Does verifying AI output actually require more work than editing human text?

Pavel Samsonov, writing for Nielsen Norman Group, lays out what verifying a chatbot's answer actually requires, as a sequence distinct from ordinary text review: "the user will need to follow the same steps as they would when reviewing any other text, and then some." The six steps he names are: review the entire output; identify items that need verification; validate the claims being made; verify that the arguments follow from the claims; articulate corrections; and verify the response again.

Each step is harder for AI output than for human-authored text, in Samsonov's account, for reasons specific to how LLMs generate text. Reviewing the entire output requires building a mental model of its structure "from scratch," because unlike a piece one has written oneself, "even between prompts on the same topic ... LLM outputs are nondeterministic." Validating claims means tracking down whether a cited source "exists" at all and judging whether it is authoritative, or — if nothing was cited — doing the research from scratch. Articulating corrections means updating the prompt rather than editing directly, if the user isn't willing to rewrite the output themselves. And the final re-verification step exists because "there is no guarantee that any given error was fixed or new errors were not introduced," given the model's limited context window and nondeterminism. Samsonov allows one workaround — prompting for one small, digestible part of a text at a time — but flags it as requiring "an in-depth understanding of how AIs hallucinate" and as "possible only for expert users who already know what those parts ought to be."

This process names the specific labor that What tasks do expert data storytellers trust to LLMs? found shifting toward professionals in practice: that interview study found expert visual data storytellers moving from producing content to verifying it, and Samsonov's six steps give that verification work a general structure. It is also a different failure mode from the one in Why do human validation techniques fail against language models?: that note describes interrogation-style validation (cross-examination) as structurally unable to get traction on an LLM at all, while Samsonov's steps are a workable, if effortful, editorial procedure. The process is the workload that Does confident AI tone make users skip fact-checking? argues the chat interface gives users no incentive to actually perform.

The six steps are Samsonov's own design-practitioner enumeration, drawn from observation and professional cases rather than from a user study or task-time measurement. The excerpt gives no measured completion rate, time cost, or error-detection rate for any of the steps, so the claim that this process is more effortful than editing human-authored text is argued from mechanism, not measured. The taxonomy is useful as a design framework for where verification tooling could intervene, but a numeric how much harder claim would need separate empirical work.