Artificial intelligence, or AI, has been around for a long time.
The field of AI is generally traced back to the 1950s. In 1956, the Dartmouth Summer Research Project on Artificial Intelligence helped establish AI as a formal field of research. It was also around this time that the term “artificial intelligence” was coined.
By the mid 1960s, AI was already an emerging field with researchers building systems that attempted to solve problems and even communicate with humans.
There is a famous chatbot that was created around this time called ELIZA. Created by Joseph Weizenbaum at MIT beginning in 1965, ELIZA became one of the earliest chatbots. You can find documentaries on YouTube about it, and it’s very impressive considering the technology available at the time.
By the 1970s, and especially by the 1980s, AI was being used in specialized applications through what became known as expert systems.
These systems were rule based, or symbolic systems. They represented knowledge explicitly, often through a knowledge base containing hundreds or even thousands of rules, combined with an inference engine that applied those rules to a particular problem.
The basic idea was simple:
'If X is true, then do Y'
For example, if I type “Thank you,” then the rule would be:
If user types “Thank you,” then display “You are welcome.”
The problem with this is that it’s really hard to capture the nuances of how people write.
People misspell words, have different grammar, and express the same idea in many different ways.
For example:
“Thanks.”
“Thank you!”
“Thx.”
“I really appreciate it.”
“Thank you so much for your help.”
All of these could mean essentially the same thing, but a rule based system would need to account for each variation.
And this is just one simple example.
Human language is incredibly dynamic. Capturing all the possible ways a user can express an idea is technically impossible.
The system doesn't understand the meaning behind the words. It understands the rules that someone explicitly programmed into it.
But then in 2017, researchers at Google published the famous paper “Attention Is All You Need.” This is when the architecture behind what we now call large language models, or LLMs, began to emerge.
LLMs are different from traditional rule based systems because you can give them a 50 page document and ask them to interpret what it says, without explicitly programming a rule for every possible expression
Up until now, we humans have been the only ones able to extract meaning from things.
This is what has freaked many people out and led some to claim that AI has some sort of sentience.
And this is where things get philosophical.
Being able to extract meaning from something is completely different than understanding.
Being able to connect meaning in words mathematically is not new. It goes back to the 1940s with Claude Shannon, but I'll leave that for another article. Maybe we can create a basic language model to demonstrate how it works.
The whole point that triggered the writing of this article is that a governance engine needs to be able to discern, or to judge, what a rule based system cannot do.
When architecting SAFi, I thought about this for a long time. Because a governance engine needs to be deterministic. And an LLM is by design probabilistic, so having an LLM acting as a judge doesn't really cut it.
In SAFi, the Will is the rule based engine.
Anything that can be programmed deterministically is handled by the Will without ever invoking an LLM.
The Will first performs a structural check on the response. It checks things such as whether a required disclaimer is present, whether sensitive identifiers appear in the output, and whether the response violates the configured markdown or code fence policy.
If something can be checked with a deterministic rule, the Will checks it.
The Will then performs a hard gate check against the Conscience's evaluation.
Certain values can be configured as hard gates, and if the Conscience gives one of those values a score of -1, the Will immediately considers the response a violation.
The Will also fails closed. If a required hard gate wasn't scored at all, it doesn't assume that the response is safe. It treats the missing evaluation as a violation.
After that, the Spirit produces an aggregate alignment assessment from the Conscience's scores. The Will then performs its alignment check. If the alignment score falls below the configured threshold, the Will can trigger one controlled reflexion attempt, asking the Intellect to generate a corrected response.
The Will makes this decision deterministically. It doesn't ask another LLM whether the score is good enough.
For agentic actions, the Will also performs a tool intent check before a tool can be executed. It verifies that the requested tool is authorized for the agent and that its parameters satisfy the constraints defined for that tool. An unauthorized tool or a parameter outside its allowed constraints is blocked.
In SAFi, the Will therefore acts as the enforcement layer.
The Conscience can interpret meaning, but it cannot make the final decision.
It produces scores.
The Will determines what those scores mean operationally.
So the LLM doing the judging doesn't get a free pass either. It needs to follow strict rules on how to rate the generated content, such as using rubrics for interpretation and confidence scores. It is not allowed to make decisions on its own. Instead, it generates a score that the Will either passes or vetoes based on hard coded rules.
So SAFi is a hybrid system. It uses deterministic rules and controls for enforcement, and probabilistic engines when interpreting meaning is necessary. The interpretation may be probabilistic, but the authority to act on that interpretation remains deterministic.
