We have been using generative tools in production work for about two years. The honest summary is that they changed several things substantially, changed nothing about the hardest parts, and created two problems we did not anticipate.
Where the gains were real
Variation, overwhelmingly. Producing thirty versions of a headline, or twenty crops of an asset for different placements, or localised variants, is where the time went and where it stopped going.
This matters more than it sounds because variation is genuinely valuable in paid media. Creative wears out, and having many executions extends the useful life of a campaign considerably.
Previously the constraint was production cost, so campaigns ran with three executions until they stopped working. Now they run with twenty.
Rough concepting also improved. Producing a visual approximation of an idea to show a client, before committing to a shoot, is fast and cheap and removes a lot of misunderstanding early.
And unglamorous production tasks — cutting out backgrounds, extending images to new aspect ratios, cleaning up audio, generating rough subtitles — went from hours to minutes.
Where there were no gains
The strategic work, entirely. Deciding what a campaign should say, who it should reach, and why anybody should care is unchanged.
The genuinely original idea. Everything we generated was competent and derivative, which is what these systems do — they produce the expected thing, and the expected thing is by definition not distinctive.
Anything requiring specific knowledge of a client's business, market position, or the reasons behind previous decisions.
And the judgement about which of thirty options is the good one, which turns out to be the bottleneck once generating options is free.
The first problem we did not anticipate
Volume without discrimination.
When producing an option costs nothing, the number of options rises, and reviewing them costs the same as it always did.
We spent several months producing enormous quantities of adequate material and then spending longer than before selecting from it.
What fixed it was applying the same discipline as before generation — deciding the criteria first, generating a small number against those criteria, and stopping. The temptation to generate more is strong and it is nearly always a way of avoiding a decision.
The second problem
Sameness, and it is visible across the industry rather than only in our work.
These tools produce outputs clustered around the average of what they were trained on. Everybody using them converges toward the same aesthetic, the same phrasing, the same compositional habits.
Which means the work is competent and increasingly indistinguishable, at exactly the moment that distinctiveness is the thing brands are paying for.
Our response has been to use the tools for production tasks and deliberately not for anything that carries the brand's character. The photography that matters is shot. The line that matters is written by a person.
That is an expensive position and I think it is the correct one.
The disclosure and rights questions
Practical matters that have real consequences.
Rights in generated material are unsettled in several jurisdictions and vary by how the tool was built and licensed. Using generated imagery in a campaign without establishing what rights you actually have is a risk that legal departments have started taking seriously.
Several advertising regulators have issued guidance on disclosure where generated imagery could mislead about a product, and the direction is toward more requirement rather than less.
Depicting people who do not exist, or generating something resembling a real person, carries obvious risk.
Our working rule is that anything representing a product, a result or a person is shot or licensed, and generated material is used for backgrounds, textures and abstractions where nothing is being claimed.
The effect on the team
Worth being honest about since it is the part people ask about.
We did not reduce headcount. The work shifted — less time on execution, more on selection, direction and the strategic work that did not change.
The junior roles changed most, because a lot of what junior people used to do to learn the craft is now automated. That is a genuine concern for how anybody develops, and I do not have a good answer to it.
The people who have gained most are the ones with strong judgement and moderate technical skill, because the technical constraint has weakened and the judgement constraint has not.
What we tell clients
Since disclosure comes up in every conversation now.
We state which elements were generated and which were not, in the delivery documentation, as a matter of routine.
Nobody has objected. Several have asked for more generated variation once they understood the cost implication.
The clients who have restricted it have done so for specific categories — anything depicting a product in use, anything depicting people, anything making a claim — which is broadly the same line we had arrived at independently.
Being the party that raises it rather than the party that gets asked has been consistently better for the relationship.