When Artificial Intelligence Became a Co-Producer of Knowledge, Not Merely a Tool

For most of the history of management thought, one assumption was so obvious that it was rarely written down: the knowledge inside an organization is produced by its people. Machines stored knowledge, moved it, sorted and retrieved it—but the act of knowing belonged to human beings. Our best theories of organizational learning are built on that premise. When Nonaka and Takeuchi described the knowledge-creating company, the engine of creation was social: tacit understanding made explicit and back again, through conversation, apprenticeship, and shared work among people. When Huber catalogued the processes of organizational learning—acquisition, distribution, interpretation, memory—each was something humans did, aided by tools but not shared with them.

Generative artificial intelligence changes the verb. It does not merely retrieve what people have written; it drafts the strategy memo, proposes the hypotheses, synthesizes the literature, and produces the first version of the analysis a manager or analyst once wrote unaided. The machine has moved from the filing cabinet to the table. It is no longer only a place where knowledge is kept; it is a participant in the making of it.

This has two consequences that our theories are not yet equipped to absorb. The first concerns provenance. When a recommendation is half-human and half-model—drafted by a system, edited by a person, approved by a committee that read neither the prompt nor the training data—whose knowledge is it, and who is accountable for it? The tidy chain that ran from an expert's judgment to a decision now passes through an author that cannot be cross-examined and does not remember what it produced.

The second consequence is deeper, and it is the one worth sitting with. Human-produced knowledge always arrived with a kind of built-in friction. It cost effort to make, which limited its volume. It came attached to a person who could be questioned, whose competence could be gauged, and whose confidence usually bore some relationship to their command of the subject. Fluency, in other words, was a rough proxy for reliability. That proxy has now broken. A capable model can produce a thousand fluent, confident, well-organized paragraphs in an hour—including when it is entirely wrong. The polish that once signaled effort and understanding now signals neither.

Organizations therefore inherit a burden they have never had to carry at this scale: separating knowledge that is plausible from knowledge that is valid, quickly, and for an ever-growing volume of material that arrives already sounding true. The classic models offer little help here, because they quietly assumed the producer and the interpreter of knowledge were the same human, and that plausibility tracked competence closely enough to ignore the gap. Neither assumption survives contact with a system that generates without understanding.

Organizations have so far responded to this in one of two unsatisfying ways. Some adopt the tools uncritically, folding machine-generated analysis into decisions as though it carried the same warrant as the work of a trusted expert, and discovering the errors only after they have already propagated into commitments. Others respond with prohibition, banning the systems outright and forfeiting their genuine value along with their risks. Both responses share a hidden assumption: that the question is whether to trust the machine. It is not. The machine will be used; the useful question is how an organization decides, case by case and at scale, which of its outputs to rely upon—and that is a question about the organization's own judgment, not about the technology. Treating it as a procurement decision, or as a policy to be issued once and forgotten, mistakes a permanent new condition of knowledge work for a passing controversy about a product.

None of this is an argument against the tools. The productivity is real, and the genie will not return to the bottle. It is an argument that we have adopted a new kind of participant in the organizational mind without updating our account of how that mind is supposed to work. We still teach that the central challenge of organizational learning is to create knowledge. Increasingly, the harder challenge is to decide which of the knowledge now offered to us, at volume and at speed, deserves to be believed—and to build that judgment into how the organization works, rather than leaving it to whoever happens to be reading. That problem does not yet have a settled home in our theories. It should.