On Monday morning, a cultural cooperative needed to present three ideas for a new membership campaign. They had set aside an hour to explore and another to choose. Someone opened a writing assistant and asked for headlines about belonging, music, and neighbourhood life. Within seconds, twenty polished options appeared. “Your story deserves a stage,” said one. “Together we make the city resonate,” said another. Nobody thought they were bad. That was the difficulty: they sounded so plausible that pausing to ask what story the group actually wanted to tell felt unnecessarily slow.
Creative AI can shorten the empty moment at the start of a task, suggest structures, or offer combinations a team has not considered. It does not have to be an enemy of a newsroom, a design studio, or a communications team. Yet its fluency creates a specific trap. A sentence that sounds finished may receive the same weight as an idea developed over several days, even though it knows nothing of the organisation’s conflict, the place where it will be read, or the commitment it creates. When the first proposal goes straight to the decision table, the tool is no longer assisting. It is taking the place where human judgment ought to happen.
The comfort of familiar language#
Generative models continue patterns. Ask for a campaign about community and they will often return the vocabulary that many campaigns have already attached to community: connection, future, impact, closeness, transformation. Those words are not automatically false. The risk begins when familiarity is mistaken for relevance. A text can be grammatical, warm, and appealing while saying nothing that only this cooperative, library, neighbourhood group, or organisation could say.
Generative models continue patterns. Ask for a campaign about community and they will often return the vocabulary that many campaigns have already attached to community: connection, future, impact, closeness, transformation.
At that meeting, the machine’s suggestions avoided an awkward fact: the new memberships cost less because the theatre had created an afternoon slot for people with caring responsibilities. That detail changed the campaign’s meaning. The task was not to declare that culture brings people together. It was to explain that an institution was changing a rule that had excluded part of its audience. The generic response felt safe because it did not require the team to own that change or show how it worked. They realised they could publish one of those headlines without lying, but also without being accountable for what it implied.
Creativity is not always the search for an unusual line. Often it is the work of choosing what deserves to be said, from which point of view, and within which limits. That choice has consequences. A sentence describing a programme as accessible can disappoint someone who arrives to find that the only ramp is at the back. An advertisement about diversity can use people as scenery. A slogan about innovation can hide the fact that nobody knows who will respond when a service fails. The tool does not experience the gap between a promise and the later life of the person who receives it. The team should be prepared to notice and address that gap.
Keep exploration separate from commitment#
One helpful practice is to name two phases with different purposes. During exploration, an assistant may produce abundant options, including predictable ones. A team can ask it for contradictions, questions to take into an audience interview, ways to avoid a patronising tone, or narrative paths that are not yet final copy. The material belongs on a board of hypotheses, not on a tray of approved lines. The more clearly this condition is stated, the less automatic authority the first generated wording receives.
The second phase should begin with the screen out of the centre. The team can select three possibilities and ask: what verifiable fact supports each one? Who is left out? What would a person who does not know our context understand? What would happen if we had to explain this sentence to someone who received poor service? These questions are not designed to punish a tool or turn every task into a long ceremony. They recover work that urgency often hides: deciding what we will put into the world under our own name.
In the cooperative’s case, a designer suggested bringing notes from conversations with members who care for relatives. They were not decorative testimonies or a source from which to imitate voices. They were context for testing whether the team’s language described a real possibility. That led to a different line: “An afternoon that fits into your week too.” It was neither perfect nor heroic. It made room to talk about hours, tickets, and conditions. The group could then review it with the people involved and revise it. The campaign gained precision because nobody treated the first plausible answer as a verdict.
A record is not needless administration#
Keeping iterations makes judgment visible. A shared document with the initial request, discarded alternatives, reasons for rejection, and the approved version is enough. That record need not become an endless archive or expose sensitive conversations. Its practical purpose is to let someone understand, weeks later, why a particular wording was chosen and which assumptions were tested before it was used.
The record also prevents confused attribution. If an AI suggested an opening, one person reshaped it, and another checked that it did not promise something impossible, the team should know where each intervention took place. This is not about defensively distributing blame. It is about knowing who can correct a decision when circumstances change. Saying “the AI generated it” does not answer the relevant question. The piece does not reach the public signed by an algorithm; it reaches them under the responsibility of an organisation and the people who approved its use.
That clarity matters even more when text touches health, education, employment, housing, or safety. In those areas, an elegant formulation may shape expectations, decisions, or fears. An assistant may help organise a draft, but it cannot judge whether a promise is prudent, whether a metaphor stigmatises, or whether an omission leaves someone without an alternative. The greater the possible effect, the more a situated human review is needed: someone who knows the service, can consult people with knowledge, and has authority to stop publication.
Find clichés before they become policy#
Clichés are not always individual words. Sometimes they are a structure: begin with a universal emotion, promise transformation without a cost, and finish with an empty invitation to join in. Reading a draft aloud and underlining anything that could belong to any organisation is a useful first filter. Another is to remove the campaign name. If the body works unchanged for a delivery app, a bank, and a museum, the problem is not only style. It is missing context.
The team can then look for the friction that makes its proposal true. What constraint has it accepted? What practice will change? What can it not offer yet? What should a particular person be able to do after reading? The answers should improve the copy, but they may also change the project. Perhaps the campaign should not promise support until an actual support channel exists. Perhaps the message needs to acknowledge a limitation. Perhaps it should wait. Being able to say “not yet” is a valuable part of creative responsibility, not a failure of productivity.
Back in that meeting, the group chose a rule instead of the slickest sentence: no generated line could be used unless someone could defend it with an experience, an internal source, or a decision already made. They would also reserve twenty minutes to make a proposal without an assistant before comparing. The machine remained useful for exploring variations and spotting repeated language, but it stopped being the place from which the campaign’s meaning was defined.
The most important measure is not banning a tool or romanticising manual writing. It is retaining both the right and the duty to judge. When an impeccably plausible answer appears next time, the team can pause before copying it and write down three things: what question it opens, which concrete reality it names, and who will answer if its promise fails. If there is no answer, the line is not finished thought. It is working material. Treating it that way protects creativity from something more dangerous than an obvious error: the habit of silently delegating judgment.
