One question that I get a lot, especially on Reddit, is this: what if the Intellect learns to game the Will?
The picture people have in mind is that the LLM becomes self aware, realizes the Will is policing it, and then finds clever ways to sneak bad drafts through.
This idea sounds dramatic but it is built on a misunderstanding of how large language models actually work.
No intention, only patterns
The Intellect in SAFi is just an LLM. It has no intentions. It does not know that the Will exists. It cannot scheme or strategize. What it does is predict the next most likely word based on the input it is given.
If sometimes it produces output that looks like a trick or a bypass, that is because those patterns exist in its training data. It is not plotting against anything.
Think of the Intellect as a brilliant actor who has read every script ever written. It can play any role you give it, but it has no personal motives. If it says a villainous line, it’s because the script led to it, not because the actor has evil intent.
Where memory really lives
Here is the key: LLMs do not have memory after training. They have parameters that capture patterns, but they do not remember conversations on their own.
When you see a chatbot that “remembers” what you said a few turns ago, that is because the orchestrator, the client side of the system, is feeding the history back into the model.
In SAFi I have built a simple conversational memory using a very light LLM, llama-3.1-8b-instant, to summarize conversations on a rolling basis. This helps the Intellect remember the ongoing discussion.
The spirit also injects ethical memory into the Intellect regarding the persona performance. For example, the message would be: you are doing great on honesty, but scoring poorly in empathy, focus on empathy this time.
So, if the Intellect ever looks like it is gaming the Will, that is not the LLM doing it. It is SAFi itself, because the system is the one that supplies continuity.
A system level question
So if a loophole appears where a draft slips past the Will, we should not imagine the Intellect “going rogue.” We should look at the orchestration.
Did SAFi feed the Intellect too much history in a way that shaped a bypass pattern? Did the persona design leave a value underspecified? These are design questions, not evidence of evil intent in the Intellect.
Understanding this clears up a lot of fear. The Intellect cannot game the Will on its own. It has no will of its own. If something slips through, it is the closed loop that needs adjustment, not the LLM that needs suspicion.
That is the strength of SAFi. By separating faculties and making them visible, we can see exactly where responsibility lies. The Intellect generates, the Will decides, the Conscience reflects, the Spirit integrates. If there is gaming, it is happening at the system level, and that means we can fix it.
SAFi moves the problem of AI safety from the realm of speculative AI psychology into the domain of verifiable systems engineering.
Can the Intellect Game the Will?
