AI chatbots fall into the same groupthink: the startup trying to break the mould

A feature in MIT Technology Review exposes a little-known limitation of large language models: their startling lack of variety when answering open-ended questions. The author illustrates it with a simple game: asking Claude, ChatGPT or Gemini for a random number between 1 and 10 almost always produces a 7; asking for a second number usually gives 3 or 4; a third, 8 or 9. The pattern repeats with other open-ended questions: asked to name a type of car, mainstream models tend to answer Toyota or Honda; asked for a slogan for a pair of New Balance sneakers, both Claude and ChatGPT proposed literally the same line, "Run your way".

Australian startup Springboards, co-founded by Pip Bingemann (CEO) and Kieran Browne (CTO), has turned that problem into its starting point, flipping the usual debate about hallucinations on its head. "Most language models are fighting hallucinations. We welcome them," Bingemann sums up, referring to the need to encourage less predictable answers when what you are after is creativity rather than factual accuracy.

The article links this anecdote to a more rigorous academic study: the paper "Artificial Hivemind: The Open-Ended Homogeneity of Language Models (and Beyond)", presented in November and awarded best paper at NeurIPS, one of the most important AI conferences. The researchers asked 25 different models —including the leading US systems and open-source models from China and other countries— to each write, 50 times over, a metaphor about time. Of the 1,250 responses generated, the vast majority boiled down to variations on "time is a river" or "time is a weaver". The author of the feature contrasts this with the real diversity of his human colleagues, who, asked the same question, gave six different answers among six people.

According to the researchers, it is not entirely clear why this happens, although their hypothesis is that most current LLMs are trained in similar ways, on similar data and for similar tasks, which pushes them to converge on the same high-probability answers. Browne, Springboards' co-founder, offers another everyday example: ask a chatbot for a band name and words like "glass", "neon", "velvet" or "static" are almost certain to appear. On testing it, ChatGPT generated a list of 56 names headed by "Glass Harbor" and including "Static Empire", "Neon Hearts" and "Velvet Echo"; Gemini offered 15 suggestions, among them "Static Horizon". Even an apparently original ChatGPT proposal, "Sofa Astronauts", turned out to belong to a real, existing band.

Asked about the phenomenon, OpenAI replied that training models to give reliable, coherent answers tends to make them converge on familiar, high-probability responses, and that pushing harder for novelty can produce weaker or less reliable results. The company also noted that the "Artificial Hivemind" study analyzed 2024 models that have since been updated.

Against this diagnosis, Springboards has developed Flint, a model built on Qwen 3, the open-source system from Chinese giant Alibaba, since the company —as Browne explains— is too small to consider training its own foundation model from scratch. The team first tried the obvious route: turning up the "temperature" parameter, which governs the randomness of responses. But it found that maxing out the temperature on an OpenAI model produced incoherent answers, including abrupt switches from English to code mid-sentence. Springboards concluded that temperature is too blunt a tool: there is no point increasing randomness uniformly across the whole text, only at the specific points where there really is room for variety —for example, in the name of the destination when answering "where should I go in Europe?", not in every word of the sentence.

To achieve that, they trained their version of Qwen 3 so that the model itself identifies the points in a response where more variation fits and fills them with slightly more unexpected words or phrases. Maximilian Weigl, co-founder of the marketing agency Uncommon, describes the result this way: "Flint is programmed to slip in something unexpected. It's more of an invitation to think more broadly".

Springboards also sells a creativity-support tool aimed at advertising and marketing professionals, which combines several models —including ChatGPT and Claude— and lets users drag and combine fragments of text generated by each to build new ideas; Flint is offered within that tool as an alternative when more variety is wanted. Zoe Scaman, founder of the strategy consultancy Bodacious and chief strategy officer at 77X (the direct-to-fan marketing platform backed by basketball player Luka Dončić), says she put Flint, Claude, Gemini and ChatGPT through the same MBA-style case study —how to reinvent a financial company for a young audience— and that the three conventional models all landed on the hackneyed idea of "teaching financial literacy in a fun way", while Flint went straight to rethinking the very concept of wealth accumulation, an idea Scaman found genuinely interesting. Even so, she warns that Flint is still a prototype that "sometimes falls over" when pushed too hard.

Weigl, whose team uses Flint alongside ChatGPT, Claude and Gemini, sums up the usefulness of the approach by noting that "you can't create something boundary-breaking with tools that drag you toward the average", though he qualifies that in most real use cases that average is perfectly valid: people usually settle for something "good enough" and familiar to a mass audience. He also issues a broader warning about over-reliance on any AI, Flint included: if he saw someone on his team simply copying and pasting model-generated output, he would remind them to think, talk to other people and use their own voice.

For now Flint is aimed at advertisers and marketing professionals because they are Springboards' customers, but its founders insist that the lack of variety in chatbot responses is a problem that affects any user, not just that sector. Bingemann sums up the project's philosophy as a matter of offering options and letting the person judge whether the result is good or not: variety, he says, is valuable for sparking ideas, and the alternative is to let machines do everything and end up in "a gray, boring world".