How we make this list: we compare pricing pages, product documentation and public user reviews. Ratings come from G2, Capterra or Trustpilot, with the date we checked them.
Key takeaways
- Start with provenance, not classifiers: Content Credentials Verify reads the signed C2PA manifest some cameras and generators attach, reporting a verifiable record when one exists, and Google's SynthID Detector checks for Google's watermark in content made by Google models.
- Hive Moderation and Sightengine are the production classifiers: enterprise APIs with generator attribution in Hive's case, and in Sightengine's a generous free tier of 2,000 API checks a month and paid plans from $29 a month.
- AI or Not and Illuminarty are the accessible checks: a free instant verdict from AI or Not, and from Illuminarty a heatmap showing which regions drove its judgement, with 5 free scans a day in current comparisons.
- The recommended professional workflow from 2026 testing is an ensemble: provenance first, then three to five classifiers, treated as a majority vote on a weak signal, never as a verdict.
Quick comparison
Is your product missing or out of date?
We review this list regularly. Vendors can request a listing, send updated pricing or ask about featured placement.
Starting priceFree
Free planYes
Verify is the correct first step in any serious check because it answers a different question from the classifiers: not does this look generated, but what does its signed history say. When a manifest is present and intact, you have provenance, the strongest signal available in this field.
What we like- Reads the signed Content Credentials manifest attached to a file at creation: who or what made it, which tool, whether AI was involved, a record rather than a guess, so it cannot produce a statistical false positive, free to use, backed by the Content Authenticity Initiative and the C2PA standard that camera and software makers are adopting, the only entry here whose positive result is actual evidence
Watch out for- Most images online carry no manifest, and the tool reports nothing in that case, which proves nothing either way, manifests can be stripped by screenshots and many platform uploads, coverage grows with adoption, which is still partial in 2026
Starting priceFree
Free planYes
SynthID is the other half of the provenance layer: where Verify reads attached metadata, the watermark is woven into the pixels, so it survives the handling that strips manifests. For content made in Google's tools, it is close to a definitive origin check.
The trade is the boundary in its name. SynthID knows one family's work. The detector's silence about everything else is the most misread result in this field, routinely presented as proof an image is real when it proves only that Google's models did not watermark it. Use it for the question it answers, and no other.
What we like- Checks for SynthID, the imperceptible watermark Google embeds in content from its own AI models, a provenance signal that survives cropping and compression per Google's published work, free to use through Google's detector surface, a watermark hit is strong evidence of Google model origin
Watch out for- It answers only for Google models: a clean result says nothing about an image from Midjourney, Stable Diffusion or any other generator, or from a camera, access is tied to Google's ecosystem and partner surfaces rather than a general public API, like all provenance, a missing watermark proves nothing
Starting priceSee website
Free planYes
Hive is what platforms buy when detection is infrastructure: high volume, many media types, one API, with attribution that helps a moderation team understand what it is seeing. Independent comparisons in 2026 placed it in the top group of detectors on their test sets, which is the strongest kind of praise this category honestly supports.
The trade is access and certainty. An individual cannot buy a plan, and an enterprise cannot buy certainty: Hive's score is a probability from the best resourced guesser in the field, and the correct production pattern is thresholds plus human review on the consequential cases, not automated verdicts on a score.
What we like- The enterprise heavyweight: detection across images, video, deepfakes, voice, music and text, used to moderate content for major platforms, generator attribution attempts to name which model likely made an image, a distinction rivals rarely offer, a free web demo and a genuinely free browser extension for casual checks, fast API responses in independent testing
Watch out for- Paid access is enterprise API only, with custom pricing and no individual plan, accuracy figures quoted for it come from tests, not a published verified rate from Hive itself, as a classifier it shares the category's failure mode: heavily processed real photos are its known weak case
Starting price$29 per month
Free planYes
Sightengine is the pragmatic build: if detection must live inside your product, its free tier lets a developer evaluate properly, its documentation names the generators it covers, and its pricing is readable arithmetic. That transparency about coverage is worth more than any accuracy slogan, because coverage is what decides your real world results.
The trade is explainability. When Sightengine flags a user's photograph, your support team has a number and nothing to show. Products that act on flags, suspensions, accusations, rejections, should pair it with a tool that explains, or with a human review step that supplies the judgement the API cannot.
What we like- The developer's choice: a clean API detecting outputs from more than 20 named generators, alongside deepfake and broader moderation models, the most generous free tier here, 2,000 API operations a month, paid plans from $29 a month with per image costs falling to fractions of a cent at volume, ranked at the top of an independent academic benchmark on 80,000 images in one comparison's account
Watch out for- No visual explanation of verdicts, so a flagged image arrives without evidence a human can inspect, operation based pricing assumes an integration project, not casual checking, published generator coverage lists matter: an unlisted generator is an untested one
Starting price$5 per month
Free planYes
AI or Not is the tool most people will actually use, because it asks nothing: no account, no integration, an answer in seconds. As the opening move of an ensemble check, or a private sanity check before sharing a suspicious image, it earns its popularity.
The trade is that its convenience invites misuse. A one click verdict feels like knowledge, and it is the furthest thing from it in this list: a single model's guess, weakest exactly on the processed, compressed images people most often check. Let it raise questions. Never let it settle them.
What we like- The fastest first pass: upload an image and get a verdict with no account and no setup, free checks for casual use, with paid credits from $5 a month and an API for volume, covers images, video, voice, music and text, with generator identification attempted, scored well in one vendor run 2026 test, 8 of 10 at a strict confidence threshold
Watch out for- A single classifier's verdict, with all the category's error modes and no heatmap to inspect, free limits and credit pricing have shifted across comparisons, so check the current terms, the strong test result quoted for it came from a company that sells detection, a reason to discount it, not to repeat it as neutral fact
Starting price$10 per month
Free planYes
Illuminarty's heatmap is the most honest interface in the classifier field, because it admits the verdict has a geography and lets a human argue with it. In education and publishing disputes, where the output will be shown to the accused, that inspectability is an ethical feature, not a luxury.
The trade is the seduction of the picture. A glowing region feels like evidence even when the underlying call is wrong, and the benchmark record here includes the embarrassing direction of error, AI passed as human. Show the heatmap, by all means, as the start of a conversation. Do not laminate it as proof.
What we like- The explainable entry: its heatmap shows which regions of an image drove the AI judgement, the only widely available tool that shows its working, free basic detection, with 5 free scans a day without an account in current comparisons, paid plans from $10 a month, useful exactly where a bare score fails: conversations with the person whose image was flagged
Watch out for- Explainability is not accuracy: in one 2026 benchmark it returned a false negative on a generated portrait, passing AI content as real, a heatmap can make a wrong verdict look scientific, which is a persuasion hazard of its own, a smaller operation than the enterprise pair, with coverage to match
Starting priceSee website
Free planNo
Reality Defender represents the serious end of the market's shape: detection as one layer in a fraud and impersonation defence, where a flag triggers verification through other channels, a callback, a second factor, a human analyst, rather than an automatic punishment. That architecture is the correct lesson of this whole guide, sold as a platform.
The trade is that the platform cannot repeal the mathematics. Its buyers are paying for integration, throughput and support, and calibrating thresholds against their own traffic. Any enterprise expecting a purchase to settle the deepfake question will be disappointed at renewal; one buying a well run signal will not.
What we like- An enterprise platform aimed at deepfake defence for organisations: detection across image, video, audio and text, delivered by API and integrations for security and fraud teams, probabilistic outputs designed for risk workflows rather than binary shaming, the buyer is a bank, a media company or a public figure's office defending against impersonation
Watch out for- Enterprise sales only, with no public pricing or self serve tier, shares the classifier category's underlying error rates, so its value is in workflow and scale, not certainty, overkill and inaccessible for individuals and small teams
Starting priceFree
Free planYes
Is It AI is listed because free matters: the teacher, the marketplace buyer, the moderator of a small community will use a free tool or none, and a weak signal used humbly beats a strong opinion held blindly.
The trade is the error rate at the bottom of the table. A tool that is wrong most often is most dangerous exactly where it will be used, by people with the least ability to second guess it. If a decision about a person hangs on the check, this is not the tool, alone, to hang it on. Nothing is, which is the point of the guide below.
What we like- Completely free single image checks with no account, the zero friction end of the market, a reasonable second opinion in an ensemble for users with no budget, simple enough for a student or teacher to use without training
Watch out for- Ranked lowest among the tested tools in one 2026 comparison's accuracy table, free services carry the least accountability for their error rates, no API story for production use, treat its verdict as the weakest signal in any ensemble
How we chose
We mixed the two honest families: provenance tools, which report verifiable records, and statistical classifiers, which guess, and we weighted published pricing, free access, explainability and the quality of a tool's own claims about itself. No accuracy percentage from any vendor is repeated in this guide as established fact, because none is independently verified at scale; where test results are mentioned they are attributed to their tests, several run by interested parties on small image sets. We did not claim hands-on testing, and we deliberately ranked verification above detection, because that ordering is the field's own professional consensus in 2026.
How to check an image without harming anyone
Work in this order. First, provenance: run Content Credentials Verify, and if the image might come from Google's models, the SynthID check. A positive provenance result is real information. A negative one is silence, and silence must be reported as silence.
Second, context, which outperforms every classifier here. Where did the image come from: an account with a history, or one created yesterday? Does it appear earlier elsewhere, which a reverse image search answers in seconds? Do the details cohere, text, hands, reflections, the physics of light, the facts of the claimed event? Journalists catch more fakes with those questions than with any score.
Third, if a score is still needed, ensemble: run three to five classifiers, including one that explains itself, and read agreement as a weak signal that raises or lowers your suspicion, never as a verdict. Processed images, screenshots, social uploads, heavy crops, degrade every classifier, so a file that has been through a platform deserves extra doubt in both directions.
Last, the human rule. If the consequence falls on a person, the detector output starts a fair process; it does not conclude one. Let them produce the original file or project history. Originals settle what scores cannot, and the scandals in this field share one shape: a score treated as proof, against someone given no route to answer it.
Frequently asked questions
How accurate are AI image detectors, really?
On clean test sets, the better tools score in the 85 to 95 percent range in the tests that exist, several of them vendor run and all of them small. On real world images, compressed, cropped, filtered, performance drops sharply, and one respected 2026 benchmark concluded most detectors perform little better than chance on hard cases. No tool here publishes an independently verified false positive rate. Plan accordingly.
Can a detector prove an image is AI generated?
No. A classifier offers a probability based on pixel patterns. Proof, in the strong sense, comes from provenance: a signed Content Credentials record, a SynthID watermark from Google's models, or the maker's own publication of the image. Everything else is suspicion with a number attached, and should be described that way to anyone affected by it.
A detector says my photograph is AI generated. What now?
It is wrong often enough that you should not panic, and neither should whoever flagged it. Produce the provenance you have: the original file from the camera or phone, with its metadata, earlier versions, the RAW file if one exists. Run a provenance check. A false positive on a real, heavily processed photo is the best documented failure in this field.
Which tool should a school or employer use?
One with explainable output, like Illuminarty, inside a written process that treats every flag as an allegation to be investigated, with the student or employee able to present originals. The tool matters less than the process. An institution that punishes on a score alone will eventually punish an innocent person publicly, and no vendor's terms will share that cost.
Do watermarks and credentials solve this problem?
They solve the provable slice of it. Content from participating cameras and generators carries evidence of its origin, and adoption is growing. But most images carry no credential, watermarks cover one maker's models each, and bad actors strip what they can. Provenance is the future of verification; classifiers are the imperfect present. Use both, in that order.
Sources we checked (0)
Is your product missing or out of date?
We review this list regularly. Vendors can request a listing, send updated pricing or ask about featured placement.
Shihab Mia · Founder, Rankite
Shihab has 5+ years in B2B and SaaS SEO. He writes Rankite's software buying guides from pricing pages, documentation and public user reviews. More about Shihab.
Content Credentials Verify
SynthID Detector
Hive Moderation
Sightengine
AI or Not
Illuminarty
Reality Defender
Is It AI