Dispatches
field notes from the frontier — the full archive
GUIDESAI Hallucinates Sometimes. Here's How to Spot When It Does.
AI systems generate confident-sounding fabrications without any audible tells, but you can spot them by watching for vague attribution phrases like "research shows," suspiciously specific unsourced details, and stories that change when you rephrase the question. The real defense isn't waiting for better AI—it's developing verification habits on your end, especially for claims that matter.
GUIDESHow to Create Accurate Summaries of Long Articles, Documents, and Meeting Transcripts
Summaries are where AI hallucinations hide best because a fabricated detail blends seamlessly into accurate points, and nobody fact-checks by reading the original 40-page contract they used AI to avoid. You can fix this by giving the model a second job after summarizing: force it to identify three ways the summary could be wrong, cite specific evidence from the source, and revise accordingly.