Prompt Engineering as a Control Illusion

Key points

• Language models are probabilistic; prompts change likelihoods, not boundaries.
• Treating prompts as control is a category error: language describes and persuades, control constrains.
• Prompt injection is structural evidence that instructions and inputs share the same inference process.
• Governance defines intent; architecture defines capability. Without an execution-constraining layer, governance stays declarative.
• Control layers should be local, bounded and explicit about which actions exist, which are reversible, which are non-delegable and when a human must assert authority.

The confusion

Considerable effort goes into refining instructions and context so that models behave. In short demonstrations and narrow tasks this can look stable, which encourages the belief that better prompts will eventually deliver reliability. Across time, contexts and organisational boundaries, systems drift towards locally coherent outputs that ignore global constraints. That is not a flaw to be engineered away; it follows from probabilistic inference under uncertainty.

How it fails

When prompts serve as guardrails, failure accumulates quietly: edge cases become frequent, exceptions become normal, and responsibility is deferred until after execution, when audits and reviews describe what should have happened without changing what the system was able to do. Mature engineering disciplines do not rely on description to prevent overload or leave the flight envelope; they use constraints that make certain actions impossible.

The alternative

The answer is architectural separation, not larger models or longer instructions. Between inference and execution there should be a layer whose only job is to decide whether an action is permitted, needs escalation or must stop — making execution conditional on resolved ownership rather than inferred intent. Prompts remain valuable for interaction and expression; they are not a foundation for control.

What the paper does not claim

This is a short argument, not an empirical study. It does not claim that prompt design is useless, only that it cannot serve as a control boundary.

This page is a summary written by One World AI. The paper on Zenodo is the authoritative version.

Cite this paper

Bański, P., & Sobieski, S. (2026). Prompt Engineering as a Control Illusion. Zenodo. https://doi.org/10.5281/zenodo.18213542

More research

• Standing to Stop — A stopping point is not enough. The party at it needs presently valid, traceable authority, and the stop must actually change what can happen.
• Architectural Constraints in AI Governance — Governance of AI-enabled operational systems has to live in the architecture that decides which state transitions are admissible — including across sequences of individually acceptable steps.
• Enforceable Decision Boundaries — Inference may propose actions but never authorise them. Authority is resolved outside the model, and execution is impossible without it.
• The Missing Place for “Stop” — Systems that work well for years can still drift into decisions nobody can defend — because no one is structurally placed to stop them before execution.