Who should have the first idea? Human planning in AI-assisted creativity
Putting a human at the end of an AI creative workflow is not the same as letting them shape its possibilities.
A creative team can put a human in every approval meeting and still give that person remarkably little influence over the work. If the model proposes the initial direction, develops the options and polishes the strongest candidate, the human may be left to choose among possibilities that have already been defined. That is a form of oversight. It is not the same thing as authorship of the plan.
The distinction is worth examining because “human in the loop” describes a position in a workflow, not a guarantee of meaningful creative agency. A person who sets the problem and decides which directions deserve exploration has a different role from a person who is invited to approve a finished result.
Consider two hypothetical teams responding to the same campaign brief. In one, a strategist identifies the audience tension, chooses three genuinely different starting premises and asks a generative system to develop each. In the other, the team asks the system for a campaign, receives ten polished options and chooses the most convincing. Both teams use AI. Both involve a human. But they distribute the power to define the creative search space differently. This is an illustration of process design, not an observed experiment or proof that one team will outperform the other.
Planning is an intervention, not a ceremonial first step
Creative production involves decisions at different levels. Planning sets the direction: what problem the work will address, what assumptions it will challenge, what audience interpretation matters and which alternatives should be pursued. Writing or making produces a candidate. Reviewing judges the candidate. Revising changes it. Those actions interact, but they are not interchangeable.
An October 2026 preprint by Chenyi Li and colleagues makes the distinction unusually concrete. The researchers constructed a creative-writing process with four stages: Plan, Write, Review and Revise. Across 333 participants and 427 stories, they compared all-human, all-AI and mixed human-AI chains. In that experimental setting, all-AI chains scored higher on individual-story creativity but produced less diversity than all-human chains. When a single human was introduced into an otherwise AI chain, the role mattered. Human planning offered a favorable balance between creativity, diversity and enjoyment; the increase in diversity persisted through later AI stages. These are the authors' reported findings in a preprint, not a settled rule for professional design. https://arxiv.org/abs/2610.37916
The result changes the practical question. Asking whether a person should be included is too imprecise. A useful team-design question is where that person's judgment has the highest leverage. If the concern is a narrowing range of starting ideas, a late approval step may be an inadequate intervention because it evaluates a set of options that has already been narrowed.
But this does not mean the first human idea is automatically good, or that models cannot contribute to planning. A person can begin with a cliche. A model can generate an unexpected reframing. The relevant difference is who is accountable for noticing missing alternatives and expanding the search when the first options are too similar.
Originality is not the same as diversity
One strong idea and a diverse set of ideas are different outcomes. A model may produce an impressive individual artifact while repeatedly returning to familiar structures across many outputs. A team evaluating only its favorite candidate can miss that pattern. Conversely, a large collection of unusual ideas can be incoherent, irrelevant or impossible to execute. Neither a novelty score nor the number of generated concepts is a sufficient measure of useful creative work.
That is why the Li et al. study should be read as a domain-specific test of role assignment, not a claim that AI always homogenizes every field. Its tasks were creative stories. The participants, scoring methods and transmission-chain design constrain what can be inferred. A design studio, product strategy team or advertising agency has different materials, constraints and measures of success. Transfer to those settings is a hypothesis that needs testing.
A second 2026 preprint, <i>Humans Introduce, Models Elaborate</i>, studies narrative contributions in human-human, human-model and model-model co-writing. It describes a tendency for humans to introduce novel and persistent narrative elements while models elaborate them. This is relevant to the distinction between opening a direction and developing one, but the co-writing setup does not isolate the causal effect of assigning planning authority in a commercial workflow. https://arxiv.org/abs/2609.07920
There is also a counterweight to the simple story that AI inevitably reduces creative breadth. Research involving advertising professionals has examined how people use generative AI in ideation, including the effect of training and prompt practices. The existence of that work does not establish that every trained team produces more original work; its design, sample and observation period need to be inspected before making a present-day performance claim. It does, however, make a useful conceptual point: the quality of human-model collaboration depends on how people use the system, not merely on whether it is present. https://research.google/pubs/human-genai-collaboration-in-creative-ideation-a-study-with-advertising-professionals/
Who gets to define the alternatives?
In many organizations, the most consequential creative decisions are invisible in the final presentation. A team selects a brief, chooses which evidence to believe, decides which constraints are negotiable and determines which options never make it onto the board. A final reviewer sees the surviving candidates. They do not necessarily see the routes that were abandoned or never considered.
Generative systems can accelerate this narrowing. A prompt returns fluent possibilities quickly, and fluent possibilities are easy to treat as a reasonable representation of the option space. But a generated list is an output of the prompt, model and context, not a census of what could be imagined. Asking for more items may broaden the list; it does not guarantee a change in the underlying frame.
The practical remedy is not to ban the model from proposing ideas. It is to make the frame itself reviewable. Before production, write down the strategic premise and at least two alternatives that genuinely disagree with it. During generation, preserve the origin of each direction. At review, ask which initial assumptions survived unchallenged. This makes human planning a continuing responsibility rather than a single ritual at the start.
Design the workflow around the risk
Different creative tasks deserve different allocations of human attention. When a team needs many variations of an already validated direction, generation and rapid iteration may be the point. When it needs to discover the direction, a human should remain responsible for the problem definition and for deciding whether the range of possibilities is wide enough. When work has legal, reputational or ethical consequences, final review still matters even if the human planned the project from the start.
A workable process can be organized around four questions. First: who can change the brief? Second: who can introduce a direction that the initial prompt did not anticipate? Third: what evidence will distinguish a useful idea from a merely polished one? Fourth: when does the team stop generating and commit to a test in the world? These questions assign decision rights, not decorative roles.
The strongest case against privileging human planning is straightforward. People bring biases, habits and institutional constraints of their own. Sometimes the model's first proposal may be the one that challenges the team. That possibility argues for reversible planning and deliberately adversarial alternatives, not for removing human accountability. A human-led plan should be open to revision by new evidence and by genuinely surprising generated options.
The useful conclusion is narrower than “humans must always have the first idea.” A creative process should protect the ability to question its starting point. Where a person enters the process changes what they can influence. If they arrive only after the options have been produced, the work may be supervised without ever having been meaningfully directed.
Research boundaries
- Li et al., October 2026 preprint: experimental creative-writing transmission chains, 333 participants and 427 stories; not a universal claim about professional design. Exact observation dates are not established in the public abstract.
- Humans Introduce, Models Elaborate, September 2026: separate narrative co-writing setting; no causal claim about commercial design work.
- Google Research, advertising-professional ideation study: conceptual counterpoint; no undated empirical performance claim is made.
The campaign teams are hypothetical. Workflow recommendations are author synthesis.