AI’s Secret to Creativity: Embracing Constraints for Surprising Results

By Diego Nishimura Oropeza, Head of AI & Analytics

AI took the world by surprise in late 2022 with the launch of ChatGPT. Large language models were just the beginning of a rapid wave of generative tools. Since then, AI has expanded to images, video, and audio—enabling entire ad campaigns to be produced in a fraction of the time and cost of traditional methods. 

What once took weeks can now be done in days or even hours. While early models felt gimmicky, the technology has matured. Some AI-generated content now rivals professional work; other outputs remain low-quality or unusable. Still, AI reduces the need for large teams, costly shoots, and physical locations, allowing small teams to produce full campaigns with minimal overhead. 

Despite these gains, AI still lacks true understanding. It predicts patterns based on training data, which can result in content that looks right but lacks depth, nuance, or coherence. 

To stand out in this evolving landscape, brands must use AI strategically. A helpful framework is dividing content into past-facing (assets based on existing materials) and future-facing (assets based on creativity). AI excels at past-facing content, where it draws from patterns and styles rooted in existing data. 

Past-facing content can be automated. 

Past-facing content includes assets based on existing materials, such as adaptations, versioning, and transcreation. Brands can train or prompt AI with proprietary tone-of-voice guidelines, customer data, or historical brand campaigns, resulting in more "on-brand" outputs. This type of production can be largely automated, enabling greater speed, efficiency, cost savings, and the ability to generate high volumes of personalised content.  

A great example of Past-facing assets is the: “Not Just A Cadbury Ad”, where machine learning was used to recreate India’s biggest brand ambassador face and voice to include local store names in the ads based from existing materials. 

What about future-facing content? 

True creativity – producing original, surprising, yet useful ideas – often requires breaking away from the obvious – contrary to LLMs predictable or generic answers – areas where human creativity still leads. Because AI is trained on similar datasets, many outputs sound or look similar.    

To shed light on their black-box nature, Anthropic published "On the Biology of a Large Language Model," using attribution graphs to examine how Claude 3.5 Haiku functions. 

One section in the paper explains how the model does additions. 

They asked the model to add 36+59. They found that instead of actually calculating, the model uses a heuristic, text-based approximation—associating numbers until the answer "feels" right instead of doing the math. 

When asked to explain its reasoning, Claude gave a textbook-style explanation that didn’t reflect its actual internal process. 

This shows AI is just predicting tokens, it can’t “remember” its own reasoning and it hasn’t developed mathematical abstraction. In other words abstract thinking.  

Creativity is based on abstract thinking. 

One way to measure creativity is called “Divergence of Thinking”. 

Divergence of thinking can be measured using the Torrance Creativity Test. This test asks participants to think of as many use cases as possible for an object within three minutes. The score combines the frequency of responses with their practical application. Use cases lose value as they become more frequent or as their practicality becomes absurd. 

Frequency is the keystone of AI. Asking AI for a creative response will generate the most common and least creative answer. This means AI is not inherently creative as it lacks abstract thinking. 

Without abstract thinking, can AI craft creative work? 

It already has. 

In 2006 researchers at NASA JPL evolved the "ST5 spacecraft antenna" using a computer simulation of Darwinian evolution. They set up structured rules that led to an unconventional and asymmetric antenna no designer would imagine. 

Necessity is the mother of invention 

Most of the creative work comes from constraints.  Music keeps reshaping itself to fit its speakers: Gothic cathedrals spawned soaring, slow-blooming lines from Bach; crackly AM radio crowned intimate crooners with bright , static-piercing hooks like Bing Crosby or Frank Sinatra; and today’s noise-canceling earbuds encourage producers to weave whisper-level details and immersive spatial effects for Spatial Audio. It happened the same with paintings and images, once the camera could "copy," painting began to "interpret." 

Researchers are finding that introducing structure into the generation process can paradoxically foster creativity in AI. Constraints eliminate the most straightforward or common solutions, forcing the model to explore less typical pathways. This researcher used a LLM to recreate different games by constraining the model with a 40-year-old hardware and a limited programming language to assess the model’s ability to adapt and reason creatively.  

Creative output may result not from giving models more (training data), but from cleverly limiting what they can do. 

Structured approaches in AI, especially with LLMs, show that we can steer generative systems toward greater creativity by carefully balancing guidance and freedom.  

The evolution of music, painting and most creative forms of art all reveal that structure and creativity are not opposites, but partners. A framework of rules or constraints can yield an explosion of diverse and creative outcomes when coupled with variation and feedback.