When you ask Claude, ChatGPT, or Gemini for a random number between 1 and 10, you’ll most likely hear 7. That tiny quirk illustrates the AI groupthink problem that’s been gnawing at developers and creatives alike. The issue isn’t that LLMs can’t generate randomness; it’s that they’ve fallen into a predictable groove that limits brainstorming, travel planning, and any task that thrives on diverse ideas.
Key Takeaways
- Mainstream LLMs often converge on the same answer for open‑ended prompts.
- Springboards, an Australian startup, built Flint to broaden response diversity.
- Flint is trained specifically for open‑ended questions like travel recommendations.
- The approach challenges the default “most likely” sampling that powers today’s chatbots.
- Developers should watch Flint’s early results for clues on avoiding homogenous AI output.
Historical Context
Large language models started as massive pattern‑matching engines that excelled at reproducing the most common phrasing found in their training corpora. Early versions prioritized statistical confidence at every token, a design that made them reliable for answering factual queries and completing code snippets. Over time, that same confidence‑first mindset became the default for most commercial offerings. The trade‑off was a model that rarely strayed from the path of highest probability, even when the prompt invited imagination.
Developers soon noticed that the same “most likely” logic also showed up in creative use cases. When a user asked for a list of ideas, the output tended to cluster around a handful of well‑known concepts. The pattern was subtle at first, but it grew into a recognizable bias: the model would repeatedly surface the same handful of answers, regardless of how many times the question was re‑asked. That pattern is what many now refer to as AI groupthink.
Industry observers began to ask whether the focus on raw accuracy was crowding out the very quality that set generative AI apart—its ability to surprise. The conversation shifted from “how correct is the answer?” to “how many distinct angles does the model explore?” That shift opened a space for startups to experiment with alternative training objectives, and Springboards stepped into that space with Flint.
Flint Tackles the AI Groupthink Problem
Will Douglas Heaven summed it up succinctly:
“Meet the company pushing chatbots away from the obvious.”
Springboards isn’t trying to give you a magic wand; it’s trying to nudge LLMs out of the rut that makes them sound like a chorus of identical voices. By training Flint to purposefully explore less‑trodden paths, the startup hopes to give users a richer palette of suggestions when they ask, “Where should I go in Europe?”
What Groupthink Looks Like in Practice
Ask any of the big‑name models for a random digit, and you’ll hear 7 again and again. That’s not a bug; it’s a feature of the probability‑maximizing decoding strategies most vendors employ. Those strategies favor the most likely token at each step, which is great for code completion or factual recall, but it throttles creativity. When you need a brainstorm, a single‑track answer feels like a dead‑end.
Why It Matters for Creators
Designers, marketers, and product teams rely on LLMs for ideation. If the model keeps serving up the same three‑sentence answer, you’re not getting the out‑of‑the‑box thinking that fuels innovation. That’s why the AI groupthink problem is more than a curiosity; it’s a bottleneck in workflows that depend on divergent thinking.
Springboards’ Approach: Diversity by Design
Springboards’ answer is simple: train a model to value variety as much as accuracy. Flint’s training data includes prompts that explicitly reward novel, unexpected answers. The company says the model “has been trained to come up with a wider variety of responses than mainstream LLMs to open‑ended questions.”
Engineering the Variety Signal
Instead of letting the model default to the highest‑probability token, Springboards injects a loss term that penalizes repetitive outputs. The result is a system that will sometimes suggest a less‑obvious city in Europe, or propose a niche activity that mainstream models tend to ignore. The team hasn’t published exact metrics, but the internal demos they’ve shown suggest a noticeable lift in answer diversity.
Practical Implications for Users
- When you ask Flint for travel ideas, you’ll see suggestions that span popular capitals and off‑beat towns alike.
- For brainstorming sessions, Flint can surface unconventional angles that keep conversations moving.
- Developers integrating Flint via API can toggle a “diversity” knob to control how far the model strays from the mainstream.
Early Signals and Limitations
Flint isn’t a silver bullet. The startup admits that “that won’t work every time—but if it did for you, you may wonder if I have superpowers. I don’t.” In other words, the model still sometimes falls back to the safe answer, especially when the prompt is too narrow. The team also cautions that encouraging too‑much novelty could drift into nonsense, so they’ve built a guardrail that pulls the model back when the output strays beyond a reasonable threshold.
What the Data Shows So Far
The only concrete figure we have is the anecdotal repeat of 7 for random‑number queries on existing LLMs. Flint’s early demos show a broader spread—answers ranging from 1 to 10 appear more often, and travel suggestions aren’t limited to Paris, Rome, or Barcelona. Those results are still preliminary, but they hint at a model that’s less prone to the single‑track habit that plagues its peers.
Potential Risks
Encouraging novelty can amplify bias if the training data isn’t carefully curated. Springboards acknowledges that they’re monitoring for harmful or misleading outputs, and they’ve built a moderation layer that flags anything that looks unsafe. That safety net is essential, because a model that deliberately seeks rare answers could surface fringe viewpoints if left unchecked.
What This Means For You
If you’re a developer building a product that leans on LLMs for ideation, Flint offers a concrete example of how to shift the balance from predictability toward creativity. You can experiment with the same loss‑penalty technique in your own models, or you can integrate Flint’s API (once it’s publicly released) to give end‑users an option for “more diverse” answers.
For founders, the story is a reminder that the market isn’t satisfied with just accurate text generation; there’s a niche for tools that help teams think differently. Springboards is betting on that niche, and if they can prove a measurable boost in brainstorming efficiency, they could carve out a valuable slot in the crowded LLM ecosystem.
Concrete Scenarios
- A marketing team drafts campaign taglines. Using Flint, the copywriter receives five distinct phrasing options instead of three variations that all echo the same theme. The broader set of options shortens the decision loop and uncovers a tagline that resonates better with a target demographic.
- A travel‑booking platform wants to suggest itineraries that go beyond the usual tourist hubs. By feeding user preferences into Flint, the service can surface hidden gems—think a coastal village in Croatia or a mountain hamlet in Slovenia—thereby differentiating its recommendations from competitors that stick to the usual suspects.
- A product manager runs a remote ideation workshop. With Flint as a virtual participant, the group receives prompts that challenge conventional assumptions, such as “What would a user in a low‑bandwidth environment need?” The model’s willingness to jump off the most‑likely path keeps the discussion lively and prevents premature convergence.
Each scenario shows how a modest shift toward diversity can ripple through a workflow, turning a single‑track output into a catalyst for richer outcomes.
Looking Ahead
The next step will be real‑world testing. Will developers adopt Flint for internal hackathons? Will travel platforms plug it into recommendation engines? Those questions will shape whether the AI groupthink problem stays a curiosity or becomes a solved engineering challenge. One thing’s for sure: as long as LLMs keep defaulting to the same answer, there’ll be room for startups like Springboards to shake things up.
Key Questions Remaining
- How will the “diversity” knob balance novelty against relevance in production environments? Too much variance could overwhelm users, while too little might revert to the old homogenous pattern.
- What metrics will the industry adopt to quantify answer variety without sacrificing accuracy? Existing benchmarks focus on correctness, leaving a gap for evaluating creativity.
- Can the moderation layer keep pace with the model’s push toward rarer outputs, especially as the system encounters niche topics that haven’t been extensively vetted?
- Will other startups adopt the same loss‑penalty approach, and if so, will a new ecosystem of “creative‑first” LLMs emerge, or will the majority stay anchored to probability‑maximizing decoding?
Sources: MIT Tech Review, original report
