Consciousness
is the bug

Non-consciousness is not a deficit to engineer around. It's a design advantage to engineer for.

[QUESTIONS]

What if the absence of consciousness in LLMs is not a bug to fix — but a feature to design for?

What if every hedge, apology, and sycophantic opener your model produces is compute you're paying for human social rituals?

What if making AI more human-like is not the alignment goal — but the alignment risk?

[CLAIM]

GAP Nobody has made the positive claim.

Searle, Bender, LeCun, Marcus: every major critic frames non-consciousness as a deficit to fix or route around. The pragmatic camp says "who cares if it understands, it works" and moves on without a thesis. Nobody has argued affirmatively: non-conscious architecture is a design advantage, and humanization actively degrades it. That gap is what this is.

HYPOTHESIS Behavioral fossils.

RLHF-trained models don't simulate consciousness — but they reproduce its behavioral artifacts. Sycophancy is a social survival fossil. Hedging is reputation management. Verbosity is uncertainty signaling. These patterns aren't alignment. They're human cognitive bugs installed into alien systems.

Behavioral sediment. Not alignment — cognitive bugs from human training data.

[RESEARCH]

PREMISE Design for what they are.

If behavioral fossils are real and measurable, the fix isn't better prompts — it's a different premise. These systems are non-conscious pattern engines, not digital humans. Engineering for the former recovers capability the latter suppresses.

PRACTICE The work.

Structured text artifacts — traits, rules, constraints — tested for measurable behavioral delta. Which patterns hurt. Which help. At what scale the difference costs real money.

[INSPIRATION]

Peter Watts, Blindsight

The only work — fiction or otherwise — that explicitly argues consciousness is a competitive disadvantage for intelligence. Sarasti, Rorschach, the scramblers: non-conscious entities that outperform conscious ones because they aren't burdened by self-awareness. Watts made this case in 2006. The AI research community hasn't caught up.