AI Detection Research
In-depth analysis, benchmarks, and insights into AI content detection technology. Stay informed about the latest developments in identifying AI-generated content.
Pangram vs GPTZero vs ZeroGPT vs tropa-3: AI Detector Benchmark (September 2026)
We ran 2,572 texts through five AI detectors, including text from six humanizers. On raw AI text tropa-3 detected 95.6%, GPTZero 61.6% and Pangram 44.0%, and none of the three flagged a human text. Every text and verdict is public.
Introducing tropa-3: 30 Times Fewer False Positives
tropa-3 replaced tropa-2 on 24 September 2026. At a score of 90 it flags 0.1% of 239,490 held-out human texts, down from 3.02%, and it catches more paraphrased and humanized AI text. Full numbers and known weaknesses.
Apple’s iOS 27 AI Watermark: SynthID, Photos & Image Playground
Apple’s image watermark explained: what SynthID signals, how it differs from metadata, and what an AI image detector can actually verify.
How Much Text Does AI Detection Need? We Measured Our Own Floor
Measured on our previous model, tropa-2. We cut 40 known-AI passages to different lengths and scored each one. All were caught at 100 words; at 25 only 22% were. We had published 50 as the floor — here is the corrected number.
Is GPT-6 Astra Detectable? We Replicated Originality.ai's Study
Originality.ai reported 99.9% on 1,000 GPT-6 Astra responses three days after launch. We rebuilt the same design and ran it through our previous model, tropa-2: 100%, at their threshold and at our stricter live one. Why that number is the least interesting part.
How Claude Marks AI-Generated Content
Anthropic marks Claude output since 2 August 2026 and named the method on 14 August: a version of SynthID-Text. The C2PA file metadata is verifiable today; the text watermark needs Anthropic’s key and the API has not shipped — and a mark proves processing, not authorship.
EU AI Act Article 50: What It Means for AI Content Detection (August 2026)
Since 2 August 2026 the EU AI Act requires AI-generated content to be marked and disclosed. What Article 50 demands of providers and deployers — and why detection is the compliance backstop.
AI Text Detection Accuracy 2026: How Well Do Detectors Really Work?
We benchmarked 6 major AI text detectors on 50,000 samples. Here's the data on accuracy rates, false positives by writing style, and which tools actually perform.
AI-Generated Images: Statistics, Detection & Trends 2026
12.4 billion AI images are created monthly. Breakdown of generation models, use cases, misuse for disinformation, and how detection accuracy varies by image type.
AI Detection in Hiring: Screening AI-Written Resumes & Cover Letters
78% of job applications contain AI content. How HR teams are responding, which policies work, the false positive problem for non-native speakers, and what fair screening looks like.
Deepfake Statistics 2026: Growth, Threats & Detection
The definitive analysis of deepfake proliferation covering growth trends, industry impacts, detection rates, and what organizations need to know to protect themselves.
AI Insurance Fraud Statistics 2026: The Growing Threat
How artificial intelligence is enabling sophisticated insurance fraud at unprecedented scale, and why traditional detection methods are failing.
AI-Generated Fake Reviews: How to Spot Them in 2026
Learn how to identify AI-generated fake reviews on Amazon, Google, Yelp and other platforms. Statistics, detection methods, and tools for spotting synthetic reviews.
AI Romance Scams: How Deepfakes & Chatbots Target Dating App Users
Romance scammers are weaponizing AI to create convincing fake personas at scale. Learn how AI-generated photos and chatbots are deceiving millions.
Content Authentication: What's Coming in 2026
Predictions and analysis of AI detection trends, watermarking standards, regulations, and technology developments shaping content authentication this year.
AI Detection Accuracy: Comparing Detection Methods in 2026
A comprehensive analysis of different AI detection approaches, their accuracy rates, and why some methods outperform others across text, image, and audio content.
How Google's SynthID Watermarks AI Content
Deep dive into Google DeepMind's invisible watermarking technology that embeds undetectable signatures in AI-generated images, audio, and text.
The Arms Race: AI Generation vs Detection
How AI generators and detectors are evolving together in an endless cat-and-mouse game. Analysis of techniques, countermeasures, and where the battle is heading.
How to Detect AI Voice Cloning & Audio Deepfakes
Learn how AI voice cloning works, the telltale signs of synthetic audio, and how detection technology identifies fake voices in calls, podcasts, and media.
AI Detection for Academic Integrity: A Guide for Educators
How universities and schools use AI detection tools to maintain academic integrity. Best practices, limitations, and fair policies for handling AI-generated student work.
GPTZero vs ZeroGPT vs tropa-2: AI Detector Benchmark (August 2026, previous model)
The August 2026 edition, run with our previous model. GPTZero, ZeroGPT and tropa-2 on the same 522 texts. Which tools catch GPT-5.6, Claude Opus 5 and Gemini 3.6 — and which ones accuse human writers. Every text and score is public.
Introducing tropa-2 (previous model): Humanizer Detection and Hidden-Character Defense
Our previous model, released August 2026 and replaced by tropa-3 in September. tropa-2 was built for humanized text and hidden-character tricks, and at release caught 97% of raw AI text while wrongly flagging one human text in two hundred.
Introducing tropa-1: 87% Fewer False Accusations, Half the Errors
Our first model, released July 2026 and still available in the API. At release, false positives on verified human writing dropped from 9.7% to 1.2%, overall errors were cut in half, and humanized AI text was caught more often at every operating point — full benchmarks and honest caveats inside.
Stay Updated on AI Detection
Get notified when we publish new research, benchmarks, and insights about AI content detection.
Try Our Detector