Chunking long transcripts catches AI deception that full-context monitoring misses
Monitoring long AI transcripts in chunks instead of all at once catches deceptive behaviors that full-context monitors miss, new research shows. Chunk-level deception signals get diluted over long contexts, so chunked monitoring is a cheap, practical upgrade for AI safety evaluations and jailbreak detection.