The Artificial Hivemind: Why Your Chatbot Is Less Creative Than You Think

Ask your favorite chatbot to pick a random number between one and ten. Chances are, it will say seven. Ask it to name a car brand, and you’ll likely get Toyota or Honda.

When we type a prompt into an interface like ChatGPT or Gemini, it feels like a private, highly personalized brainstorming session. In reality, you are likely receiving the exact same statistically safe answers as millions of other users.

This phenomenon is becoming known as the "Artificial Hivemind"—a term coined by a team of researchers who recently won a best paper award at the major AI conference NeurIPS. To test the creative limits of modern algorithms, they asked 25 different large language models to write a metaphor about time. Out of 1,250 responses, the vast majority essentially parroted the same two concepts: "Time is a river" or "Time is a weaver."

The root of this extreme predictability lies in how modern AI is built. Because developers train these models to be helpful, reliable, and coherent, the algorithms naturally gravitate toward the highest-probability responses. OpenAI itself has noted that pushing a model too hard for novelty can compromise its reliability. While this safety-first approach is perfect for summarizing legal documents or writing Python code, it severely bottlenecks creative thinking.

An Australian startup named Springboards is trying to build an escape route from this groupthink. They have developed a new model called Flint, designed specifically to prioritize variety over predictability. Unlike mainstream tech giants that spend millions trying to eradicate AI "hallucinations"—instances where the model makes things up—Springboards deliberately embraces them.

The results are noticeably different. When asked to write a tagline for New Balance shoes, both Claude and ChatGPT generated the exact same phrase: "Run your way." Flint, however, offered "Built to last, run to win." In a more complex test, business strategist Zoe Scaman asked several models how to reinvent a finance company for young people. While the major models uniformly suggested teaching financial literacy in a "fun" way, Flint proposed a radical rebranding of the very concept of wealth accumulation.

Flint is still an early prototype, and its creators admit that pushing the boundaries of AI creativity sometimes causes the model's logic to break down entirely. However, its existence highlights a crucial shift in how we might use AI in the future. Rather than relying on a single omnipotent chatbot for everything, we may soon need a diverse toolkit: strict, predictable models for our spreadsheets, and slightly unhinged, highly creative ones for our blank canvases.

Key Points

• Major LLMs suffer from groupthink, often giving identical answers to open-ended prompts.
• This predictability is a side effect of training models to be highly reliable and coherent.
• A NeurIPS-winning study showed 25 different models largely gave the exact same metaphors for 'time'.
• A startup called Springboards created Flint, an AI that prioritizes creative variety by embracing 'hallucinations'.

Why It Matters

Recognizing the predictable nature of mainstream AI prevents us from outsourcing our creativity to algorithms that are mathematically designed to be average.

Sources:

• LLMs are stuck in a groupthink groove. This startup is trying to get them out. — MIT Technology Review - AI