The meaning of words and actions often depends on the context in which they occur, including what has happened before, who is involved, their relationship, and the activity they are engaged in.

Modern AI can infer a great deal from what is available to it, but some meaning depends on tacit understandings embedded in the personal and social world people share.

Wittgenstein’s language games are useful here. He argued that words often take their meaning from the activities in which people use them, and from the shared expectations that make those activities intelligible.

In other words, understanding a sentence is not just knowing what its words mean. It is knowing what someone is doing with them in that situation.

Humans learn this through participation. We use language, see how others respond, are corrected when we get it wrong, and build common ground through repeated interaction.

The prosaic “do what you think is best” could express trust, resignation, hurt, or a hope that the other will reconsider. Context may make one reading more likely, but not certain.

Frontier models can get that judgement wrong in both directions. They can miss an implication that the context supports, or infer one that is not sufficiently justified. Scaling and reasoning help, but do not remove the problem.

This becomes more consequential with agents, which may act before this ambiguity is resolved.

Humans navigate this partly through common ground, learned understanding and a shared form of life. Agentic AI works from traces of this social world, so a plausible interpretation is not always a sufficiently grounded basis for action. Where meaning depends on tacit social understanding, the response may be to make intentions explicit, limit autonomy, and retain human judgement.