Human Provenance: What AI Cannot Be Allowed to Own

A document can now be assembled from human notes, proprietary data, web sources, model retrieval, generated analysis, synthetic examples, and automated formatting without leaving visible seams.

The result may be accurate.

It may also be impossible to tell where evidence ended, inference began, and a human choice was made.

Most provenance discussions focus on content origin: Was the image generated? Which model produced the text? Can the source file be verified? Those questions matter. NIST identifies content provenance as a primary generative-AI consideration and recommends traceability, documentation, testing, and mechanisms for evaluating authenticity and origin.[1]

Strategic work requires another layer.

Human provenance is the trace of authorship inside the decision.

AI can perform judgment-shaped tasks

A system can rank options, summarize trade-offs, estimate sentiment, recommend a strategy, draft a refusal, and explain why the preferred direction appears strongest.

The output resembles judgment because judgment also produces rankings, reasons, and choices.

But the system does not bear the institutional, ethical, reputational, or human consequence in the way the decision owner does. The organization cannot transfer accountability by allowing the recommendation to arrive in completed language.

This is why “human in the loop” is often too weak. A person can remain in the loop as an approver while contributing almost no meaningful authorship. Clicking accept is not the same as deciding what counts.

Provenance should follow significance

The most important decisions are rarely hidden in the final sentence. They occur earlier:

• which sources were considered credible;
• which stakeholders were treated as relevant;
• which harms were excluded as immaterial;
• which comparison class was selected;
• which interpretation was framed as primary;
• which objective was allowed to outrank another;
• which uncertainty was tolerated.

These are significance decisions.

AI can assist with each one. It should not make them disappear.

A strategy becomes unauditable when the team can reproduce the text but not the reasoning that gave certain evidence more weight.

The provenance ledger

For material decisions, maintain a compact ledger with four layers:

Evidence: What entered the analysis, from where, and with what limitations?

Inference: What interpretation was drawn, and what competing interpretation remained viable?

Machine contribution: What did AI retrieve, generate, rank, simulate, or transform?

Human authorship: Who selected the direction, rejected alternatives, accepted the trade-off, and owns the result?

This is not bureaucratic paperwork. It is an insurance policy against institutional amnesia and automated certainty.

Provenance protects the human from the machine—and from the room

The risk is not only that AI will make an error. Humans also hide behind systems, committees, and consensus.

A provenance record prevents a familiar sentence: “That was what the data told us.”

Data does not tell an organization what to value. Models do not decide which consequence is acceptable. A system can expose a likely outcome and the room can still choose against it for legitimate reasons.

The obligation is not to obey the model.

It is to make the departure visible.