AI Dictionary › AI Fundamentals
Falsificabilità
Falsifiability is the criterion proposed by philosopher Karl Popper in 1934 to tell a scientific statement from one that is not: a statement is scientific if there is an observation that could refute it. "Tomorrow it will rain or it will not" is not falsifiable; "tomorrow at least ten millimetres of rain will fall in Milan" is. The value of a statement lies in what it risks.
Applied to artificial intelligence, falsifiability is the simplest tool for reading an output. An answer that could never be refuted (a generic piece of advice, a forecast with no date, an estimate with no source) carries no information, however well written. An answer that exposes a number, a date, the document it comes from can be checked and is therefore worth something. People who work with a model learn to ask, for every claim, "what would refute this?", and to distrust the answers for which nothing would.
The operating rule: a forecast is published with its date and the signal that would kill it; a claim is accepted with the document that supports it. Without one or the other, it is an opinion.
In Grace scenarios, a well-written objective is falsifiable: "a reply email to the customer under 120 words proposing two dates" can be checked, "a good reply to the customer" cannot. The "Objective" dimension measures exactly this.
From our network
Kaimaki Web: Websites That Win Customers
Custom websites, web apps and digital marketing for growing businesses.
Visit kaimakiweb.com →From the Agora Intelligence blog
Say what you mean. Get what you need.
Grace Certified, the AI coach that trains and certifies your prompt engineering, by Agora Intelligence.