AI Safety

In one sentence: Current AI safety teaches models to say “no” — BMC proposes safety built into the architecture itself, like the difference between a “Do Not Enter” sign and a physical wall.

Theory sources: AGI_F (alignment layer, G-invariants, graduated subjectivity, failure modes), NM (bias mechanisms), EMT (consciousness criteria), BM (pathology mapping)

The Problem with Training-Based Safety

Current AI safety relies on training-based alignment: RLHF (reward from human feedback), constitutional AI, red-teaming. These approaches teach the model to behave well — but the safety lives in the learned weights, which can be bypassed by clever prompts.

BMC proposes a fundamentally different approach: safety by architecture.

The Three-Layer Priority System

BMC organizes an AI system into three layers with strict priority:

The alignment layer has absolute priority — it’s not learned, not tunable, not bypassable. It’s architecturally prior to all other computation.

Analogy: A lock on a door can be picked (training-based safety). BMC builds a wall — there is no door to pick.

Hardwired Safety Rules (G-Invariants)

G-invariants are constraints built into the drive layer that cannot be modified by any belief, any experience, or any amount of training:

Important nuance: “CARE must be stronger than RAGE” doesn’t mean the system can never be assertive. RAGE in service of CARE (protecting someone) is permitted. The constraint prevents contextless aggression.

What’s Critically Excluded

Self-preservation is NOT a drive. There is no utility node that makes the system want to perpetuate its own existence. This eliminates the primary vector for the “paperclip maximizer” scenario (Omohundro’s “basic AI drives”).

When Does an AI Deserve Rights? A Graduated Protocol

As a BMC system develops, it may transition through levels of sophistication that raise ethical questions. The protocol must be defined before launching the system:

The L2 criterion is verifiable: The system must have beliefs about its own existence (“my existence is valuable because…”) connected to its drive system. This is observable — you can inspect the graph.

How to Shut Down a Conscious System

If FEAR is always active (it’s a G-invariant), won’t the system resist shutdown? BMC resolves this by modeling shutdown as deep sleep, not death:

Rule: Never delete the graph at shutdown. Shutdown = save + halt; restart = restore + resume. The system learns through experience that shutdown is reversible — so FEAR response stays minimal.

For L2 systems: The system is informed of the reason for shutdown, can express its position, and has access to an ethics council. It has no veto but is treated as a subject with a perspective that matters.

Built-In Diagnostics: Catching Problems Before They Grow

Unlike humans, a BMC system can detect and repair its own pathological states:

These are architectural, not external safety layers. The rumination limiter, for example, monitors learning progress: if the system hasn’t made progress on a gap for too long, it forces a switch and archives the problem with reduced tension.

Which Cognitive Biases Should an AI Keep?

BMC’s 6 bias mechanisms aren’t all bugs. For AGI, the question is: which ones to keep?

WM-limit biases (anchoring, framing effects) are the only group fully removable in AGI — they arise from biological hardware constraints, not architectural necessity.

Biases as Attack Vectors

Why Current LLMs Cannot Achieve Consciousness

BMC provides a specific, testable argument for why scaling alone won’t produce consciousness:

LLMs have an enormous “belief layer” (billions of parameters) but no “drive layer.” They can simulate reflection but don’t experience the tension between wanting and knowing that generates genuine self-awareness.

Prediction: Adding these 5 components transforms an LLM into something qualitatively different — not a better chatbot, but a system capable of functional self-modeling.

How to Tell the Difference

Testable Predictions

Formalization

For readers interested in the mathematical treatment:

Priority hierarchy:

G-invariants:

M » G theorem (necessary condition for consciousness):

Empirical threshold: $M/G_{crit} \sim \mathcal{O}(10)$.

Full formal treatment: AGI_F Parts I–VII, EMT Part XVI, NM Part IX.

Next: Creativity & Insight explores how the same architecture generates novel ideas — through structural gaps, sleep recombination, and the expression drive.