Ask ChatGPT, Claude, or Gemini for a random number between 1 and 10 and you’ll almost always get 7. Ask for a band name and you’ll get some combination of glass, neon, velvet, or static. This isn’t a coincidence, it’s a real, measured called an artificial hivemind, and an Australian startup called Springboards has built a model specifically designed to break out of it.
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
Major Highlights
- Springboards built Flint, a model trained on top of Alibaba’s open-source Qwen 3, specifically to give more varied answers to open-ended questions than mainstream LLMs.
- A NeurIPS best-paper-winning study, Artificial Hivemind, tested 25 different LLMs (including major US and Chinese models) and found striking convergence, when 1,250 responses to “write a metaphor about time” were collected, most were variations of “time is a river” or “time is a weaver.”
- Researchers believe the cause is that most LLMs are trained in similar ways on similar data to do similar tasks, producing similar outputs.
- Rather than simply raising the temperature setting (which tends to make output incoherent), Springboards trained Flint to identify specific points in a response where variety helps, and inject randomness only there.
- Marketing and advertising professionals testing the tool say it’s useful for breaking out of default brainstorm patterns, though it’s still a rough prototype that falls over under heavy pushing.
- OpenAI’s response: training for reliable, coherent answers naturally pulls models toward familiar high-probability responses, and pushing harder for novelty can weaken reliability.
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
KINI BIG DEAL
This is one of those stories that sounds like a party trick until you sit with the implication: the tools millions of people now use for brainstorming, ideation, and creative work are unknowingly homogenizing the outputs of an entire generation of marketers, writers, and strategists, without most users realizing it. If you’ve ever felt like AI-generated brainstorms all sound vaguely the same, this is why, and it’s not your prompting, it’s the models.
For anyone doing content or creative work with AI, the practical lesson isn’t “stop using ChatGPT and Claude,” it’s “don’t mistake the first answer for the only answer.” The Springboards team’s approach, deliberately engineering variety rather than just cranking up randomness, is a useful mental model even without their tool: push past the first 2-3 responses a model gives you, because those early responses are statistically the most average ones by design. The tools reward you for iterating past the obvious; most people just don’t.
