Social products break when a model hides or promotes a post and nobody can say which policy, which signal, or which review produced that outcome. Ranking, recommendations, and moderation assistance all need a record, a fallback to a person, and a hard limit on what the system will not decide alone.
What makes AI work different here
- Moderation needs a reason. Removing or downranking a post without the policy clause and the signal that triggered it creates an appeals pile you cannot answer.
- Ranking is a product decision. What the feed optimizes for, time, replies, or something else, has to be chosen and measured. A model should not invent the goal.
- Safety edge cases go to people. Self-harm, harassment, and child-safety reports are not a confidence-threshold problem. They route to a trained reviewer.
- Personal data is the product's fuel and its risk. Training on private messages or contact graphs needs an explicit scope. We do not treat the whole social graph as free model input.
Where we typically start
We start by listing the decisions the product wants automated, the ones that must reach a person, and the log an appeals reviewer will read. The published social study is a moderation queue that cites the policy clause. Account bans stay with a person.
Services we typically provide
Case studies from this industry
Frequently asked questions
Can the model ban a user on its own?
No. It can classify a defined policy label and queue the item. Account bans and the hard safety cases stay with a reviewer.
Will you train on private messages?
Only if that data is in the written scope, with retention and access agreed first. It is not a default.
Do you have a published social case study?
Yes. The social study is a moderation queue that cites the policy clause and leaves account bans with a reviewer.