AI, Meaning, and Representational Failure

Short papers on how AI systems preserve fluent representations while meaning, context, and source fidelity weaken.

Modern AI systems operate through layers of representation rather than direct contact with reality. Retrieval, compression, summarization, reconstruction, and generation can preserve recognizable information while weakening context, provenance, constraints, and meaning. The papers below examine how this happens.

Compression and Meaning Loss

How Abstraction Causes Information to Lose Meaning (PDF)
Examines how selection, compression, normalization, recontextualization, reconstruction, and recursive compression can preserve recognizable information while weakening the relationships that originally gave it meaning.

Why an AI Summary Can Be Factually Accurate and Still Misleading (PDF)
Explains how an AI summary can contain facts supported by its source while still changing the conclusion, confidence, scope, or practical implication a reader takes from the original material.

AI Interpretation

Why ChatGPT Can Sound Right While Getting the Answer Wrong (PDF)
Examines how fluent AI responses can blur the distinction between source material, inference, generalization, and advice, making an answer appear better grounded than the evidence supports.

Retrieval and Evaluation

Why Retrieval and Citations Can Still Lose the Meaning of a Source (PDF)
Explains why retrieving and citing the correct source does not guarantee faithful interpretation, and how a source can remain present in a workflow while losing its ability to constrain the answer.

How to Evaluate Semantic Fidelity Across an AI Workflow (PDF)
Presents a diagnostic method for tracing meaning across retrieval, intermediate representations, generated answers, agent handoffs, and downstream reuse while testing whether the original evidence can still correct the system.