abs/2507.18640
6. DeepfakeDetector.ai — Best AI Image Detectors 2026: 8 Tools Compared & Ranked: deepfakedetector.ai/blog/best-ai-image-detectors
7. TruthScan — AI Image Detector: truthscan.com/ai-image-detector
8. Sightengine — Detect AI-generated images: sightengine.com/detect-ai-generated-images
9. DeepAI — AI Image Detector: deepai.org/ai-image-detector
10. Winston AI — AI image detector: gowinston.ai/ai-image-detector/
11. QuillBot — AI Image Detector: quillbot.com/ai-image-detector