The Human
Factor.

The web split into two audiences.

For most of its history the web had one kind of reader: a person. Ranking, advertising, design — the entire economy of discovery — was tuned to human attention, because attention was the scarce resource being competed for.

That is no longer the only audience. By 2026, automated requests crossed a threshold forecast for years and arrived early. Agents now consume the web in a way humans never did — software acting on a person's behalf: visiting pages, comparing options, filling forms, completing transactions.

Share of measured web requests, 2026 · Cloudflare network data — attribute

The two audiences do not run on the same scarce resource. The human-facing front runs on attention and trust — a person has finite time and weak defenses against volume. The agent-facing front runs on provenance and verifiability — an agent has unlimited attention but must evaluate what it consumes against sources and constraints. Hold that separation. It is the frame for everything below.

Where a person is reading.

Start where the reader is human. Synthetic content is now a large share of what gets published.

On this front the human input shows up as origination and taste — the decision about what is worth making — and as discernment about what reads as true. Reported consumer preference for content known to be AI-made fell sharply as the novelty wore off, with authenticity re-emerging as the thing people say they value.

The variable that moved trust was not whether a machine was involved. It was whether a human had exercised judgment about presentation, disclosure, and legitimacy.

Across thirteen experiments, disclosing that you used AI reduced trust in the discloser — in some settings, a person who disclosed was trusted less than an autonomous system doing the same task. The mechanism was legitimacy, not accuracy. So "just label it" is not a clean fix; transparency is itself a judgment call with a cost.

Where an agent is acting.

Now cross to the other front, where the consumer is a machine acting for someone. The scarce resource is no longer attention — "AI-generated" is not even a defect category here, since everything in this lane is machine-produced by definition. The only questions are whether it is verifiable, and whether what the agent does stays accountable.

The established answer is a named function, not a mood. In the NIST risk framework, Govern is the cross-cutting requirement — it establishes accountability and oversight above the work of mapping, measuring, and managing. On this front the human input is governance: the decision trace, the escalation threshold, the answer to "who authorized this."

The obvious objection is that oversight is a tax on speed. The most useful single finding says otherwise:

Central estimate 91% · Kang, “Governed AI-Assisted Engineering” 2026 — peer sample

A 2026 framework for regulated code generation routes each task, by blast radius, into human-in-the-loop, human-over-the-loop, or automated-with-monitoring — and preserves nearly all agentic velocity while keeping compliance evidence intact. Read plainly, the human factor is not the brake. Keeping the human at the load-bearing points — and only there — retains almost all the machine speed.

The same constant, twice.

Set the two fronts side by side and something recurs. On the human front, the input that moved outcomes was origination and the judgment of how to earn trust. On the agent front, it was the design of oversight and the judgment of what stays accountable. Different registers — taste and authenticity in one, architecture and governance in the other — but the same two capacities: creativity and judgment. They are not assigned one per front. They appear on both, wearing the clothes of each.

The failures rhyme for the same reason. Slop is ungoverned generation aimed at a human feed; the shadow agent is ungoverned generation aimed at a P&L. Whether that pattern is a coincidence of two immature markets or something more durable, I will leave to the reader.

Where the lens does not reach: not every problem on either front is a story of underweighted human input. Some are plain capability limits. Some are adversarial: deliberate misuse that more human judgment on the defender's side does not, by itself, prevent. And the two fronts are not cleanly separable — agents are trained on and retrieve from the same human web, so pollution in one lane leaks upstream into the other. A lens that claimed to explain everything would be a pitch. This one explains a lot, and stops where the evidence does.