Journal entries are never available for model training
ContextThe journal is where people write what they would tell no one else. Reflection features need a model to read an entry at the moment of use. Training on entries would improve those features — and would turn the most sensitive data on the platform into a product input.
Options considered- A
Train on entries from users who opt inRejected
Consent given in distress is not a reliable basis for reuse. - B
Train on a de-identified journal corpusRejected
Free-text journals cannot be dependably de-identified. - C
Inference-only access, encrypted end to end, nothing retainedChosen
The feature works; the entry stays the person's.
Trade-off acceptedReflection quality improves more slowly, from licensed and synthetic data only. In exchange, the platform can state its privacy commitment without qualification.
- Owner
- Product engineering lead
- Reviewed by
- Foundation privacy owner
- Linked
- Privacy-first architecture · User-controlled sharing