Does Shellf train AI on my reading data?

No. Shellf doesn’t train any AI or machine-learning model on your reading data. There is no training pipeline in the codebase — only inference. Your data is used to compute recommendations and stats for you, and that’s it.

There’s a useful distinction between training a model on data and using a model to infer something for you. Training means feeding data to a model so the model’s weights change — the model literally learns from your data and gets better at the task across all users. Inference means running an existing, fixed model against your data to produce a specific output (a recommendation, an explanation, a category guess).

Shellf does only the second. When you ask for recommendations, your library is fed to the recommendation pipeline and the output is computed for you. None of the underlying models change as a result. They are pre-built models that Shellf uses; Shellf doesn’t update them.

Shellf calls external AI services for inference — a text-embedding model to map books and your taste into a structured representation, and an LLM to write the short explanations that come with each recommendation. The providers Shellf uses are configured for inference-only use:

Building your library, generating your recommendations, computing your stats, and remembering what you’ve already seen so the engine doesn’t resurface the same titles. All of this is your-account-only.

We also use all users’ aggregated stats (e.g. number of books read) to show comparison stats to people — such as how your number of books read compares to others.

Explicitly not used:

If at some future point the policy changes in order to do things like show ads, share data, or make things public by default, I will make it an explicit opt-in policy.

And if you ever want to walk away entirely, the full CSV export and account deletion are designed to be friction-free.