"Where Provenance Ends, Knowledge Decays" Reflections

Author: Yauheni Kurbayeu
Published: Mar 17, 2026

Here is another strong argument on something that has been quietly breaking beneath the surface of the AI wave - the relationship between knowledge and its origin.

I recently read “Where provenance ends, knowledge decays” by Jessica Talisman, and what makes it powerful is not that it criticizes AI, but that it reframes the problem entirely.

It shows how, as language models become better at producing fluent, convincing outputs, they also dissolve the chain that makes knowledge trustworthy in the first place.

Not just citations, but the full lineage - who created something, under which assumptions, in what context, and why those decisions made sense at the time.

The result is subtle, but structural.

We are not losing information.
We are losing the ability to understand where that information came from and whether it should still be trusted.

Knowledge starts to shift from something grounded and traceable into something that is simply plausible.

And the more we rely on these systems, the more this effect compounds.

What I find particularly important is that the article does not reduce the problem to “AI needs citations.”

It goes deeper.

Provenance is not a formatting issue, it is an integrity layer.

It is what allows systems, human or technical, to validate, audit, and evolve knowledge safely over time.

Without it, even correct outputs become fragile, because their reasoning cannot be reconstructed.

Reading this, I kept mapping it to something I’ve been exploring from a slightly different angle - decision provenance in software delivery.

In engineering, we have always been good at preserving outcomes.

We store code, APIs, infrastructure, documentation.

But the reasoning behind them, the decisions, trade-offs, rejected alternatives, tends to disappear into meetings, chats, and individual memory.

It was already a problem, but a manageable one while systems evolved slower.

AI changes that completely.

Now decisions are not only made faster, they are increasingly co-produced by humans and agents.

And if the reasoning behind those decisions is not captured as part of the system itself, we end up with the same kind of decay the article describes, except now inside the systems we build, not just the knowledge we consume.

This is where I see a strong overlap with the idea of a provenance-native approach.

The article frames the problem at the level of knowledge ecosystems.

My work tries to push it into execution.

What would it mean if every meaningful action in the SDLC produced a structured trace, not just what changed, but why it changed, what constraints existed, what options were considered, and who or what made the call?

In that sense, provenance stops being documentation and becomes infrastructure.

At the same time, there is also an extension here.

The article focuses on the loss of provenance as a risk.

But once you start capturing decision lineage systematically, it becomes more than mitigation.

It turns into a new kind of capability, the ability to query reasoning, to audit AI participation, to replay decisions, to evolve systems without guessing.

And that’s where a few interesting spin-offs start to emerge.

Provenance is not only about preserving trust, it becomes a foundation for governance, for explainability, and eventually for building systems that can safely collaborate with other systems.

What this article makes very clear is that we are not just scaling knowledge production.

We are reshaping the conditions under which knowledge remains valid.

And if provenance is not part of that foundation, decay is not an edge case.

It becomes the default.