
A frontier AI lab trains on real conversations, with consent on every record.
Naturally occurring human conversation, contributed with consent and documented record by record.
The data
Personal messages from messaging platforms and apps, contributed by thousands of people.
A leading San Francisco-based AI lab training conversational models already used synthesized datasets, human-simulated dialogue and structured corpora. It wanted to add naturally occurring human conversation.
What people agreed to
Each contributor chose what to share and removed anything they wanted kept private. An automated check then flagged any remaining identifiers.
Contributors connected their message data through the consent layer in a single permission step. The automated check is AI-assisted and runs after each contributor’s own review.
The record
Consent and redaction are documented for every record. The lab could see where each record came from and how consent was obtained.
Provenance can be audited end to end.
What it made possible
The dataset passed the lab’s internal compliance review, which its synthetic and simulated sources had not.
Naturally occurring human conversation now sits in the lab’s training mix, with a consent record behind every item.