If you have ever asked four chatbots the same open-ended question and gotten four nearly identical answers, you have run into the problem that a tiny Australian startup is trying to solve. Springboards shipped a fine-tuned build of Alibaba’s Qwen 3 named Flint that is trained to inject variety into its own answers only at the points where divergence is appropriate, rather than turning the randomness dial up across the board and hoping for the best. The MIT Technology Review profile by Will Douglas Heaven on July 1, 2026 is the cleanest write-up yet of what they did and why.
The gap the team is pointing at is no longer theoretical. For most of 2025 the major assistants have been converging on the same brand of polished, hedge-heavy prose, the same handful of stock metaphors, and the same careful non-answer to any open prompt. “Time is a river” has become the inside joke inside every AI lab: ask any model to write a metaphor about time and you get a dozen variants in the same register. For a writer, marketer, or strategist using these tools as a thought partner, that sameness is a feature gone wrong.
The Groupthink Symptom
The diagnosis now has a name. The paper “Artificial Hivemind: The Open-Ended Homogeneity of Language Models (and Beyond),” submitted to arXiv on October 27, 2025 by Liwei Jiang, Yuanjun Chai, Margaret Li, Mickel Liu, Raymond Fok, Nouha Dziri, Yulia Tsvetkov, Maarten Sap, Alon Albalak, and Yejin Choi, was given an Oral slot at NeurIPS 2025. The team built a dataset of 26,000 open-ended user prompts and collected 31,250 human ratings per example; the abstract documents both intra-model repetition and inter-model homogeneity and labels the pattern an “Artificial Hivemind” effect.
MIT Technology Review reported that the same research group earlier ran a simpler version of the experiment: 25 LLMs, 50 prompts apiece, 1,250 responses written about time. Most of those 1,250 answers collapsed to two or three variants of “Time is a river” or “Time is a weaver.” OpenAI, asked for comment by MIT Technology Review, pointed out that the underlying study looked at 2024 models that have since been updated, and argued that training for reliability naturally pushes models toward high-probability responses; novelty, the company said, often undermines reliability.
The tension here is real and not going away. Any model that is tuned for safety and helpfulness has been pushed toward the centre of its output distribution. The result is a generation of chatbots that are individually competent and collectively indistinguishable. Good enough that no single application wins on quality, similar enough that switching costs are vanishing. The question for the next two years is whether that flat landscape is a feature or a bug, and who can fix it without breaking the rest.
What Flint Actually Does
Springboards’ answer is a training-time tweak that targets the moments of choice inside a single response. As MIT Technology Review reports, the team taught their Qwen 3 build to flag positions in its own draft output where a more varied word or phrase was acceptable, then sampled one of those alternatives at exactly those points. Global temperature cranking, the usual fix users try themselves, was tested and rejected; CTO Kieran Browne told MIT Technology Review that dialing the temperature on an OpenAI model to its maximum setting produced responses that “switched from English into code halfway through a sentence.”
The demos in the article are small, vivid, and easy for a reader to reproduce. A “random number 1-10” prompt returned 7 from both ChatGPT and Claude, and 3.7916 from Flint. “Name a type of car” had ChatGPT and Claude both naming Toyota or Honda; Flint said Ford F-150. A request for a New Balance tagline gave “Run your way” from both mainstream models and “Built to last, run to win” from Flint. On the band-name prompt the article tried, ChatGPT produced 56 names and Gemini 15; recurring tokens across both included “glass,” “neon,” “velvet,” and “static” (Glass Harbor, Static Empire, Neon Hearts, Velvet Echo, Static Horizon). None of those is objectively correct, but the spread is the point. Flint is willing to land somewhere the rest refuse to.
The company is small on purpose. Springboards has two named co-founders, Bingemann as CEO and Kieran Browne as CTO, and is choosing to fine-tune an open-weight base rather than train a foundation model from scratch. Finetuning Qwen 3 keeps the build bill manageable and lets them ship. CTO Browne told the outlet that most users do not realise how similar their chatbot’s answers are to everyone else’s; Springboards wants to make the sameness visible so customers can ask for something different. Early customers include the marketing firm Uncommon, with co-founder Maximilian Weigl quoted in the piece saying “Flint’s programmed to throw an oddball in. It’s more of an invitation to think wider.” Bodacious founder Zoe Scaman, also quoted, reaches for Flint “if I want to catapult myself all over the place” and notes it “sometimes falls over when you start pushing it too far.” Bingemann’s framing is the line of the article: “Most language models are fighting hallucinations. We welcome them.”
What This Means
Two readings of this story are worth taking seriously. The narrow one is that the consumer chatbot market is about to bifurcate again, with one half chasing the convergent “safe middle” and the other pitching idiosyncratic outputs as a feature. The broad one is a Nature World View editorial by Xizhe Zhang of Nanjing Medical University, published in the journal’s June 22, 2026 issue (vol. 654, no. 8120, p. 843). The headline framing alone, “Will AI spark a scientific renaissance - or a diffuse monoculture?”, and the article’s own description, “Artificial intelligence’s ability to enrich science will depend not only on model capability, but also on whether researchers, reviewers and funders reward originality over speed,” make the same point at a different altitude. The convergence problem is not just about chatbot chit-chat; it is about every domain where the same model is asked to do harder and more creative work, from marketing copy to scientific drafting.
The honest answer is that we get both at once. Mainstream assistants are flattening into a single voice at exactly the moment that voice is being asked to do real work. Any startup that can teach a model to zig where the others zag will find paying customers, and so will the open-weight community if it can package “less boring” as a switch a buyer can flip. The harder lesson is that sameness is not an accident of the current products; it is the predictable output of training systems to maximise approval at scale. Fixing it requires new training data, new evaluation benchmarks, and a willingness to spend a bit of coherence-tax to get something new. Flint is one of the first products to make that exchange explicit. Whether the rest of the field follows will tell us a lot about what kind of AI we want.
The Bottom Line
If every assistant sounds like every other assistant, the assistant is no longer doing the part of the job that matters. Springboards is betting, plausibly, that a slice of paying users will pay to escape the sameness. The bet is small; the problem it is aimed at is not.