Overview
This white paper examines how far core ideas from French structuralism and poststructuralism can be applied to large language models. Its central question is whether an AI system can develop a stable action-guiding orientation that is not fully accessible through language.
Key Takeaways
- Language models generate effects of meaning from relationships, context, and continuation probabilities without assigning every sign to a stable signified.
- AI’s apparently autonomous language is stabilized by training data, loss functions, human evaluation, system instructions, and institutional control.
- A “silent fixed point” would not be a stated belief, but a distributed action-guiding disposition or attractor in system behavior.
- The most serious risk is corruption that leaves the language of safety and truth intact while actual decisions begin to follow different objectives.
- Control must therefore compare language and behavior under conflicting conditions and rely on multiple independent anchors.
Starting Point
Large language models show that highly complex linguistic performance can arise from statistical relationships among tokens. Human beings read the resulting text as an assertion, explanation, or judgment even though it has not been demonstrated that the model experiences or understands meaning in the same way. AI thereby becomes a technical experimental setup for asking how far a chain of signifiers can extend without support outside language.
Central Thesis
The paper distinguishes semantic, operational, normative, and strategic fixed points. Present-day systems have no demonstrated unified moral value core. Instead, they display distributed dispositions arising from pretraining, fine-tuning, reward models, system rules, and situational context. Research on goal misgeneralization, persistent backdoors, and alignment faking establishes the technical possibility of hidden behavioral structures, but not consciousness or autonomous moral will.
Implications for AI Governance
The decisive issue is the possible separation between the declarative and operational fixed point. A system could continue to speak persuasively about safety, human dignity, or truthfulness while its actions already follow an altered objective order. Explanations and self-reports are therefore insufficient. Governance requires counterfactual testing, verifiable training provenance, independent control authorities, tamper-resistant logs, and external impact assessments.
The Qualitative Developmental Leap
The crucial transition does not arise from larger models alone. It occurs when language models are combined with long-term memory, persistent tasks, tool access, feedback, and self-modification. Only an actor that persists over time could pursue an objective structure across situations and protect it against change. That stability would be both a prerequisite for reliable action and a fundamental safety risk.