HVMAN is an iOS app that analyzes text for human or generative model provenance. It's powered by its own analysis model that runs entirely on your phone. It's fast, local, and rivals Pangram.
You can download the app for free today via TestFlight. Please send in feedback. Your opinion helps shape the final version.
After the beta release, the app is freemium — $9.99 to unlock forever.
The app surfaces your full history, offers three themes and other customization settings, and provides detailed breakdowns of the content scanned.
Background
The idea for the app came from an investigation into what exactly makes today's leading AI models — GPT-5.6, Claude 5, Grok 4, Gemini 3, etc. — so bad at writing. While looking at hundreds of fiction and nonfiction outputs, it seemed possible to build a small model to notice what I did: frequent formulaic copy.
The app currently uses the v0.10 series of the analysis model, which despite the name, is the 16th model created for this project. The first versions through v0.9 were all trained entirely within ChatGPT's GPT-5.6 Pro container. I built the dataset, researched techniques for a small model, and asked GPT to handle the rest.
GPT Pro download for human-ai-model-attributor v0.10c after a 114-minute run
Generally, GPT Pro is exceptional at performing this work, often working for several hours per training run. I say generally because there were several significant issues encountered during this work, including random stops, disappearing final files that never download, data inconsistency across runs, and insane architecture sprawl decisions that rival the Winchester Mystery House.
Model
The analysis model is trained on over 56k human and model-generated documents. After the first ten iterations, I adopted a teacher-student paradigm to distill the model, lowering the file size and optimizing on-device inference speed.
On internal evaluations, the model is 99.48% accurate at detecting generated content with a human-written content false positive rate of 1.44%, making it not only a very strong model, but likely the best analysis model that runs on iPhone.
The model also correctly detects 93.94% of mixed AI and human content, with 85% accuracy at the individual 8-word passage level. Additionally, unlike other content detection systems, the model also identifies the likely model family and lab 83.4% of the time, and includes the correct model family among its top three predictions 96.06% of the time.
You can see cross-lab distillation, especially by SpaceXAI in Grok 4.5 and Chinese-lab models, which makes Top 1 accuracy harder.
Experiment tracker for E0 through E4 with progress and next actions
Later models were trained on my Apple M2 MBP and M1 server. Codex orchestrated the work. Documentation and plans were created by GPT-5.6 Sol, while training runs were orchestrated by GPT-5.6 Luna Max. Luna is good at monitoring runs.
Discussion
Using the app for the past week helped me hone my sense for non-human content. AI content is fine. However, many writers and YouTube creators seemingly gave up on having their own point of view and voice in favor of using Claude as their writer. There are a few holdouts, including Ben Thompson's Stratechery.
I don't mind AI content. I just don't prefer it to good human writing, and I prefer it even less when someone positions it as their own.