Most large language models will give you the number 7 if you ask for a random integer between 1 and 10. This is not a glitch. It is a pattern that holds true for ChatGPT, Claude, and Gemini. An Australian startup called Springboards has built a new model named Flint to break this habit.
Breaking the pattern
Pip Bingemann, the chief executive of Springboards, demonstrated the issue during a meeting. He asked the three major models for a random number. Each returned 7. He restarted the session and asked again. The results were 7, 7, and 3.7916. Bingemann noted that the mainstream models were predictable, while Flint offered a distinct alternative.
The problem extends beyond numbers. When asked to name a car, the major models consistently suggested Toyota or Honda. Flint proposed a Ford F-150. Bingemann stated that the other models are biased against less common answers, even though they possess the capability to generate them.
In a test involving a tagline for New Balance running shoes, the results highlighted the divide. ChatGPT and Claude both produced the phrase “Run your way.” Flint generated “Built to last, run to win.” Bingemann described the approach as welcoming hallucinations rather than fighting them.
Artificial hivemind
Researchers published a paper in November titled “Artificial Hivemind: The Open-Ended Homogeneity of Language Models (and Beyond).” The study found that different models converge on very similar answers when prompted with open-ended questions.
A team of researchers asked 25 different models to write a metaphor about time. Out of 1,250 responses, most were variations of “Time is a river” or “Time is a weaver.” The team won the best paper award at NeurIPS in 2025.
Kieran Browne, the chief technology officer at Springboards, says this repetition is visible in everyday use. He notes that users often do not realise they are receiving the same suggestions as everyone else. For example, when asked for band names, most models returned lists containing words like glass, neon, or velvet.
When I asked ChatGPT for 56 band names, the top suggestion was “Glass Harbor.” Other options included “Static Empire” and “Neon Hearts.” Gemini provided 15 suggestions, including “Static Horizon.” I found that a band called Sofa Astronauts already exists, despite the name appearing in a list generated by ChatGPT.
OpenAI states that training models to provide reliable and coherent answers leads them to converge around familiar, high-probability responses. The company noted that the “Artificial Hivemind” paper studied models from 2024 that have since been updated.
Creative catapult
Springboards has developed a tool for creative professionals in advertising and marketing. The system allows users to drag text produced by different models and combine the parts they like. Flint acts as an alternative option for users seeking more variety.
Zoe Scaman, founder of the business strategy startup Bodacious, uses the tool to explore different directions. She described the experience as catapulting herself all over the place.
In a test, Scaman asked the models to reinvent a finance company for today’s youth. The mainstream models suggested teaching financial literacy in a fun and funky way. Flint suggested rebranding the concept of wealth accumulation entirely. Scaman found the latter suggestion particularly interesting.
She noted that Flint is still a prototype and does not work every time. It sometimes fails when pushed too far. However, she believes the premise behind the model is powerful.
Taking the temperature
Springboards built Flint on top of Qwen 3, an open-source model from Alibaba. Browne explained that training a foundation model is too expensive for a small team.
Most LLMs have settings to adjust randomness, often called temperature. Increasing this setting can make models incoherent. Browne noted that dialing up the temperature on an OpenAI model caused it to switch from English into code halfway through a sentence.
Springboards realised that adjusting parameters across the board is not effective. The goal is to boost randomness at specific points in the output. For example, when asking where to go in Europe, the model only needs to tweak the randomness before naming a destination.
To achieve this, Springboards trained its version of Qwen 3 to identify points where more variety was possible and fill those spots with slightly random words. Maximilian Weigl, a chief strategy officer at Uncommon, described the approach as an invitation to think wider.
Weigl’s team uses Flint alongside ChatGPT, Claude, and Gemini. He stated that you cannot create something boundary-breaking with tools that pull you back to the average. He also noted that nine times out of 10 the average is fine, as most people want to see mass-market familiar things.
Weigl cautioned against relying on AI output too much. He said that if he saw people on his team copy-pasting from AI, he would tell them to think, talk to other people, and use their own voice.
Flint is currently aimed at advertisers and marketers. Bingemann and Browne insist that a lack of variety is a problem for anyone using chatbots. The goal is to give people the choice to decide if the result is good or not.
Bingemann said variety is great when trying to spark ideas. He argued for exploring new routes instead of letting machines do everything and ending up in a gray, boring world.
