Who Has Contact With the Thing?

In 1990, the cognitive scientist Stevan Harnad asked a question that now sits under every enterprise AI project: how can the symbols in a formal system mean anything from within the system, rather than only in the heads of the people reading them? He compared it to trying to learn Chinese from a Chinese-to-Chinese dictionary. Every definition leads to more symbols, and none of them leads to the world.

Three decades later, Emily Bender, Timnit Gebru, and colleagues made a related point about language models in a paper whose phrase stuck: "stochastic parrots." The model produces fluent sequences because of statistical patterns in text, not because it has contact with what the words refer to.

In plain business terms: a model can write "customer retention is improving" with complete fluency whether or not a single customer has stayed. It works entirely on the sign side of the line. The thing, the customer, is somewhere it has never been.

That is not a reason to avoid the tools. It is a description of the job they leave to you. Someone has to supply the referent: to go to the account, the warehouse floor, the call, and check that the sentence is about something real.

Leaders who understand that stop asking "is the model accurate?" and start asking "who in this process has contact with the thing?"

Who in your AI workflow has direct contact with what the output describes?