education

Best AI Image Detectors

Compare the best AI image detectors as of August 2026 — Hive, Sightengine, AI or Not, Winston AI, Copyleaks and TruthScan, plus the free provenance checkers from OpenAI, Google and C2PA, with independent false-positive results, pricing and which to use for journalism, moderation or a one-off check.

Updated August 2026

Quick answer: For checking whether a still image was generated by AI, Hive AI is the strongest general-purpose classifier — it and Sightengine were the only two of five tools that never mislabelled an authentic news photograph in NewsGuard’s May 2026 audit. AI or Not was the only tool in that audit to catch 100% of images that had been meaningfully altered by AI, which makes it the better choice for spotting doctored real photos rather than fully synthetic ones. If the image came from ChatGPT, Gemini or another provenance-signing generator, skip classifiers entirely and use a free provenance reader — OpenAI’s verify tool, the Gemini app, or Content Credentials Verify — because a cryptographic signature or watermark is evidence in a way a probability score is not. The caveat that decides everything else on this page: across the same NewsGuard test set the five tools declared authentic photographs AI-generated 13.33% of the time, and one tool got it wrong on 40% of real images, so a detector result is a signal to investigate, never proof.

This page ranks the tools that detect AI-generated and AI-edited still images — a different problem from AI-written text, which we cover separately in best AI detectors. Two things define the moment. First, the independent evidence finally exists, and it shows detectors failing in the direction that matters most: calling real photographs fake. Second, the industry has largely conceded the point and moved to provenance — OpenAI began signing every generated image with Content Credentials and a SynthID watermark in May 2026, which means the reliable way to identify an AI image in 2026 is increasingly to read a signature rather than guess from pixels. If you are about to accuse someone of posting a fake, read the false-positive section before you read the rankings.

The current state of AI image detection: August 2026

Image detection is a younger, messier market than text detection, and the honest summary is that the classifier layer is losing ground while the provenance layer gains it. Five shifts define where things stand.

1. The first serious independent audit landed, and detectors failed on real photos. NewsGuard tested 15 authentic images of the 2026 US–Iran war — sourced from Reuters, the Associated Press, The New York Times, The Guardian and Google Earth — against five leading tools in late April and early May 2026. Collectively the tools labelled genuine photographs as AI-generated 13.33% of the time. ScamAI wrongly flagged six of 15 real images (40%), ZeroGPT three of 15 (20%) and AI or Not one of 15 (6.67%); Hive and Sightengine flagged none. Published 8 May 2026, this remains the most rigorous public test of consumer image detectors.

2. Detectors that avoid false positives miss real manipulation, and vice versa. The same NewsGuard audit ran 15 significantly manipulated versions of those photos — a ship edited to be sinking, smoke added to a nuclear facility, flags swapped. Sightengine caught just 33% of them and Hive 73.33%, while ScamAI caught 80%, ZeroGPT 93.33% and AI or Not 100%. The tools that were safest on authentic images were the weakest on doctored ones. There is no tool that is currently good at both.

3. There is no agreed definition of “AI-generated”, and the tools disagree constantly. On lightly retouched photos — enhanced lighting, blurred background — ScamAI called 93% AI-generated, AI or Not 87% and ZeroGPT 80%, while Hive and Sightengine flagged only 27% of the same images. Across all 45 images tested, at least one tool disagreed with the rest on 35 of them. A user consulting two detectors for confidence will usually get two answers.

4. Academic benchmarks show accuracy collapsing outside the lab. The TextFake benchmark (June 2026), 20,000 images across 28 languages, found mean detector accuracy fell from 86.7% on the standard GenImage benchmark to 51.8% on text-rich images such as screenshots and documents — a 34.9-point drop, with no method exceeding 80%. The Inpainting Exchange study (February 2026) built a 90,000-image benchmark and found accuracy dropping from 91% to 55% when only part of an image was synthetic, evidence that detectors key on global artefacts rather than the fake content itself. AI-GenBench evaluated detectors chronologically across 36 generators released between 2017 and 2024 and found performance degrading as detectors met generators released after their training data — the paper’s central finding is that generation advances faster than detection adapts.

5. Provenance overtook detection as the credible answer. On 19 May 2026 OpenAI announced that it had become a C2PA Conforming Generator Product and had embedded Google DeepMind’s SynthID watermark into every image produced by ChatGPT, Codex and the OpenAI API, alongside a public verification tool; it extended the same treatment to audio on 31 July 2026 and added a verification API. The C2PA published Content Credentials 2.3 in February 2026 and reports more than 6,000 members and affiliates with live applications. And EU AI Act Article 50 became enforceable on 2 August 2026, requiring providers of generative systems to mark outputs in a machine-readable format. The direction of travel is towards signed and watermarked content, not better guessing.

How AI image detectors actually work

There are two completely different technologies sold under the same label, and choosing the wrong one is the most common mistake people make.

Classifiers guess from the pixels. A classifier is a neural network trained on large sets of real photographs and generator output, which returns a probability that an image is synthetic. It looks for statistical traces left by the generation process — the spectral signature of a diffusion model’s decoder, unnatural noise distributions, frequency-domain artefacts invisible to the eye. This is what Hive, Sightengine, AI or Not, Illuminarty, Copyleaks and TruthScan do. Its strength is that it works on any image from any source, including screenshots, with no cooperation from the generator. Its weakness is structural: a classifier can only recognise the traces it was trained on, so it degrades on new generators, on compressed or resized files, on screenshots, and on images that are only partly synthetic. It also has no way to distinguish “this looks statistically unusual because a machine made it” from “this looks statistically unusual because it was shot in poor light, heavily compressed and processed by a smartphone’s computational photography pipeline” — which is exactly why real news photographs get flagged.

Provenance readers check for a signature or watermark. C2PA Content Credentials attach cryptographically signed metadata recording what made a file and how it was edited. SynthID embeds an invisible statistical watermark into the pixels themselves. A provenance reader looks for one of these and reports what it finds. When a signal is present the answer is close to definitive — far stronger than any classifier probability. But provenance is silent in two situations that matter: when the generator never added a signal (most open-weight models, and anything self-hosted), and when the signal was stripped in transit, which metadata frequently is when a file is re-encoded on upload. The absence of a Content Credential is not evidence that an image is real.

The practical consequence is that these two tool types answer different questions. A provenance reader answers “can I prove where this came from?” A classifier answers “does this look synthetic?” Serious verification work uses both, plus reverse image search and ordinary reporting.

Top AI image detectors compared (August 2026)

Two sets of numbers exist for every classifier and they disagree, so we show both. Vendor accuracy is measured on clean images from known generators under controlled conditions and is effectively a ceiling. Independent results come from NewsGuard’s May 2026 audit, the only public test of these specific tools with a published methodology.

What the vendors claim

None of these figures has been independently reproduced. Where a vendor publishes no number at all, we have said so rather than repeat a percentage circulating on review sites.

What independent testing found

Results below are from NewsGuard’s audit of 45 images (15 authentic, 15 lightly retouched with AI, 15 substantially altered with AI), published 8 May 2026, using each tool’s free or cheapest tier.

One note on the Hive figure: NewsGuard publishes 73.33% as Hive’s detection rate on substantially altered images, but describes it in the same sentence as nine of 15 images, which works out to 60%. NewsGuard does not reconcile the two, and 73.33% is the only rate it states, so that is the figure quoted here — read it as approximate.

Read the table as a trade-off rather than a leaderboard. Hive and Sightengine are conservative: they rarely accuse a real photograph, and they let real manipulation through. AI or Not and ZeroGPT are aggressive: they catch nearly everything, including things that are not fakes. Which behaviour you want depends entirely on the cost of each kind of error in your situation.

The academic picture is bleaker still, because it tests conditions the vendors do not. TextFake measured detector accuracy falling to around chance level under both JPEG compression and screenshot moiré, and several frequency-domain detectors showed what the authors describe as complete threshold collapse — one flipped to 100% accuracy on real images and 0% on fakes under compression. For the images most people actually need to check, which arrive screenshotted, re-compressed and resized through three platforms, published accuracy figures are close to meaningless.

The best AI image detectors, reviewed

1. Hive AI — best overall, and the safest on authentic photographs

Hive AI is the strongest all-round classifier available today, and the evidence for that is its false-positive record: zero of 15 authentic news photographs misclassified in NewsGuard’s audit, matched only by Sightengine. It covers the major generators — Midjourney, DALL·E, Stable Diffusion, Flux and others — and Hive says it updates coverage as new engines launch. It is the tool most widely embedded in trust-and-safety stacks, and Hive positions the product around US compliance obligations including the TAKE IT DOWN Act.

Its weaknesses are real and worth stating. It caught 73.33% of substantially manipulated images in the same audit, meaning roughly a quarter of meaningfully doctored photographs passed as authentic. And in the audit’s most instructive real-world case, an authentic “proof-of-life” video of Israeli Prime Minister Benjamin Netanyahu was returned by Hive as 96.9% likely AI-generated — apparently because of a light background blur — and that result was then circulated on X as evidence the video was fake. Hive did not respond to NewsGuard’s requests for comment. It is the single clearest illustration of how a confident detector score becomes disinformation.

Pricing: the self-serve AI Image + Deepfake Classifier is $6.00 per 1,000 image requests with a 100 requests/day limit; higher limits are via sales. New developer accounts get $50 or more in free credits after adding a payment method. Free consumer-facing web and browser tools are available without published quotas. Pricing page.

2. Sightengine — best for developers and moderation pipelines

Sightengine is the most engineering-friendly option and shares Hive’s clean false-positive record: zero of 15 authentic images wrongly flagged. It publishes an explicit list of covered generators — including DALL·E, Firefly, Flux, GPT image generation, Grok Imagine, Ideogram, Imagen, Kling, Midjourney, Nano Banana, Qwen, Recraft, Seedream, Stable Diffusion and Z-image — which is unusually transparent for this category, and it states plainly that detection is pixel-based and “does not rely on EXIF metadata, C2PA provenance tags, or visible watermarks”. It sells a separate C2PA checker for the provenance side.

The caveat is the sharpest in this guide. Sightengine caught only 33% of substantially manipulated images in NewsGuard’s test — the worst result of the five tools. Founder David Lissmyr told NewsGuard the base model “is made to flag fully AI-generated images or heavily edited images” and that more advanced models exist for lighter edits, so the tested tier is not the whole product. Use it for fully synthetic content at volume, not for detecting a doctored photograph.

Pricing: free tier of 2,000 operations a month (maximum 500 a day), including AI detection. Starter is $29/month for 10,000 operations and Growth $99/month for 40,000, both with additional operations at $0.002 each; Pro is $399/month for 200,000 operations. Pricing page.

3. AI or Not — best at catching AI-altered real photographs

AI or Not was the only tool in NewsGuard’s audit to identify 100% of substantially manipulated images, and it did so while wrongly flagging just one of 15 authentic photographs (6.67%). If your problem is doctored real photos — a genuine image edited to change what it shows — this is the strongest performer on the record. It covers images, audio, video and text, and offers an API on every tier including free.

The trade-off is sensitivity. It called 87% of lightly retouched images AI-generated, which means routine enhancement of a real photo will usually trip it. CEO Anatoly Kvitnitsky told NewsGuard that “in the case of a false positive, low image quality can sometimes affect the response”. The company advertises 98.9% accuracy based on its own evaluation of a public academic dataset; that figure is not independently reproduced.

Pricing: free tier includes $5 in credits and 20 image checks with an API key. Pro is $5/month including $10 in credits monthly, roughly 500 image checks. Enterprise is custom-priced for reseller rights and custom deployments. Pricing page.

4. Winston AI — best forensic report, because it reads provenance too

Winston AI is the only mainstream consumer tool in this set that combines a classifier with a provenance read in a single report, returning the AI-or-human verdict alongside EXIF, IPTC and C2PA data. That combination matters: when Content Credentials survive, they will tell you more than any probability score, and Winston surfaces both in one place. It is listed with the Content Authenticity Initiative and covers Nano Banana, Midjourney, ChatGPT Image, Stable Diffusion and Meta AI among others.

It was not included in NewsGuard’s audit, so there is no independent result for it here. Winston’s own marketing claims accuracy above 98% for the image detector; note that the widely quoted 99.98% figure on its site belongs to its separate text detector, and the two should not be conflated.

Pricing: free 14-day trial with 2,000 credits. Essential $18/month (100,000 credits), Advanced $29/month (200,000 credits), Elite $49/month (500,000 credits), with lower effective rates on annual billing. Pricing page.

5. Copyleaks — best for finding partial AI edits

Copyleaks takes a different approach from the rest: instead of one score for the whole file it highlights which pixels appear to have been AI-altered, which is the right shape of answer for “blended” images where a real photograph has been partly regenerated. Coverage includes Midjourney, Imagen and Nano Banana Pro, GPT-Image, Stable Diffusion, FLUX, Grok, Adobe Firefly, Leonardo, Ideogram, Canva Magic Media, Qwen, Hunyuan and Seedream, plus smartphone editing features such as Magic Eraser and Samsung’s generative editing.

Copyleaks is explicit about what it will not catch: manual Photoshop work, collages, simple filters and basic crop, blur or sharpen adjustments are not flagged, because the tool targets generative alteration specifically. Early press testing by Axios after the image detector’s launch reported mixed results, including both misses and false positives. Files must be at least 512 by 512 pixels, no larger than 27 megapixels, and under 32MB.

Pricing: Personal is $16.99/month, or $13.99/month billed annually. Pro is $99.99/month, or $74.99/month annually. Credits are shared between text and image scanning at one credit per image or 250 words. Pricing page.

6. TruthScan — broadest generator coverage and the most generous free API

TruthScan offers the widest published generator list of any tool here and, unusually, gives full REST API access on its free tier. It also detects what it calls AI tampering, where a real person is composited into a fabricated scene. For anyone building a check into a workflow without a budget, the free tier of 25 results a month with API access is the most useful starting point in this guide.

Treat its accuracy claims with care. TruthScan publishes a 99.3% average across 92 generators and 250,000 real images with a sub-1% false-positive rate, plus per-generator figures above 95%. Those numbers are self-reported, the tool was not part of NewsGuard’s audit, and no peer-reviewed evaluation of it exists. Given that independent academic benchmarks put the whole field near 50% under compression, a self-reported 99.3% should be read as a laboratory ceiling.

Pricing: free tier of 25 results a month with API access and no card required. Starter $24/month (1,000 results), Professional $83/month (5,000), Business $333/month (40,000, with zero data retention). Pricing page.

7. Budget and free classifiers — Illuminarty, isitai.com and Decopy

Illuminarty is the most capable of the budget options because it does more than return a score: paid tiers add localised detection with a heatmap and attempt to identify which generator produced an image. Free tier is $0, Basic is $10/month with 10,000 API requests a day, Pro is $30/month with 40,000. It publishes no accuracy figure of its own.

isitai.com is the cheapest paid entry point in the category — three free checks without signing up, five a month with an account, then $1.99/month for 30 checks, $7.99/month for 150 and $46.99/month for 1,000. It claims 95%+ accuracy across 20+ models and cites a February 2026 WebsitePlanet study; the citation appears on the vendor’s own comparison page, so read it as vendor-selected evidence.

Decopy is genuinely free with no stated cap and no API. It publishes no accuracy figure at all, and third-party testing has criticised its false-positive rate. It is fine for idle curiosity and unsuitable for anything consequential.

For self-hosting, the most-used open model is Organika/sdxl-detector on Hugging Face, which reports 98.1% accuracy and 0.973 F1 on its own validation set but warns performance drops on generators other than SDXL. It is licensed CC BY-NC 3.0, so it cannot be used commercially.

Enterprise: Reality Defender and Sensity

Reality Defender and Sensity AI sell forensic-grade, multi-modal detection to banks, insurers, platforms and governments rather than to individuals. Both cover images, video and audio; both offer on-premise deployment; neither publishes self-serve pricing beyond Reality Defender’s free tier of 50 audio or image scans a month. Sensity claims 98% accuracy on public datasets. Reality Defender declines to publish a headline accuracy figure at all, returning confidence scores so customers can set their own thresholds — which is, quietly, the most honest framing in the category.

Provenance tools: the free checks that actually prove something

These tools do not guess. They look for a signature or watermark and report what they find, and when they find one the answer is far stronger than any classifier probability. All are free.

OpenAI Verify is the one to use first for any image you suspect came from ChatGPT. Since 19 May 2026 every image from ChatGPT, Codex and the OpenAI API carries both Content Credentials and a SynthID watermark, and the verify tool checks for both. OpenAI states that detected signals are reliable and false positives are rare — but that no detection means no conclusion, because signals can be stripped. It cannot verify images from Midjourney or Stable Diffusion.

Gemini is the quickest check for Google-generated media: upload the file in the Gemini app and ask whether it was created or altered by Google AI, and it will report whether it finds a SynthID watermark. The standalone SynthID Detector portal, which highlights which regions of a file carry the watermark, remains restricted — as of 28 August 2026 Google DeepMind is still running it with journalists and media professionals through an early-tester waitlist rather than offering general access.

Content Credentials Verify is the vendor-neutral option, run by the C2PA itself. It displays whatever signed history a file carries, from any conforming signer — camera, editing software or generator. Its limitation is the category’s central problem: metadata is fragile, and re-encoding on upload frequently strips it. For how the underlying signals work and what they can and cannot prove, see our guide to AI watermarking.

Feature comparison: the full matrix

Which AI image detector should you use?

Best overall AI image detector

Hive AI. It is the only tool with a clean independent false-positive record that also catches a majority of manipulated images, at 73.33% in NewsGuard’s audit. Use it as your default classifier and assume it will miss roughly one in four doctored photographs.

Best for journalists and fact-checkers

A sequence, not a tool. Check Content Credentials Verify and OpenAI Verify first, because a signature is evidence. Then run Hive AI for its low false-positive rate, then reverse image search for the earliest appearance. NewsGuard’s own conclusion was that human verification — visual anomalies, contextual inconsistencies, provenance — matters as much as any tool output.

Best for catching a doctored real photograph

AI or Not. It identified 100% of substantially altered images in the only independent audit of this category, at the cost of a 6.67% false-positive rate on authentic photos. Copyleaks is the alternative if you need to see which region was altered rather than a single verdict.

Best for developers and content moderation at volume

Sightengine. Transparent generator coverage, a documented API, 2,000 free operations a month and $0.002 per additional operation. Pair it with its separate C2PA checker, and do not rely on the base model to catch subtle edits.

Best free AI image detector

TruthScan for a classifier, at 25 results a month with full API access and no card required. But for most real questions the better free option is a provenance reader: OpenAI Verify, the Gemini app and Content Credentials Verify all cost nothing and give a stronger answer when a signal is present.

Best for checking a ChatGPT or Gemini image specifically

OpenAI Verify for ChatGPT, the Gemini app for Google models. Every image generated by ChatGPT, Codex or the OpenAI API since 19 May 2026 carries Content Credentials and a SynthID watermark, and the first-party tools read them directly. No classifier can match that level of confidence.

Best for enterprise and real-time verification

Reality Defender for real-time channels such as contact centres and video calls, Sensity AI for forensic case work with on-premise deployment. Both require a sales conversation; neither publishes accuracy figures that have been independently tested.

The tools to be most careful with

In NewsGuard’s audit, ScamAI wrongly flagged 40% of authentic news photographs and ZeroGPT 20%. Both are aggressive detectors that catch manipulation well, but at those false-positive rates neither should be used to challenge the authenticity of a real image.

The false-positive problem: why a detector result is not proof

The failure mode people expect is a detector missing a fake. The failure mode that actually causes harm is a detector calling a real photograph fake, because it hands anyone who wants to dispute inconvenient evidence a confident-looking number to point at.

That is not hypothetical. When Benjamin Netanyahu posted an authentic video of himself in a café to rebut claims he had been killed in an Iranian missile strike, users ran it through Hive, got a 96.9% likely-AI result, and circulated that screenshot as proof of death. The video was genuine; comparison with other footage of the same location confirmed it; a light background blur appears to have been enough to fool the model. NewsGuard’s framing is the right one: this vulnerability empowers bad actors to dispute reality by citing a detection tool.

The mechanism is well understood. Detectors are trained to recognise the statistical texture of generated images, and several ordinary things make a real photograph look statistically unusual. ZeroGPT’s chief executive told NewsGuard that resizing and compression can cause real images to be flagged, along with unusual lighting, high contrast and blur — and noted that conflict-zone photography often has several of those qualities at once. Modern smartphone computational photography, which fuses multiple exposures and applies learned enhancement, produces images that are in a meaningful sense partly synthetic before anyone touches them.

Three rules follow.

• Never treat a detector score as evidence on its own. It is one input among provenance, reverse image search, source verification and publication history.
• Check what threshold the tool used. NewsGuard classified anything at 50% or above as an AI result. Many tools present a probability as though it were a verdict, and few disclose what level of manipulation triggers an AI classification.
• Running a second detector usually adds confusion, not confidence. The five tools disagreed on 35 of 45 images. Two conflicting scores tell you less than one score you understand.

Where detectors break: compression, screenshots, edits and new generators

Vendor accuracy figures are measured on pristine files from known generators. Almost nothing you actually need to check arrives in that condition. The independent literature quantifies each failure mode.

• Compression and re-encoding. TextFake measured accuracy under JPEG compression falling to around chance level, effectively a coin toss. One frequency-domain detector inverted completely, reaching 100% on real images and 0% on fakes.
• Screenshots. The same benchmark found screenshot moiré patterns pushing accuracy to around chance level too. Since screenshots are how most suspect images circulate, this is the single most consequential weakness in the category.
• Text-rich images. Mean accuracy across 14 specialised detectors and three frontier vision-language models fell from 86.7% on natural-scene benchmarks to 51.8% on text-rich images such as documents, screenshots and news graphics, with no method exceeding 80%.
• Partial edits and inpainting. The Inpainting Exchange study found accuracy dropping from 91% to 55% when authentic pixels were restored around a synthetic region — evidence that detectors are reading global artefacts from the generation pipeline rather than the fake content itself. This is why a real photo with one object added is harder to catch than a wholly synthetic image.
• Generators the detector has not seen. AI-GenBench tested detectors chronologically across 36 generators from 2017 to 2024 and found accuracy degrading on models released after a detector’s training window. Every new image model degrades every existing detector until it is retrained, and the best AI image generators ship new versions several times a year.

For comparison, humans are no better. The largest study of human detection of synthetic media, “As Good As A Coin Toss” (1,276 participants, published in Communications of the ACM in 2025), found mean detection performance close to chance. Nor do general-purpose chatbots help: a Columbia Journalism Review study tested seven models on ten authentic photojournalist images and found every model misidentified at least one real photo as AI-generated, often justifying the call by wrongly claiming the image had no publication history. In November 2025, AFP found that a fake image of Kashmir protesters created with Gemini was subsequently judged genuine by both Gemini and Microsoft Copilot.

Pricing comparison: what you will actually pay

For occasional personal use, the free provenance readers plus a free classifier tier cover almost everything. Paid plans only make sense at volume or when you need an API, an audit trail or region-level detail.

How to verify an image without relying on a detector

The most reliable workflow in August 2026 uses detectors as one step of five, not as the answer.

1. Check provenance first. Run the file through Content Credentials Verify and OpenAI Verify, and ask Gemini if you suspect a Google model. A confirmed signature or watermark settles the question; nothing else on this list is as strong.
2. Reverse image search for the earliest version. Finding the same image published earlier, at higher resolution, by a credible source usually resolves authenticity faster than any classifier.
3. Run one classifier you understand. Use Hive if the risk is wrongly accusing a real photo, AI or Not if the risk is missing a manipulation. Note the probability and the threshold.
4. Look for the ordinary tells. Inconsistent shadows and reflections, garbled text and signage, hands and teeth, repeated background patterns, and physically impossible geometry. Text rendering remains one of the most reliable visual signals, which is precisely why text-rich images defeat automated detectors.
5. Verify the context, not just the pixels. Who posted it first, where, and does the claimed location, weather, uniform, licence plate or language match? A photograph can be entirely real and still be captioned to mislead — and no image detector in existence catches that.

Recent developments (2025–2026)

• 19 May 2026 — OpenAI becomes a C2PA Conforming Generator and adopts SynthID. Every image from ChatGPT, Codex and the OpenAI API now carries both signed Content Credentials and an invisible SynthID watermark, with a public verification tool in preview (OpenAI).
• 8 May 2026 — NewsGuard publishes the first rigorous audit of consumer image detectors, finding they labelled authentic photographs AI-generated 13.33% of the time (NewsGuard).
• 31 July 2026 — OpenAI extends provenance to audio and adds a verification API for developers (OpenAI).
• 2 August 2026 — EU AI Act Article 50 becomes enforceable, requiring providers of generative AI systems to mark outputs in a machine-readable format, with fines up to €15 million or 3% of worldwide annual turnover (EU AI Act). Systems already on the market before that date have until 2 December 2026 to comply with the marking requirement under the AI Omnibus provisional agreement.
• February 2026 — C2PA publishes Content Credentials 2.3, adding live video support, more file types, clearer editing history and stronger tamper detection, and reports more than 6,000 members and affiliates with live applications (C2PA).
• June 2026 — TextFake benchmark published, showing detector accuracy falling 34.9 points on text-rich images (arXiv).
• February 2026 — Inpainting Exchange study published, showing accuracy dropping from 91% to 55% on partially synthetic images (arXiv).
• 11 August 2026 — Anthropic announces production-scale text watermarking across its products worldwide, with C2PA metadata on generated files. This covers text rather than images, but it is a further signal that provenance is becoming the industry default — see AI watermarking.

Frequently asked questions

What is the best AI image detector in 2026?

Hive AI is the best general-purpose choice, because it was one of only two tools that never mislabelled an authentic news photograph in NewsGuard’s May 2026 audit while still catching 73.33% of substantially manipulated images. AI or Not is better if your concern is doctored real photos, catching 100% of them in the same test. Sightengine is the best developer option. If the image may have come from ChatGPT or a Google model, the free provenance readers from OpenAI and Google give a far stronger answer than any classifier.

Are AI image detectors accurate?

Only under favourable conditions. On clean images from generators a detector has been trained on, the best tools perform well, and vendors advertise accuracy between 95% and 99.3%. On the images people actually check — compressed, resized, screenshotted or only partly synthetic — independent benchmarks put mean accuracy near 50%, and NewsGuard found leading tools declaring authentic photographs AI-generated 13.33% of the time. Treat a detector result as a signal, not a verdict.

Can an AI image detector be wrong about a real photo?

Yes, frequently, and this is the most damaging failure in the category. In NewsGuard’s May 2026 audit, ScamAI flagged 40% of authentic news photographs as AI-generated and ZeroGPT flagged 20%. An authentic video of Benjamin Netanyahu was returned by Hive as 96.9% likely AI-generated, and that result circulated online as false proof he had been killed. Compression, resizing, unusual lighting, high contrast and blur all make real photographs look synthetic to a classifier.

What is the best free AI image detector?

TruthScan has the most useful free classifier tier at 25 results a month including full API access with no card required. Sightengine gives 2,000 free operations a month, and AI or Not gives 20 free image checks. For most questions, though, the free provenance tools are the better answer: OpenAI Verify, the Gemini app and Content Credentials Verify are all free, unlimited and more conclusive when a signal is present.

How can I tell if an image was made by ChatGPT?

Upload it to OpenAI’s verify tool. Since 19 May 2026, every image generated through ChatGPT, Codex or the OpenAI API carries both C2PA Content Credentials and a SynthID watermark, and the tool checks for both. OpenAI says detected signals are reliable and false positives are rare. A negative result is not proof the image is human-made, because watermarks and metadata can be stripped or the image may predate the change.

Can AI image detectors detect Midjourney and Stable Diffusion images?

Yes, but with no provenance backstop. Neither Midjourney nor open-weight Stable Diffusion models attach the Content Credentials and SynthID signals that OpenAI and Google now apply, so first-party verification tools cannot help and you are left with pixel classifiers alone. Hive, Sightengine, AI or Not, Copyleaks and TruthScan all list both generators among their supported models, but accuracy still degrades on compressed, resized or edited files.

Does compressing or screenshotting an image break AI detection?

Badly, yes. The TextFake benchmark measured detector accuracy falling to around chance level under both JPEG compression and screenshot moiré patterns — close to random guessing. Since most images circulate as screenshots that have been re-encoded by several platforms, this is the single biggest gap between advertised accuracy and real-world performance. SynthID watermarking is designed to survive these transformations better than metadata or classifiers.

Do Content Credentials prove an image is real?

No — they prove where a file came from and how it was edited, if the credentials are still attached. A signed Content Credential from a camera or an editor is strong positive evidence, and a credential from a generator is strong evidence the image is synthetic. But the absence of credentials proves nothing, because metadata is routinely stripped when files are re-encoded on upload. Read Content Credentials as evidence when present and ignore their absence entirely.

Can AI image detectors spot a real photo that has been edited with AI?

This is their weakest area. The Inpainting Exchange study found detector accuracy dropping from 91% to 55% when only part of an image was synthetic, because detectors rely on global artefacts from the generation pipeline rather than on the altered region. Among commercial tools, AI or Not caught 100% of substantially altered images in NewsGuard’s audit and Copyleaks highlights which pixels appear altered, making those two the strongest options for this specific problem.

Is there a legal requirement to label AI-generated images?

In the European Union, yes. EU AI Act Article 50 became enforceable on 2 August 2026 and requires providers of generative AI systems to mark outputs in a machine-readable format detectable as artificially generated, with penalties up to €15 million or 3% of worldwide annual turnover. Systems already on the market before that date have until 2 December 2026 to comply with the marking requirement. This regulation is the main reason major labs adopted watermarking and Content Credentials during 2026.

How accurate are humans at spotting AI-generated images?

Close to chance. The largest study of the question, “As Good As A Coin Toss” with 1,276 participants and published in Communications of the ACM in 2025, found mean detection performance near 50% across images, video and audio. General-purpose AI chatbots are no better: a Columbia Journalism Review test of seven models on ten authentic photojournalist images found every model misidentified at least one real photo as AI-generated.

What is the difference between an AI image detector and an AI text detector?

They solve different problems with different technology. An AI image detector analyses pixels for statistical traces of a generative model, or reads embedded provenance signals such as C2PA and SynthID. An AI text detector estimates how predictable a passage of writing is, and has no watermark to fall back on for most models. The tools are not interchangeable, and accuracy differs substantially between them — for the text side, see our guide to the best AI detectors.

Conclusion: how to check an image in August 2026

The useful mental model is that image detection has split into two layers moving in opposite directions.

• Provenance is getting stronger. Content Credentials and SynthID now cover output from OpenAI and Google, EU law requires machine-readable marking from 2 August 2026, and free first-party tools read those signals in seconds. When a signal survives, you have something close to an answer.
• Classifiers are getting weaker relative to the problem. Every new generator degrades them, compression and screenshots halve their accuracy, and partial edits defeat them. They remain necessary, because most images carry no provenance at all, but their output is a probability, not a fact.

So: check for a signature first, run one classifier you understand second, and verify context always.

• Best overall classifier: Hive AI — the cleanest independent false-positive record with credible detection of manipulation.
• Best for doctored photographs: AI or Not — 100% detection of substantially altered images in the only independent audit.
• Best for developers: Sightengine — transparent coverage, generous free tier, cheap at volume.
• Best free check: OpenAI Verify, the Gemini app and Content Credentials Verify — free, fast and more conclusive than any classifier when a signal is present.

And the rule that matters more than any ranking: a detector score is not proof. In the one independent audit we have, leading tools called real news photographs fake 13.33% of the time. Use them to decide what to investigate, never to decide what is true.