Polished Is Not Verified
The vocabulary I use for judging AI work, and the measured reason every briefing on this site ships with a ledger.
The better AI output looks, the less people check it. Anthropic's first behavioral study of AI fluency found that when conversations produce polished artifacts, fact-checking and challenges to the AI's reasoning measurably drop. Polish is a feeling. Verification is a record. That is why the ledger publishes with every briefing here.
Four words for the actual skill
I did not coin the vocabulary I use for this. The AI Fluency framework, written by Rick Dakan (Ringling College of Art and Design) and Joseph Feller (University College Cork) with Anthropic, names four competencies that decide whether AI-assisted work is any good. In my words:
Deciding what to hand the AI, what stays human, and what you do together.
Communicating the goal, the constraints, and how you want the AI to work.
Evaluating what comes back before you trust it: the output, the process, and the behavior.
Owning the result: honesty about AI's role, and accountability for what ships.
Those are my one-line versions. The official definitions, the twelve sub-competencies, and the free courseware live at aifluencyframework.org.
The artifact paradox
In February 2026 Anthropic published the AI Fluency Index, a behavioral baseline built on the framework: thousands of real conversations scored against observable fluency behaviors. The finding that matters most for leaders: in conversations that produce polished artifacts (code, documents, tools), people get more directive and less critical. Fact-checking, challenges to the AI's reasoning, and catches of missing context all decline, at the exact moment the output looks most finished.
The direction matches what I see in enterprise rooms: a confident-looking draft sails through a review that a rough one would never survive. Whole approval chains run on polish as a proxy for quality, and AI now produces polish on demand.
The caveats travel with the claim: this is Anthropic measuring users of its own product, the findings are correlative rather than causal, and the sample is a single week of early-adopter conversations. I treat the direction as credible and the exact numbers as directional.
What Discernment looks like here
Each verified briefing publishes its ledger: claims checked, fabricated figures removed, popular versions corrected from the primary source, shaky claims demoted.
Demoted claims stay visible in the ledger with the reason. A clean-looking page with hidden doubts is exactly the artifact paradox at work.
Every briefing states what would change our mind. One already fired, and the page says so; holding a published claim accountable is Discernment applied to my own work.
Judgment stays human before anything ships. That rule, and the system it governs, are on the practice notes: the operating system behind this research.
Where to learn this
AI Fluency: Framework & Foundations. Dakan and Feller's course on the framework itself. Free, with a certificate on completion.
The framework home. Official definitions, the full free courseware, and the open license.
The applied tutorial library. Includes the observable behaviors behind the Fluency Index.
This list stays short: only what I am actually using. Certificates appear when earned, not before.
The AI Fluency framework is by Rick Dakan and Joseph Feller with Anthropic; the course materials are released under CC BY-NC-SA 4.0 at aifluencyframework.org. This page explains the framework in my own words with attribution, reproduces no course materials, and has no affiliation with or endorsement by Anthropic.
Facing one of these decisions right now? No pitch. A short conversation about what matters first for your situation.
Start with one question