Every detector vendor says it is 99% accurate. That number only means something on a test set someone else built.
So we ran WasItAIGenerated V3, the model behind our image detector and our image detection API, on the five public benchmarks Pangram used when it launched Pangram Image 1.0. Same test sets, same settings. Then we added what public benchmarks miss: images from social media and the newest generators.
None of the test images were used to train our model.
TL;DR
- Public benchmarks: V3 matches or beats Pangram Image 1.0 on all five. On Mirage-Test (49,000 images) V3 reaches 99.6%, Pangram 99.5%.
- Open models: V3 reaches an AUROC of 0.996. The best of nine open-weights detectors reaches 0.914.
- New generators: V3 caught 78 of 79 ChatGPT Images 2.6 images on release day, and 99.5% of the images from 19 current generators. ZeroGPT (TruthScan) caught 97.2%.
- Compressed copies: after WhatsApp and Instagram, V3 still catches 96.2% of AI images.
Five public benchmarks
These are the five benchmarks Pangram reported for Pangram Image 1.0, the strongest published results of any commercial detector. We ran V3 on the full test set of each one, with the same metric and the same threshold (0.5). The chart at the top of this page shows the result, the table below has the details. "Best academic" is the best academic detector on each benchmark, as listed by Pangram.
| Benchmark | Metric | WasItAIGenerated V3 | Pangram Image 1.0 | Best academic |
|---|---|---|---|---|
Mirage-Test 49,000 images · Images from 5 content domains (people, animals, objects, scenes, anime) | Accuracy | 99.6% | 99.5% | 88.4% OmniAID-Mirage |
X-AIGD 2,419 AI images · Current text-to-image generators | Recall (AI caught) | 100% | 99.96% | 95% Gram-Net |
AIGenImages2026 1,118 images · 2026 generators, official test split | Accuracy | 99.7% | 99.6% | 92.1% RINE |
Synthbuster + RAISE 2,800 images · Diffusion images vs. raw camera photos | Balanced accuracy | 99.7% | 99.5% | 94.9% B-Free |
Community Forensics (eval) 51,836 images · Thousands of generators, held-out eval set | Macro accuracy | 99.3% | 99.3% | 89.2% Community Forensics-384 |
Nine open-weights detectors
We ran V3 and nine freely available detectors on 4,480 images from six further public benchmarks. None of them was trained on these images. The score is AUROC: 1.0 means a detector separates AI from real perfectly, 0.5 is a coin flip.
The newest generators
Benchmarks lag behind. By the time one is published, newer generators are already out. So we ran two checks on what people post right now.
On the day OpenAI released ChatGPT Images 2.6, V3 caught 78 of 79 test images, although it was trained before 2.6 existed. And on our own test set of 19 current generators:
The 19 generators include GPT-Image-2, Nano Banana 2.1 and Pro, FLUX.3, Qwen-Image 3 and Seedream 5 Pro. 60% of this test set is public. The full results are in our frontier benchmark.
Internal test: 623 real-world images
Benchmark images are mostly clean originals. The images a fraud, trust & safety or OSINT team gets are not: they have been posted, compressed and cropped.
So we also ran an internal test on 623 images that were posted on Reddit and social media: 343 AI images that could pass as a photo and 280 real ones. V3, Pangram Image 1.0 and ZeroGPT got the same files. ZeroGPT uses TruthScan for image detection.
Where both gave a verdict, V3 and Pangram are level, in errors (26 and 21 on 582 images) and in false alarms (11 and 10 of 275 real photos). Against ZeroGPT, V3 made 3.7 times fewer errors: 25 against 93 on 547 images.
TruthScan, AI or Not, Sightengine and Winston AI
We have not tested AI or Not, Sightengine or Winston AI ourselves. These are the scores Pangram published for them in October 2026, on Pangram's own test of 1,200 AI images from 12 current generators and 100 real photos. Pangram reports 99.96% for its own model. That test set is not public, so we could not run V3 on it.
TruthScan is the one we did test ourselves, through ZeroGPT: see the 19 current generators and the internal test above. For text, we ran Pangram, GPTZero and ZeroGPT ourselves in our AI text detector benchmark.
After WhatsApp and Instagram
Most images reach a newsroom or fact-checker as a compressed copy, not as the original. On copies of our frontier set that went through WhatsApp, Instagram and re-sharing, V3 still caught 96.2% of the AI images at 1% false positives. The best open-weights detector dropped to 37.3%.
How we measured
- Public benchmarks: the complete official test split of each one, no subsets, at threshold 0.5. That is the setting Pangram uses.
- Internal test on 623 images: V3 at its production threshold (a score of 0.8 or higher counts as AI). Pangram through its web app and ZeroGPT through its web tool, in November 2026. The same files for all three.
- Always the production model, as it runs on our website and API. No tuning for any benchmark.
- No image from any of these test sets was used in training.
- Pangram's own numbers, the academic results and the scores for TruthScan, AI or Not, Sightengine and Winston AI on Pangram's test set come from Pangram's announcement of 4 November 2026.
These are our own measurements. If you want to check them, we fund independent benchmarks with free API credits.
FAQ
Which AI image detector is the most accurate in 2026?
On five public benchmarks, WasItAIGenerated V3 matches or beats Pangram Image 1.0, the strongest published commercial detector: 99.6% on Mirage-Test, 100% on X-AIGD, 99.7% on AIGenImages2026 and on Synthbuster, 99.3% on Community Forensics. Against nine open-weights detectors it reaches an AUROC of 0.996; the best open model reaches 0.914.
WasItAIGenerated vs Pangram: which AI image detector scores higher on public benchmarks?
WasItAIGenerated V3 is ahead on four of the five and level on the fifth: Mirage-Test 99.6% against 99.5%, X-AIGD 100% against 99.96%, AIGenImages2026 99.7% against 99.6%, Synthbuster + RAISE 99.7% against 99.5%, Community Forensics 99.3% for both. Pangram's numbers are the ones Pangram published. We measured ours on the same test sets with the same threshold.
How accurate are TruthScan, AI or Not, Sightengine and Winston AI?
We have only tested TruthScan ourselves, through ZeroGPT, which uses it for image detection: on images from 19 current generators it caught 97.2% of the AI images at 1% false positives, against 99.5% for WasItAIGenerated V3. For the others we can only cite Pangram, which measured them on its own test set in October 2026 (balanced accuracy): TruthScan 96.3%, AI or Not 91.4%, Sightengine 88.3% and Winston AI 80.3%.
Are these numbers measured on the full test sets?
Yes. Every public-benchmark number is on the complete official test split, at the same threshold (0.5) Pangram uses. None of these images were used to train our model.
Does the detector work on brand-new generators?
Yes. On the day OpenAI released ChatGPT Images 2.6, V3 caught 78 of 79 test images, although it was trained before 2.6 existed. On our test set of 19 current generators it caught 99.5% of the AI images at 1% false positives.
What about images shared on WhatsApp or Instagram?
After WhatsApp, Instagram and re-sharing, V3 still caught 96.2% of the AI images in our frontier test set, at 1% false positives.
Check an image yourself
The detector on this page is the one running on WasItAIGenerated and in our API.