CONCEPT
The Inscription Error
Brian Cantwell Smith’s name for the methodological trap in which a researcher projects ontological assumptions onto a system, then reads those very assumptions back off the system’s performance as though they had been discovered there empirically.
The inscription error, coined by Brian Cantwell Smith in his work on the metaphysics of computation, describes a specific and consequential form of circular reasoning that runs through the history of artificial intelligence. A researcher must decide what the world contains before building a system to reason about it—what objects exist, where boundaries fall, what relations hold. This decision is a human act of
registration, the achievement of carving continuous reality into the discrete entities that representation requires. The researcher then encodes this registration into the system as its foundational ontology and runs the system on problems defined within that ontology. When the system performs well, the temptation—nearly irresistible—is to announce that the system has demonstrated something about intelligence. But it has demonstrated no such thing. The hard part, the carving of the world into the right objects for the situation, was solved by the researcher in advance and then quietly attributed to the machine. Smith called this
pre-emptive