Marketing Strategists Confront AI's Sameness Problem
Marketing strategists increasingly rely on generative AI tools like ChatGPT and Claude for research, brainstorming, and idea critique. But as these tools embed deeper into agency work, a constraint has emerged: their outputs trend toward predictability, limiting the fresh thinking clients actually pay for.
The problem stems from how large language models work. Tools like ChatGPT select the most probable sequences of words and data, which means their answers gravitate toward the average. Popular LLMs also train on similar data sources, compounding the issue.
"You don't get those sideways, magic moments," said Lorna Hawtin, chief strategy officer at creative agency Zeal. "There's an electricity that can sometimes pull two thoughts together from seemingly distant poles in your brain. You miss that edge with AI."
How Strategists Are Working Around It
Some practitioners have developed workarounds. Zoe Scaman, founder of strategy studio Bodacious, spent 18 months customizing prompts and data to "jailbreak" Claude into producing less predictable outputs. Nick Myers, chief strategy officer at Oliver's U.K. office, said his team now uses multiple LLMs to build AI agent personas, spreading the net wider to dodge what strategists call the "sameness trap."
Others use AI selectively. Scaman built custom projects within Claude, including one called "My Own Worst Critic" designed to reject weak ideas. About 80 percent of her writing gets killed by that project. She created another tool based on sci-fi author Ursula K. Le Guin's published work to push her writing into unexpected directions.
At Zeal, the team designed a custom "Banging Brief Bot" within Claude to critique strategy briefs against house standards. The tool handles fact-finding, cultural research, and semiotic analysis.
A Technical Solution Emerges
But custom prompts can't fix what's built into the architecture of major models. Researchers at Carnegie Mellon University developed NoveltyBench in 2025, an open-source test for measuring response variety across LLMs. They found that larger models produced less diverse answers, with tools like Llama and Gemini showing "a fundamental lack of distributional diversity."
Sydney-based startup Springboards built a smaller language model called Flint, trained on 30 billion parameters compared with the 2 trillion used by the largest Llama models. The model prioritizes variation and unpredictability.
On the NoveltyBench framework, major LLMs averaged 2.88 out of 10. Flint scored 7, according to Pip Bingemann, co-founder and CEO of Springboards. "It's designed for people to expand their latest knowledge or spark an idea in a way which doesn't give you the same answer as every single person," he said.
Maximilian Weigl, co-founder and chief strategy officer at Uncommon, tested Flint before its release. "It pushes you out of your comfort zone," he said. Rather than becoming another mental crutch, Weigl suggested such tools could provide "healthy competition" that challenges strategists to do better work.
The Underlying Question
Still, some practitioners question whether strategists should chase technical fixes at all. Clients hire agencies for original thinking - the very thing ChatGPT and similar tools struggle to provide.
Nick Myers put it plainly: "Are strategists going to be replaced? If you're not feeding yourself, then yes. But if you can bring something to the table and use the tool as a partner, you have a better chance of having a future."
Your membership also unlocks:
