LLMs Default to the Same Answers. Springboards Built Flint to Fix That.

Most large language models, including ChatGPT, Claude, and Gemini, tend to cluster around the same predictable outputs. Ask any of them for a random number between 1 and 10 and you will almost certainly get 7. Australian startup Springboards has identified this as a real limitation and built an LLM called Flint, trained specifically to produce a wider range of responses. The goal: make AI more useful for open-ended tasks like brainstorming and creative planning, where repetitive, consensus-driven answers are a liability.

What happened

MIT Technology Review surfaced a concrete, testable flaw in today’s leading AI chatbots: they are far more predictable than most users realise. Ask ChatGPT, Claude, or Gemini for a random number between 1 and 10, and you will almost certainly receive 7. Ask again and you will typically get 3 or 4, then 8 or 9. The sequence is nearly scripted.

This is not a quirk. It reflects how large language models are trained. They learn to produce the most statistically likely next token, which means they naturally converge on popular, safe, expected outputs. The result is a form of machine groupthink baked into the architecture.

Australian startup Springboards has built an LLM called Flint to address this directly. According to the MIT Technology Review report, Flint has been trained to produce a wider variety of responses, making it better suited to tasks where creative range actually matters.

Why it matters

For a lot of business use cases, predictability is fine. If you are asking an LLM to check syntax, summarise a document, or pull out key data points, you want a consistent, accurate answer. Groupthink is not a problem there.

The trouble starts when you use AI for ideation. Brainstorming product names, drafting campaign concepts, generating vacation itineraries, exploring strategic options: these are tasks where the value of AI comes from surfacing something you had not already thought of. If the model keeps circling the same cluster of ideas, it is not expanding your thinking. It is just reflecting it back at you faster.

This is a practical issue for any business owner who has used ChatGPT for creative work and come away feeling like every suggestion sounded the same. The sameness is not your imagination.

The use cases most affected

• Marketing brainstorming and campaign ideation
• Product naming and tagline generation
• Content strategy and angle exploration
• Strategic planning and scenario generation
• Travel or event planning where variety is the point

Our take

The random number demo is a genuinely good illustration of a real problem. We have noticed this in client work: when you use any major chatbot to brainstorm, the first five outputs feel fresh, and then everything starts sounding like the same consultant wrote it. The model has a comfort zone and it stays there.

That said, we would be cautious about assuming Flint has solved this cleanly. Training a model to be “more varied” is easy to claim and harder to verify. More variation could mean more useful creative range. It could also mean more confident nonsense. The difference matters a lot depending on the task.

What Springboards is pointing at is real: the dominant LLMs are optimised for accuracy and coherence, not for creative breadth. A model built with a different objective could genuinely perform better for ideation tasks. Whether Flint delivers that in practice will need testing across real workflows, not just demo prompts.

For now, the more useful habit is knowing when to push back against groupthink yourself. Techniques like explicitly asking for “10 ideas that are as different from each other as possible,” or asking the model to argue against its own first suggestion, can shift outputs meaningfully even in standard LLMs. Flint may eventually make that unnecessary, but the workarounds work today.

What to do about it

If you use AI for brainstorming or creative planning, try this test: ask your current chatbot for a random number between 1 and 10, three times in a row. If you get 7, then 3 or 4, then 8 or 9, you are seeing the predictability problem firsthand. From there:

1. For creative tasks, explicitly instruct the model to maximise variety across its suggestions, not just quality.
2. Ask for a second round of ideas that deliberately avoids everything in the first round.
3. Watch Springboards and Flint as independent reviews come in. If output diversity holds up in real use cases, it is worth adding to your AI toolkit for ideation work specifically.

The bottom line: the best AI tool for writing code and the best AI tool for brainstorming a campaign may not be the same model.

Source: MIT Technology Review