What tasks do expert data storytellers trust to LLMs?

The paper interviewed 12 expert visual data storytellers to ask which activities authors entrust to LLMs and which they "protect." Its central finding is that participants "rarely treated LLMs as autonomous storytellers." They "selectively delegate execution-oriented tasks" while "retaining control over activities that shape narrative intent and story meaning." Two further findings sit alongside this: LLM assistance is "most productive after human seeding and constraint-setting," and it "shifts labor from production to verification."

The discussion gives the reasoning behind the boundary. Delegation is described as "a nuanced negotiation" rather than a binary choice, with levels that vary by the specific activity, "perceived potential risks," and "desired authorial control." Authors are advised to decide up front which story components stay under human control, which the paper frames as preserving agency "without rejecting automation." Verification is the other half. Participants "treated model outputs as hypotheses to be checked rather than facts to be accepted," especially for claims, causal explanations, or audience-facing content. The paper therefore treats checking that statements and visualizations stay grounded in data, context, and intent as part of authorship, since "maintaining ownership requires" it.

This gives practitioner-side grounding to a pattern already in the library. The point that Can LLMs generate more novel ideas than human experts? argues that generation is cheap and evaluation stays human, and that human validation becomes the bottleneck. These interviews describe the same division of labor from the author's side, but the paper attributes the boundary to authors' own risk and control judgments, not to a claim about what LLMs cannot do. The verification stance also acts as a working counterweight to Does polished AI output trick audiences into trusting it?: authors who check outputs against data are declining to let a finished-looking result stand in for their judgment. The finding that help is most productive after seeding and constraint-setting is compatible with Can designers shape LLM behavior without deep technical knowledge?, where designer judgment is put in through structured prompt authoring before the model produces anything.

The excerpt does not say which tasks counted as "execution-oriented" or which as "narrative intent," how participants were recruited, or what tools they used. It reports no measure of time saved, output quality, or verification cost, so "shifts labor" is an interview-based description and not a measured trade-off. It also does not say whether the protected boundaries track real risk or only perceived risk. The supportable implication is narrow: for expert authors in this domain, tools that assume the LLM will run the story end to end do not match reported practice. Tools that help authors mark what they retain and that make grounding checkable match it better, which is the direction the paper's own design implications take.