AI hierarchizes credible sources even when no explicit arbitration rule has been declared. The article explains how that hidden hierarchy shapes answers.
Archive
Blog — page 5
Paginated archive of Gautier Dorval’s blog.
AI often arbitrates without a central truth source. The article explains how authority, reputation, and weak signals combine under synthesis.
A canonical map for biometrics, where identification, verification, surveillance, prohibitions, and legitimate non-action must remain sharply separated under synthesis.
Biometrics becomes dangerous when AI treats identification, verification, and surveillance as interchangeable categories.
Bundles and options are structurally hard for AI to preserve. The article explains why complex offers are systematically misinterpreted.
Certain information disappears in synthesis because compression rewards portability over nuance. The article explains why that loss is structural.
Structured data can stabilize meaning, but it can also destabilize it when schemas overlap, contradict, or cancel each other out.
A contradiction between credible sources is not solved just because the model produces one answer. The article explains the hidden normalization at work.
When credible sources contradict each other, AI often chooses silently. The article explains why that silence is itself a governance issue.
A controlled lexicon stabilizes official phenomenon names and definitions so the corpus does not compete with itself through synonyms, near-synonyms, and drifting labels.
Credit governance prevents a model from reconstructing scoring logic, overextending factors, or hiding temporality and negations that remain essential to interpretation.
AI can “score” without saying so. This article examines how access gets hardened by implicit ranking rather than explicit scoring.
A validation protocol for testing an entity across models without turning model preference into the hidden variable. The goal is comparable observation, not model ranking.
In customer support, AI becomes risky when a helpful answer crosses an authority boundary and starts sounding like a commitment about conditions, guarantees, refunds, or exceptions.
You do not always need to question the LLM directly to see the drift. Misinterpretation often becomes visible through its indirect effects.
A site can lose interpretive authority without losing visibility. The answer layer may simply adopt a stronger third-party frame.
The old can dominate the new long after a change has been published. This article explains how historical salience becomes interpretive inertia.
The drift index measures the variance of formulation over time. Its object is not ranking volatility, but the stability of meaning under repeated synthesis.
E-commerce governance keeps product attributes, variants, negations, and proof conditions explicit so synthesis does not flatten a governable offer into a misleading simplification.
Education governance structures thresholds, evidence, and legitimate non-action so that generative systems do not harden contextual conditions into universal rules.
In education, AI recommendations can become de facto decisions. The article explains how advisory language hardens into direction.
When the same structure repeats often enough, AI may treat the template itself as a semantic rule. This article explains that drift.
Entity dissonance appears when the official source and the surrounding environment no longer describe the same object.
Facets and pagination do not only affect crawlability. They can dilute the semantic perimeter that AI uses to reconstruct an e-commerce offer.