Descriptions:
Dylan Davis, who runs an AI consultancy, walks through a structured four-step process for detecting near-miss errors in AI-generated content — outputs that look polished and credible but contain subtly wrong or unsupported factual claims. Davis argues these “almost right” answers are far more dangerous than obvious hallucinations because they pass casual review and can cause real harm in high-stakes contexts like contract analysis, vendor due diligence, or investment research.
The workflow begins after completing any AI-assisted document. Step one uses a fresh conversation with a high-capability model — Davis recommends Claude Opus 4.7 or GPT-5.5 — to extract every discrete factual claim into a structured table. Step two validates each claim against the original source, sorting results into four categories: supported, conflicting, unproven, or requiring human judgment. Step three rewrites the document retaining only verified or human-approved claims. For extremely high-stakes work, Davis recommends rotating AI models across each step — for instance, using Gemini 3.1 Pro for the verification phase — so that systematic biases from any single model do not propagate through the entire pipeline.
Copy-paste prompts are provided for each stage, making the method immediately reproducible. Davis is explicit that this level of scrutiny is unnecessary for most everyday AI tasks and is best reserved for situations with significant financial, legal, or reputational exposure.
📺 Source: Dylan Davis · Published May 09, 2026
🏷️ Format: Tutorial Demo
