I watched it happen progressively. Tacit knowledge got codified. Codified knowledge got structured. Structured knowledge got automated. And somewhere along the way, the investment in the people and practices responsible for keeping that knowledge honest quietly dried up. The system knew. Or so we told ourselves.

That assumption is now being stress-tested at scale.

A recent piece in Harvard Business Review by Matthias Holweg and Thomas Davenport gives a name to what’s emerging. They call it “knowledge decay” – the organisational consequence of what happens when AI-generated content proliferates unchecked across business processes. Their argument is essentially a process management one: when individuals use AI to produce polished-looking but low-quality work, and that output feeds the next step in a process, which feeds the next, errors compound. Trust erodes. Eventually what an organisation calls knowledge no longer merits the name. They frame the challenge as three interlocking problems – verification, validation, and entropy – and their recommended responses are sensible: track data provenance, constrain AI use to where it genuinely adds value, redesign processes deliberately.

It’s a solid article. But it stops short of the question that follows naturally from it.

✦   ✦   ✦

The knowledge worker's role, historically, was to do precisely what automation cannot: exercise judgement. Not just to retrieve information, but to evaluate it. To understand its source, interrogate its reliability, and place it in context. As ERP systems, process automation, and digital workflows advanced, much of the mechanical work of knowledge handling was absorbed by technology. That was the point. But a less visible consequence was the gradual erosion of institutional investment in the human capability that made knowledge trustworthy in the first place. Knowledge management as a field became unfashionable - perceived as overhead, absorbed into adjacent disciplines like IT governance or change management, or simply assumed to be solved by whatever platform the organisation had just implemented.

It wasn't solved. It was deprioritised.

Generative AI has now made those decisions consequential. LLMs have no obligation to disclose their training data. They have no peer review, no editorial board, no errata process. They produce outputs that are fluent, confident, and structurally coherent - qualities we have historically associated with reliability. But fluency is not accuracy. Confidence is not correctness. And the more those outputs feed downstream processes, the harder it becomes to trace what is signal and what is noise - which is precisely Holweg and Davenport's point about entropy.

But there is a deeper problem they don't address. Studies suggest that a substantial and growing share of internet content is now AI-generated. That content will inevitably become training data for future models. The feedback loop this creates - what the authors call "model collapse" - means that each successive generation of AI may be trained on an increasingly diluted version of the original human knowledge that gave it value. The knowledge management principle at stake is not new: garbage in, garbage out. We simply haven't applied it to LLM training data with anything like the rigour we once brought to corporate knowledge repositories. Or should have brought, at least. The UK's Information Commissioner's Office has acknowledged as much - noting that while validating ground truth in training data has real limitations, that doesn't break the link between inaccurate training data and inaccurate model outputs.

This is where the KM practitioner's perspective matters - and where the discipline's marginalisation starts to look less like progress and more like a liability. Verifying AI output requires informed human judgement. Not algorithmic checking, not AI-on-AI review, but the kind of grounded, experienced assessment that comes from knowing a domain well enough to recognise when something that sounds right is actually wrong. That capability doesn't survive in organisations that have spent years systematically reducing their dependence on it.

The recruiting example in the HBR piece illustrates this with uncomfortable clarity. AI writes job descriptions. Candidates use AI to optimise CVs against them. AI screens and ranks the applications. AI conducts initial interviews. Humans, having been largely removed from the process, are left to choose from a shortlist that has been shaped at every stage by systems optimising for keyword alignment rather than actual fit. The result - as research cited in the article confirms - is more volume, less signal, and declining trust in the process from both sides. It is knowledge decay made visible.

What strikes me about this moment is not that the problems are new. It's that they are old problems wearing new clothes - and that the discipline best placed to address them has been descoped, unfunded, or at best acknowledged in principle if not in practice. It's necessary to point out that failed KM initiatives no doubt contributed to this, and the perception KM was simply another management fad.

✦   ✦   ✦

Data quality, knowledge provenance, and process effectiveness are now converging as the critical foundations of any organisation serious about AI. These are not technology questions, though technology is implicated. They are knowledge management questions. Who owns the ground truth? How is this validated? How do you preserve the integrity of information as it passes through a chain of AI-mediated processes? How do you ensure that the human judgement required to verify outputs is present, developed, and trusted?

The HBR article closes by invoking the productivity paradox - the historical lesson that technology only improves productivity when the processes around it are designed to enable it. I'd extend the analogy. The knowledge paradox of our current moment is that we have built the most capable information tools in history, and we are deploying them in organisations that have yet to establish the critical knowledge foundations they depend on - the ground truths, the governance, and the human judgement required to make AI outputs trustworthy.

Knowledge management didn't become obsolete. It became invisible. And now, with some urgency, the need for it is rising again.