A pair of developers launched a web tool that probes how much personal data survives inside large language model weights. They built the service publicly over several weeks, enabling users to submit a query that runs against both frontier‑scale and smaller models in parallel. The system clusters the models’ responses and reports a confidence score indicating how strongly each model recognizes the input.

The motivation stems from growing traffic shifting from traditional web interfaces to AI chat endpoints, raising questions about the digital footprints left in model parameters. By comparing frontier models with more lightweight variants, the creators illustrate that even compact models retain recognizable patterns tied to individual queries. The clustering step surfaces similarities, helping users gauge exposure across the model spectrum.

The tool invites community feedback through the original Hacker News post, where it has already gathered over a hundred upvotes and dozens of comments. Its open design lets anyone experiment with multiple models without writing custom integration code. By making weight‑level recognition visible, the project adds a practical lens for privacy researchers probing AI’s latent memory today.