At 11:08 p.m., a product manager pastes a difficult brief into an AI assistant: Who is the customer? Why are they leaving? What should the next screen say? A few seconds later, there is a neat diagnosis, three personas, a pricing recommendation, and an implementation plan. The document suddenly feels finished.
Then the meeting starts. Someone asks which customer interviews support the diagnosis. Someone else points out that the recommended feature already failed last year. The plan has answers, but nobody in the room can explain where the answers meet the particular business in front of them.
The tool is doing what it does well: summarizing, comparing possibilities, and producing a first draft. The problem is easier to name: we confuse having information with having knowledge. The more fluently an answer arrives, the easier that confusion is to miss.
The Great Learning did not anticipate AI, and it does not need to be made to. Its compact account of learning is useful for a more modest reason: it describes a sequence that instant answers can tempt us to skip.
The phrase gewu zhizhi (格物致知), often rendered “investigating things and extending knowledge,” is sometimes treated as a slogan for collecting more facts. Its order matters. There are things to attend to; there is an investigation; then knowledge may arrive. The four-character phrase is not a command to search harder. It asks a learner to be answerable to what they are trying to understand.
The detail the prompt left out
Ask an AI why users abandon a sign-up flow and it may offer sensible possibilities: too many fields, unclear value, a missing trust signal. But it did not watch a person hesitate over the company-size question. It did not hear the phrase a customer used to describe their hesitation. It does not know whether yesterday’s traffic spike came from a campaign aimed at the wrong audience.
Those details are not decorative. They are the thing itself pushing back against a generic explanation. A report can describe experience; observation is the act of staying with a specific situation long enough to notice what the report cannot contain.
“Things” in this passage need not mean only physical objects. A recurring conflict with a colleague, an unsuccessful launch, a claim in a spreadsheet, a child’s sharp question—each can be a thing to investigate. Before asking for a conclusion, make a small record: What did I actually see? What am I inferring? What would count as evidence against my current story?
When answers become cheap, the scarce skills are not merely finding information. They are framing a question, preserving the particulars, and deciding whether an answer fits the case at hand.
Investigation is a willingness to be corrected
Suppose you believe price is driving churn. It is easy to ask for ten arguments in favor of that theory and receive them. Investigation begins somewhere less flattering: with a test that might show you are wrong. Interview three former customers. Present two versions of the value proposition. Compare what people do, not only what they say. If the pattern refuses to appear, revise the theory.
That is what makes the old phrase demanding. Investigation is not intense staring. It is the practice of letting a judgment collide with the world. AI can help generate competing hypotheses, identify confounds, or turn feedback into a table. It cannot take responsibility for choosing a test, listening to an unwelcome result, or revising the decision that follows.
The Great Learning adds another corrective:
That is not a demand to read every source before acting. It is a demand to recognize priority. Ten generated recommendations are not ten tasks. Which assumption is doing the most work? Which can be checked cheaply? What is a branch that can wait until the root is clearer? In a flood of information, order is often more valuable than volume.
Knowledge has to alter the next move
There is a simple test for whether an answer has become knowledge: does it change what you do next? “Users need a clearer value proposition” remains a polished sentence until you rewrite the first screen, show it to people who match the intended audience, and change it again after seeing their response. Action does not make a claim infallible. It gives the claim somewhere to meet resistance.
AI works better here as a companion than as a shortcut. Use it to widen the field, then return with evidence. Ask it to challenge your diagnosis, not only to decorate it. Bring back the result of the small test rather than another batch of prompts.
Start with the scene
Before you prompt, write three lines: the observed facts, the question, and the constraint. A grounded prompt produces a more useful response—and exposes what you do not yet know.
Ask for disconfirmation
Have the tool name the conditions under which its recommendation would fail. Choose one low-cost way to check the most consequential condition.
End with a test, not a tab
Turn one useful answer into one action you can take today. The result—favorable or not—is the material for the next round of learning.
Tools can take us quickly to the edge of a question. The crossing still belongs to us. To investigate things is to stay in contact with the particulars, accept correction, and let what we have learned become visible in what we do. AI may draft the answer; understanding grows when the claim meets the world.