AI's Groupthink Problem: A Startup's Fix and What It Means for Solo Builders

I’ve been using AI tools daily for over a year now — testing them, building with them, writing about them. And I’ve noticed a pattern that’s becoming harder to ignore: the answers are getting more alike. When I ask the same question to ChatGPT, Claude, Gemini, and a handful of smaller models, the responses are converging. Not toward “the best answer” — toward the same answer, with the same framing, the same hedges, and the same blind spots.

This is AI’s groupthink problem. And a startup I came across recently is trying to fix it in a way that actually matters for people like us — solo builders, content creators, and small business owners who depend on AI for real work.

We’ve covered AI tool overwhelm before — the challenge of choosing between too many tools. But this is different. When all the tools give you the same output, the choice becomes meaningless. And the downstream consequences affect everyone using AI for decision-making, research, and content.

What’s actually happening

AI models converge for structural reasons. During training, they’re optimized on similar datasets (much of the internet), using similar architectures (transformer-based), with similar objectives (predict the next token). The result is what researchers call “model homogenization” — as models get bigger and better, they don’t get more diverse. They get more average.

This matters in practice. When I ask AI to help me evaluate a business decision, draft a blog post on a specific topic, or find a unique angle on a news story, I’m relying on the model to see something I didn’t. If every model sees the same thing — the same mainstream framing, the same safe take, the same hedged conclusion — the diversity of perspective that makes AI valuable collapses into a single, averaged-out voice.

A recent MIT Technology Review article covered a startup working on this problem directly. The approach: instead of trying to make one model that “thinks differently,” they’re building a system that routes queries to specialized models and aggregates divergent outputs. The goal isn’t consensus — it’s structured disagreement.

Why this matters for solo builders specifically

If you’re a solo builder using AI for client work, content creation, or product development, AI homogenization creates a specific problem: your output starts looking like everyone else’s output.

This is already happening. If you’ve noticed that AI-generated blog posts, social media captions, and business strategies all have the same rhythm — the same listicle structure, the same “it depends” hedging, the same conclusion that somehow says nothing — you’re experiencing the groupthink problem at the consumer level.

We saw this when we covered AI’s groupthink problem for solo builders — the practical impact is that using the same tools as everyone else, trained on the same data, produces the same outputs as everyone else. Competitive differentiation erodes when differentiation was supposed to be AI’s promise.

For content creators, this is especially acute. If you’re using AI to draft blog posts — which, let’s be honest, is something we do — the risk is that the AI gives you the same angle, the same sources, and the same structure as every other creator who asked the same question. Google’s algorithms are increasingly capable of detecting AI-generated content that says nothing new. The bar for “valuable content” is rising, and AI convergence is pulling your content toward the average.

What you can do about it

1. Use multiple models for the same query. This is the manual version of what the startup is automating. When I’m evaluating a decision or exploring an angle, I ask the same question to at least two models and look for where they disagree. The disagreement is usually where the interesting information lives.

2. Prompt for contrarian views. Instead of asking “what should I do about X?”, ask “what would someone who disagrees with the conventional wisdom say about X?” or “what’s the strongest argument against the standard approach to X?” This nudge pushes models out of their average-response territory.

3. Feed models your own data. The more context you provide — your voice, your past work, your specific situation — the less the model defaults to its trained average. This is why RAG (retrieval-augmented generation) consistently produces better outputs than generic prompting. The model isn’t just drawing on its training data; it’s drawing on your specific context.

4. Build in human review. The final filter is still you. AI can draft, research, and organize — but the judgment about whether an angle is actually original, whether a take is genuinely interesting, whether a recommendation fits your specific context still requires human evaluation. That’s what we covered in the privacy problem nobody talks about — the uncritical reliance on AI output without human verification.

5. Stay curious about smaller models. The homogenization problem is worst among the largest models, which are trained on the broadest datasets and optimized for the most “average” correct answer. Smaller, specialized models sometimes produce more interesting and divergent outputs because they haven’t been smoothed into the mean. Check our AI Tool Advisor for model-by-model comparisons that highlight these differences.

What to stop doing

Asking AI to do your thinking for you. AI is a tool — a powerful one — but it’s a synthesizer, not an originator. If you prompt it for “the best strategy for X,” it gives you the average strategy that most people would consider reasonable. That’s not the same as the best strategy for you.

Treating AI outputs as research. When AI gives you “5 studies that show X,” verify that those studies exist, that they say what the model claims, and that they’re from credible sources. Model hallucination isn’t the only risk — model citation of real studies that say something different from what’s claimed is more common and harder to catch.

Accepting the first answer. If the first response you get from an AI tool feels generic, it probably is. Push back. Ask for alternatives. Ask for the angle that most people wouldn’t think of. The model’s second or third response to a redirected prompt is almost always more interesting than its first.

The tool question

No specific tools here — the point is strategic, not tactical. But the principle applies to every tool you use:

Use AI for breadth, not depth. Let it surface information you wouldn’t have found. Evaluate and curate that information yourself.

Build workflows that include disagreement. If you use AI for decision-making, build a step where you deliberately seek the opposing view. The startup covered in the MIT Technology Review is essentially automating this step.

Keep your voice. The most valuable thing you have as a creator or builder is your perspective. AI can amplify it, but it can’t replace it. If your AI-assisted content sounds like everyone else’s AI-assisted content, the tool is working against you.

What we still don’t know

Whether the startup’s approach — routing queries to specialized models and aggregating divergent outputs — will actually produce better decisions or just more noise is unproven. Diversity of viewpoint is valuable when the viewpoints are informed. Diversity for its own sake is just chaos.

The larger question is whether AI convergence is a temporary artifact of current training methods or a structural feature of how large language models evolve. If it’s structural — if bigger models inevitably converge on more average outputs — then the tools we’re building today may become less useful over time, not more.

What’s clear: the problem is real, it’s already affecting the quality of AI-assisted work, and the solutions need to be intentional. Relying on one model, accepting its first answer, and publishing the result is a path to mediocrity. The solo builders who thrive will be the ones who use AI as a starting point — not a final answer. If you’re building your first AI workflow, start here — and remember that the thinking part is still yours.

If you’ve noticed your AI outputs feeling increasingly “same-y,” share this with someone building with the same tools.