Deckard Uses a Local Model to Detect AI-Generated Text in Your Browser

Summary

The article introduces Deckard, a Chrome extension designed to scan text on websites automatically and flag writing that may have been generated by AI. It runs a small local model on a Mac, avoiding the need to send browser text to a third-party service. The author benchmarked several local detection models on a combination of AI-detection datasets. The best-performing listed model, Gradient running in 4-bit MLX format, falsely flagged 2.712% of human text and detected 52.35% of AI-involved text; the other models generally performed worse. The author notes that these results are substantially below Pangram's claimed 99.66% detection rate and 0.004% false-positive rate, while Pangram's larger models were not suitable for continuous background use. Deckard uses Chrome native messaging to start the local model when needed, consumes about 400 MB to 1.2 GB of memory while active, and shuts down after five minutes without use. In the author's testing, it identified known AI content such as YouTube's AI summaries and AI snippets in the author's own posts. The author considers it useful for raising suspicion, but emphasizes that its roughly 2% false-positive rate means a single flag is not proof. They expect smaller local detectors to improve as models advance and plan to replace Deckard's model when a stronger one becomes practical.