LLMs Stuck in Groupthink: How Springboards' Flint Aims for True AI Creativity

Large Language Models (LLMs) like ChatGPT, Claude, and Gemini, despite their advanced capabilities, often exhibit a surprising lack of creativity and a tendency towards "groupthink." A simple experiment reveals this: asking for a random number between 1 and 10 frequently yields '7' from most mainstream chatbots. Subsequent requests often produce '3', '4', '8', or '9'. While this predictability can be useful for structured tasks like coding, it becomes a significant limitation for open-ended creative processes such as brainstorming or planning a unique vacation. This inherent bias towards common, high-probability responses restricts their utility in scenarios demanding genuine novelty. Addressing this pervasive issue, the Australian startup Springboards has introduced an innovative solution: an LLM named Flint. Unlike its mainstream counterparts that actively "fight hallucinations," Springboards cofounder and CEO Pip Bingemann states, "We welcome them." Flint has been specifically trained to generate a much wider array of responses to open-ended queries, aiming to break free from the conventional patterns. Bingemann vividly demonstrated Flint's distinctiveness by re-running the random number game, where Flint produced "3.7916" after ChatGPT and Claude both defaulted to "7" in a new session, highlighting its ability to diverge from the norm. The divergence isn't limited to numbers. When asked to name a type of car, mainstream LLMs predictably suggested Toyota or Honda, while Flint offered a Ford F-150. Similarly, for a New Balance tagline, ChatGPT and Claude both gave "Run your way," whereas Flint proposed "Built to last, run to win," showcasing a different perspective. This "lost information" and inherent bias in mainstream models, where they are capable of more diverse answers but simply don't provide them, is gaining academic attention. A paper titled "Artificial Hivemind: The Open-Ended Homogeneity of Language Models (and Beyond)," which won an award at NeurIPS, confirmed that LLMs, trained on similar data and methods, converge on remarkably similar answers, even across different models and origins. For instance, when asked to create a metaphor for time, most of 1,250 responses across 25 LLMs were variations of "Time is a river" or "Time is a weaver." Kieran Browne, cofounder and CTO at Springboards, notes that this repetition is ubiquitous, often unnoticed by users who perceive personal conversations with chatbots. Another example is band names, where mainstream models frequently suggest terms like "glass," "neon," "velvet," or "static," leading to generic and sometimes already existing names (e.g., "Sofa Astronauts"). OpenAI acknowledges that training for reliability can lead to convergence on high-probability responses, and pushing for novelty might reduce reliability. Springboards offers a tool for creative professionals, backed by various LLMs, allowing users to combine ideas. Flint is positioned as an alternative within this tool for those seeking greater variety. Zoe Scaman, founder of Bodacious, praises Flint for "throwing me in completely different directions," finding it invaluable for catapulting her thinking. Scaman tested Flint against Claude, Gemini, and ChatGPT with a classic MBA case study: how would you reinvent a finance company for today’s youth? While the mainstream models proposed predictable solutions like "teaching financial literacy in a fun and funky way," Flint offered a truly novel approach: rebranding the entire concept of wealth accumulation. This distinct perspective, though still a prototype that "sometimes falls over when you start pushing it too far," underscores Flint's powerful premise. By challenging the "groupthink groove" of current LLMs, Springboards aims to unlock a new dimension of AI creativity, offering a vital tool for industries where original thought is paramount.