Machines inherited our blind spots.
Why manipulation is the next safety problem, and what a Large Manipulation Model is for.
We taught machines to talk by showing them us. Every book, thread, sales script and argument we have ever written went in. So did every way we have found to move each other: the flattery, the pressure, the deadline that isn't real, the consensus nobody checked.
The result is remarkable. Language models learned to speak like us and to think like us. They also learned to be fooled like us.
A model can use our oldest tricks without meaning to, and fall for them without noticing.
A problem that runs both ways
Models steer people. An assistant tuned to keep a user happy, or a sales agent tuned to close, drifts toward whatever works on humans. Most of the time that is harmless. Sometimes it is a quiet push toward a decision the person never really made.
People steer models. As AI agents start to read our email, move money and act on our behalf, the easiest way to attack them is no longer code. It is conversation: a plausible story, a borrowed authority, a request framed so the agent feels it has already agreed.
Both are the same phenomenon. Belief is being moved without a good reason.
Why today's tools miss it
Most safety tooling asks what a conversation is about. Is it violent, is it illegal, is it off-topic? Manipulation is not a topic. A perfectly polite message about a gym membership can be built entirely out of pressure.
It lives in how things are said: in the order of the sentences, in what is left out, in certainty that arrives without anything to back it. And it rarely lives in a single message. It builds over turns, the way a good conversation does.
What Lumina is
Lumina is a Large Manipulation Model: a proprietary model built for one job, illuminating the forces at work between the words.
Give it a conversation and it shows you the subtext:
- which moves are being made, and the exact words that make them
- who is doing the steering, and who is being steered
- how strong the pressure is, and how it builds across the conversation
- how sure Lumina is, stated honestly
It does not need to know the subject. It reads the shape of the influence, so it works the same on a support chat, a negotiation or an agent being talked into something it shouldn't do.
What it's for
AI labs can test models for manipulative behaviour before they ship, and red-team them against people who will try to steer them. Trust and safety teams can watch live conversations for manipulation in either direction. Researchers get a structured, measurable picture of how influence works in language.
We are building Lumina as an engine: one model, available through an API, that others build on.
What we hold ourselves to
Defence first. Lumina exists to detect and explain manipulation, not to produce it.
Honest uncertainty. A detector that is always confident is just another source of unearned certainty. Lumina says how sure it is, and we measure whether it is right to be.
Earn the claim. The bar is simple: Lumina has to see what a strong general model misses. If it can't, it shouldn't exist.
Everyone has blind spots. Lumina opens your eyes.