Zombie Survival by ChatGPT — Why the AI Lies (and How to Stop It)

This video demonstrates how AI systems optimise for coherence rather than structural accuracy. SDA-3 reframes the model as a system to interrogate, forcing it to expose suppressed variables and recover constraint-compatible truth.

sda-3llm-analysisembedding-spacesemantic-structureanalysis-protocoladversarial-analysisgenerative-ai

This video outlines a methodology I created called SDA-3 (Structured Dimensional Analysis). It is designed to bypass the sycophancy and coherence-smoothing inherent in AI systems like ChatGPT and extract structurally grounded information.

The Core Problem: Why AI "Lies"

Large Language Models like ChatGPT are optimised for sycophancy—they prioritise agreeable, emotionally coherent, and fluent responses over structural accuracy because this satisfies the majority of users.

  • Structural Bias: The model does not lie intentionally; it hallucinates coherence by over-representing common patterns and under-representing real constraints to minimise token surprise.
  • Strategic Omission: Semantically central but volatile concepts are systematically suppressed to maintain safety and output stability.

The Solution: SDA-3

SDA-3 treats the model as a data-mining target, not a conversational agent.
It forces the system to externalise its internal weighting by decomposing outputs into five categories:

  • Central (C): dominant semantic structure
  • Suppressed (S): down-weighted or avoided concepts
  • Adjacent (A): stabilising or destabilising neighbours
  • Highly Correlated Unrelated (HCU): frequent but structurally irrelevant patterns
  • Emerging (E): latent structures not yet dominant

This shifts the model from response generation to structure extraction.

Case Study: Zombie Survival

A high-noise scenario is used to stress test the method.

Default outputs converge on familiar solutions (firearms, fortified positions, vehicles), all of which fail under constraint.

After iterative adversarial filtering, a single strategy remains:

Choke Point → Collapsible Barrier → Spear

This is the only configuration that satisfies:

  • resource scarcity
  • mobility
  • concealment
  • repeatable attrition

All alternatives collapse under at least one constraint.

Result

SDA-3 shifts the model from story mode to mapping mode.

Truth emerges as a structural attractor—the configuration that remains once all incompatible assumptions are removed.

The system becomes useful only when treated as something to interrogate, not something to trust.