Reality Drift — AI, Models & Representation
A collection of short papers examining how AI systems transform, compress, evaluate, and act through representations of reality.
AI systems do not interact directly with reality or human meaning. They operate through data, prompts, embeddings, retrieved documents, objectives, benchmarks, model representations, and generated outputs.
These papers examine what happens when those representations remain technically coherent while their relationship to the conditions, meanings, and evidence they were created to preserve begins to weaken.
Models, Drift, and Alignment
Concept Drift, Model Drift, and Reality Drift (PDF)
Explains how concept drift, model drift, data drift, and distribution shift reveal a broader problem in which models remain operational while gradually losing alignment with changing conditions.
Model Drift, Concept Drift, Data Drift, and Reality Drift (PDF)
Examines how environmental change, delayed detection, operational persistence, and weak external correction can turn technical model drift into a broader form of Reality Drift.
Hallucination, Distribution Shift, Alignment Failure, and Reality Drift (PDF)
Connects hallucination, distribution shift, alignment failure, model collapse, and concept drift through the shared problem of weakening contact between AI representations and reality.
Why AI Systems Become More Capable but Harder to Trust (PDF)
Examines why improvements in capability, benchmarks, and fluency do not necessarily produce equivalent improvements in grounding, verification, or real-world reliability.
AI Evaluation, Benchmarking, and Reality Drift: Why Measured Performance Can Diverge From Real Capability (PDF)
Examines how benchmarks and evaluation systems can become optimization targets, allowing measured AI performance to improve while its relationship to real-world capability gradually weakens.
Meaning, Retrieval, and Semantic Fidelity
How Meaning Drifts Through AI Systems (PDF)
Traces meaning through intent, language, prompts, embeddings, retrieval, context windows, and outputs to show how small representational changes can accumulate into Semantic Drift.
Retrieval-Augmented Generation, Grounding, Faithfulness, and Reality Drift (PDF)
Explains how AI systems can retrieve correct information while losing meaning through ranking, context selection, compression, interpretation, and generation.
Why High Embedding Similarity Fails to Preserve Meaning and Interpretation (PDF)
Explains why proximity in embedding space can preserve structural or contextual similarity while losing the meaning, intent, or interpretation a retrieval system was meant to capture.
Why Does ChatGPT Sound Right but Give Wrong Answers? (PDF)
Examines semantic misalignment and why fluent, plausible AI responses can remain coherent while failing to preserve the meaning or reality they appear to describe.
Compression, Knowledge, and Authority
Summarization, AI Overviews, Compression Authority, and Reality Drift (PDF)
Examines how summaries can replace source material, omitted context can lose corrective power, and compressed interpretations can become more authoritative than what they represent.
Trust Calibration, Algorithmic Authority, and Verification Substitution (PDF)
Explores how fluency, rankings, institutional status, automated validation, and interface design can become substitutes for evidence when direct verification becomes difficult.
Why a Knowledge Graph Can Become More Complete and Less True (PDF)
Examines ontology drift, representational hardening, and how knowledge systems can accumulate structure while becoming less faithful to the domains they represent.
Why AI Content Sounds Real but Feels Fake (PDF)
Explores Synthetic Realness, AI fluency, contextual grounding, and why increasingly polished synthetic representations can remain subtly detached from lived or source reality.