AI is not creative direction

Many projects now begin by opening a generator before the problem has been written down. Images, shots, temporary logos, textures and possible worlds appear within minutes. Volume feels like progress, but it does not clarify who the work is for, which tension should hold the idea together or what response the project is meant to produce.

That reversal is the most interesting risk of generative AI. It does not only replace an execution step; it moves forward the moment when a project takes shape. If the group mistakes the first convincing image for a decision, the model has already written part of the brief. Not because it has intent, but because it has made one possibility far more visible than the rest.

Generating many routes is useful when the team knows what it is looking for. It can surface contrasts, materials, compositions or registers that an orderly moodboard would have excluded. At this stage the value lies in divergence: the image is not a proposal to polish, but an object that makes the discussion more precise.

Creative direction starts afterwards. It means recognising what genuinely belongs to the project, removing what is seductive but generic, and turning an intuition into a system that can support language, motion, interfaces, formats and real production. A model can widen the table; it cannot take responsibility for ending the conversation.

A reference should name a quality: the density of a light, the rhythm of an edit, the imperfection of a material or the emotional distance of a scene. When it becomes a complete visual target, the project stops interpreting and starts imitating. AI makes that shift even easier because resemblance can be produced before the team has found its own language.

To avoid that loop, separate references by function and state what should be retained from each. One may describe tone, another structure, a third behaviour. None should supply the entire answer. If the prompt already contains the style, the author and the expected outcome, the search has probably ended before it began.

Since 2 August 2026, some European transparency rules for AI-generated or manipulated content have been in application. They do not make every creative project the same case, but “we used a little AI” is no longer an adequate production note. A brief should identify tools, source material, human interventions, review steps and where the work will appear so the production team and client can assess the specific obligations and risks.

Provenance is also becoming a design choice. The C2PA standard provides a way to attach verifiable information about a file’s history. It does not certify that content is true or good, but it preserves useful signals about origin and edits. The US Copyright Office report, while not determining European law, highlights the same operational issue: using a tool is not the same as delegating the expressive contribution to it.

Before generating anything, write one sentence the project must make true. Then give AI a bounded task: explore material variations, stress-test a composition, produce elements that will be redrawn or previsualise a transition. Avoid vague commissions such as “find the idea” or “make it premium”, because they hand the model the exact part that should distinguish the work.

During exploration, keep prompts, inputs, versions and the reasons behind decisions. Every file does not need a formal audit trail, but the route to the outcome should be reconstructable. Before delivery, someone other than the generator checks consistency, rights, any required disclosures and behaviour across formats. Speed remains an advantage only when review is not compressed alongside production time.

Finally, rebuild the result with the tools and expertise the project needs. Design, photography, CGI, editing, writing and development do not become redundant steps. They are where a statistical possibility gains intent, precision and continuity. If nobody can explain what was decided after generation, there is probably no direction yet.

Adding small imperfections to hide an image’s origin is a weak strategy. Work does not become more credible because it imitates an analogue flaw or avoids the model’s most recognisable mistakes. It becomes credible when every choice serves the same idea and someone can answer for it.

AI can play an explicit, even central role. But the project should be able to explain why the technology is there, what it makes possible and where the boundary was set. Creative direction is not the amount of manual control. It is the quality of decisions and the ability to make them legible.