Glossary: drifts and interpretive inertia
This page groups the phenomena that degrade the fidelity of an interpretation produced by AI systems (LLMs, generative engines, agents, RAG) when meaning drifts, smooths, or freezes.
These phenomena are not isolated “bugs”: they result from a probabilistic reconstruction of meaning, fed by partial signals, successive aggregations, and unstable contexts.
Each entry links to:
a canonical definition (if it exists), a framework (if applicable), and related pages for deeper understanding.
Quick access
Terms in the “drifts and inertia” family
Interpretive hallucination
Production of a plausible but unenforceably anchored response, often stabilized by form rather than by evidence.
- Definition: Interpretive hallucination
- Clarification: Hallucinations, attribution, and interpretive risk
Interpretive smoothing
Tendency of an AI system to erase rough edges, nuances, and negations in order to fit a concept into a standardized category.
- Definition: Interpretive smoothing
Interpretive inertia
Persistence of a prior interpretation, even after modification of sources, due to the progressive stabilization of an algorithmic “truth”.
- Definition: Interpretive inertia
Interpretive remanence
Reappearance of an old interpretation in certain contexts, even when a correction seems established elsewhere.
- Definition: Interpretive remanence
- Doctrine: Version power
Interpretive tail
Intermediate phase where a correction progresses unevenly: some outputs correct themselves, others remain frozen or ambiguous.
- Definition: Interpretive tail
State drift
Divergence between a real state (price, availability, policy, status) and the state returned by AI, when an update does not propagate.
- Definition: State drift
Compliance drift
Progressive gap between expected constraints (canon, rules, response conditions) and observed outputs, despite an apparently stable documentary base.
- Definition: Compliance drift
- Doctrine: Interpretive observability
Recommended links
Next page: Glossary: canon, authority, non-response
