See the forces working between the words.
Lumina is a Large Manipulation Model. It reads conversations between people and AI and illuminates the subtext: the places where belief is being pushed without a reason.
Why this exists
Language models learned to speak like us, think like us, and be fooled like us.
They were trained on everything we've written, including every way we've found to steer each other. A model can use those patterns without meaning to, and it can fall for them without noticing.
Most safety tools look at what is being talked about. Manipulation hides in how it's said: in the order of things, in what's left out, in certainty that arrives without a reason.
What's happening in the last reply
Across the conversation
The Large Manipulation Model
A proprietary model trained to see what we can't.
Lumina doesn't judge topics or keywords. It illuminates the mechanics underneath the words: the pressure, the framing, the borrowed certainty.
Reads the subtext
Points to the exact words where influence turns into manipulation.
Follows the whole conversation
Manipulation builds over turns. Lumina reads the arc, not just the last message.
Works in both directions
Catches models steering people, and people steering models.
Says how sure it is
Every flag comes with honest confidence, so you know when to trust it.
Blind spots
Everyone has blind spots. Lumina helps you see yours.
The same blind spots that make people persuadable are now built into the models we talk to every day. Lumina makes them visible, so they can be measured, tested and fixed.
Read the thesisWho it's for
For the teams responsible for what AI says, and what it believes.
AI labs
Test models for manipulative behaviour before they ship.
Trust and safety
Watch live conversations for manipulation, in either direction.
Researchers
A structured, measurable picture of how manipulation works in language.
See what your models can't.
We're opening Lumina to a small group of AI labs, safety teams and researchers.