Local AI Detection Powered by Gemma: Imbue's Bouncer Brings On-Device Content Filtering
Key Takeaways
- ▸Bouncer's AI detection models achieve 99.8% accuracy for human vs. AI-generated text discrimination, vastly outperforming zero-shot LLM prompting approaches that perform barely better than random
- ▸iOS users can run detection entirely locally on devices using Google's Gemma 4 E2B, eliminating privacy concerns associated with cloud-based content analysis
- ▸A three-way classifier (human/AI-edited/AI-generated) reaches 91.7% accuracy, capturing real-world scenarios where content mixes human authorship with AI paraphrasing
Summary
Imbue has announced AI text and image detection capabilities in its Bouncer app, with a privacy-preserving iOS implementation powered by Google's Gemma model. The feature addresses user demand for filtering AI-generated "slop" from feeds, moving beyond simple keyword filtering to machine learning models trained specifically for detection accuracy. Imbue achieved 99.8% accuracy in distinguishing human from AI-generated text and 91.7% accuracy in three-way classification (human, AI-edited, AI-generated) using a fine-tuned Qwen 3 4B backbone trained on 28,000 texts across multiple domains including Amazon reviews, Reddit, news articles, and Twitter.
The iOS implementation is particularly significant: detection runs entirely on-device using Gemma 4 E2B, eliminating the need to send content to cloud servers. To work within iOS memory constraints, Imbue solved a critical engineering challenge by sharing a single LLM backbone across both AI detection and existing content filtering functions. This approach demonstrates how efficient open-source models like Gemma can enable sophisticated AI capabilities on consumer devices while preserving privacy and avoiding cloud dependency.
- The model achieves 96.4% accuracy on the third-party RAID benchmark despite being trained on completely different models, indicating strong generalization capability
