Why AI Defaults to Predictable Advertising Ideas (Part 1)

The core problem we observed is that, for a given brand, product or context, these models tend to circle the same few creative concepts. Their output can quickly become a sea of sameness.

Using advanced prompts and multi-step workflows might appear to be an obvious solution, but the result is frequently a variation of a predictable theme or of a limited range of themes.

Take a simple example: an ad for a relaxation product.

• The AI's predictable idea: In our tests, most suggestions were shots of people meditating, serene in nature or lounging on a sofa with a peaceful expression. It is the most obvious, common association.
• What an LLM is unlikely to suggest on its own: Imagine a close-up on a person's forehead, glistening with sweat. The background is a blur of intense, fast-paced urban life. Suddenly, the scene transitions. The sweat is gone, the expression is relaxed, the background is calm. The product is the bridge between chaos and tranquility.

In our experience, an LLM rarely generates that second concept on its own: it relies on contrast and narrative tension rather than on the most common association. When we asked for many variants, most of them were tweaks of the first, stereotypical idea.

The problem can compound in multi-step workflows: when each step defaults to the most common option, the final output is even more generic.

Why Your AI Thinks Inside the Box

Two characteristics of how LLMs work help explain this: their training data and their probabilistic nature.

1. The Echo Chamber of Training Data

LLMs are trained on very large collections of existing text, much of it from the internet. They are very good at recognizing and recombining patterns from data they have already seen. This tends to produce what can be called a machine-oriented way of generating ideas.

Their suggestions gravitate toward the concepts that are most represented in their training data. The more common a concept, the more examples the model has seen, and the more likely it is to propose it. Common concepts are, by definition, already well known, and therefore rarely original. Left to its defaults, the model tends to favor the most statistically prevalent ideas.

2. The Trap of Probability

LLMs are probabilistic systems: they generate text by predicting a likely next word, based on the input they receive. Creativity is often about choosing a less likely, but still relevant, option to create surprise and intrigue. Default generation settings favor likely continuations, which works against this principle: the model tends to follow the beaten path rather than blaze a new trail.

Common Fixes and Their Limits

Many users try to coax creativity out of these models with a few common techniques. In our tests, they helped only marginally.

“Be More Creative” Instructions

Telling an LLM to “think outside the box” or “be super creative” has a limited effect. These instructions may slightly broaden the range of its suggestions, but the model still works within the patterns of its training data. In our tests, the result was usually a slightly unusual version of the same familiar idea rather than a genuinely new concept.

Parameter Tuning (Temperature and Top-P/K)

This is a more operational approach, but it is also limited.

• Top-P and Top-K limit the pool of candidate options the model samples from; raising them widens that pool.
• Temperature is a randomness parameter. A higher temperature makes the model more likely to pick less probable options from that pool.

Temperature is often mistaken for a “creativity” dial; it is closer to a “randomness” dial. Raising these parameters does not add new, imaginative ideas to the pool of options: it makes the selection among existing options less predictable. The output may be more “exotic”, but it is still drawn from the same set of possibilities, and at high values it can also become less coherent.

Using “Memory” to Exhaust Bad Ideas

Another common tactic is to use “memory” in a conversation. The strategy is to generate ideas, critique them and add them to a list of exclusions, forcing the model to exhaust the most obvious concepts and dig for more unusual ones.

Intuitively, this seems clever. In our experience, it is a brute-force approach that does not fix the core problem. You are still navigating the same map of machine-oriented ideas, crossing off the main roads first. You tend to reach a slightly less common destination by a longer, more convoluted path, rather than a genuinely original one.

This method also introduces friction. Either you engage in a long manual back-and-forth in a chat, which demands constant supervision, or you automate the exclusion process, which can lead to a slow and complex workflow that is costly to run and maintain, for a modest gain in originality.

So how can AI help produce more creative and diverse advertising assets? It takes a different approach: Part 2 describes how to inject curated creative knowledge into the workflow.