Why Asking the Same Question Limits Solutions

Why do almost all AI models keep giving similar answers on creative tasks or writing? I have been noticing this for a while. Whenever I ask for something creative or exploratory, Claude, GPT-5 and Gemini all converge on a similar answer template. The wording shifts a little, but the underlying idea feels identical. I just finished reading Artificial Hivemind, a new paper out of UW and the Allen Institute, and it confirms this is not just a vibe. To prove this, the authors collected 26,000 real world open-ended queries, ran them across 70+ models, and used the results to study how reward models behave when there is no single correct answer. The usual expectation would be that the responses scatter across embedding space. Instead, almost every model collapses into the same tight cluster. Once we see how little variation there is in that answer space, it becomes obvious why they all sound alike even when the prompts are creative. The examples in the paper make this very noticeable. Ask dozens of models for a metaphor about time and nearly every single one, regardless of architecture or lab, converges on “Time is a river” or “Time is a Weaver”. This has real implications for the current interest in model swarms. The idea behind multi model swarms where multiple LLMs debate, self criticise and jointly converge on a better answer. That only works if the models bring genuinely different perspectives. If every model in the swarm is pre-conditioned to seek the exact same local minima of “safety” and “helpfulness”, then it’s just a single model with higher latency. The bottleneck seems to be the Reward Models. These are supposed to help models align with human judgement, but in practice they are actually penalising idiosyncratic creativity. They flatten the natural diversity in human preference and push models toward one homogenized style of answer. If billions of people start using these tools for ideation and every tool keeps suggesting the same safe and statistically favoured ideas, we risk shrinking the collective search space of thought. My takeaways from the paper: 1. We should treat disagreement as signal rather than noise. We are currently training models to collapse distributions into a single best answer. 2. We should be training them to represent the full distribution of human preference. 3. Diversity needs to be an explicit evaluation metric. If an alignment method increases average quality but reduces the variety of answers, that is a regression. 4. For agents that help with research, design and ML work, we should build systems with intentional heterogeneity. This paper is worth reading carefully for anyone working with language models, especially the sections on inter model similarity and reward model calibration. This paper makes a strong point to show that the immediate risk of AI is a future where every model is intelligent in exactly the same way.

Imagine you hired three summer interns today. One is a top computer science grad from Stanford University. Another studied data science at Harvard University. The third is an engineer from MIT Sloan School of Management. They’re all genuinely smart, competitive, and motivated. They’re young, single, and perfectly happy working long hours without worrying too much about burnout or balance. The upside is obvious: raw horsepower. These are people who can absorb complexity fast, learn new domains quickly, and move with speed. The downside is also obvious if you’ve ever worked with interns, or frankly, with very smart people early in their careers. They make assumptions. They miss context. They confidently walk past edge cases. Not because they’re careless, but because they haven’t yet built the scar tissue that comes from seeing things break in the real world. So the mistake would be to give all three the exact same task and hope the “best” answer wins. What you actually do is assign them adjacent work and force overlap in review. You ask one to do the primary analysis, another to validate the logic and assumptions at a high level, and the third to sanity-check outcomes against real-world constraints. Not line-by-line edits, but conceptual scrutiny. That’s how they cross-train. That’s how blind spots surface early. That’s how quality goes up without slowing things down. That’s also how I think about using Anthropic’s Claude, OpenAI’s ChatGPT, and Google’s Gemini. They’re all extremely capable. They’re all fast. And they all make different kinds of mistakes. If you ask the same question to all three and blindly accept the first confident answer, you’re treating them like a search engine. That’s a waste. The leverage comes from orchestration. One model drafts. Another critiques the structure and assumptions. The third checks for gaps, hallucinations, or missing constraints. You’re not choosing “the smartest model.” You’re designing a system where smart models keep each other honest. That’s not prompt engineering. That’s management.

You answered the question in the all-hands. You covered it in the leadership team meeting. You put it in the deck. And this week, they're asking again. "What's the priority?" "Are we still doing this?" "Did that decision change?" Your first instinct: they're not listening. Or you're not communicating clearly enough. 👉🏾 But here's what's actually happening: If your team keeps asking for clarity you've already given, the issue isn't communication, it's trust in steadiness. Repetition is often a symptom of leaked uncertainty. They're not asking because they forgot. They're asking because they've learned that what you said last week might not hold this week. Research shows that frequent shifts in leaders' priorities are strongly associated with burnout and disengagement. When leaders change direction without clear rationale, 💥 teams protect themselves, 💥 they do the minimum, 💥 wait out the next shift, 💥 discount current directives as temporary. Studies analyzing over 80,000 360-degree reviews found that leaders who act inconsistently erode trust roughly three times faster than those who behave predictably. So when your team keeps asking the same question, → They're not being dense. → They're being rational. → They're checking whether this version of the answer is the one that will actually hold.  🚩Here's where uncertainty leaks: → Tone and body language. You say "we're committed," but your tone is tight, your body language closed. They feel the hedge. → Decision reversals without narrative. You change course without explaining what changed or what you learned. It registers as unreliability. → Hedging language. "Maybe," "sort of," "we'll see." Softening feels collaborative but overuse reduces perceived authority. → Say/reward gap. You say innovation matters but penalize failed experiments. They watch what you reinforce, not what you announce. People track coherence across your words, tone, and actions at a fine-grained level. When those don't align, they adjust. Not because they're bad listeners. Because you've trained them that clarity isn't stable. ✅ Here's the shift: 1️⃣ When you give clarity, ask yourself, "Do I actually believe this will hold?" If not, don't give false certainty. Say: "Here's where we are. Here's what's still forming." 2️⃣ When you change course, name it explicitly: "Last week I said X. Here's what changed. Now we're doing Y." 3️⃣ Check your tone. If your words say "we're locked in" but your body says "I'm not sure," they'll trust the body. 👉🏾 Steadiness isn't rigidity. It's coherence, your words, tone, and decisions pointing in the same direction. If your team keeps asking for clarity you've already given, look for where your uncertainty is leaking. Because repetition isn't about them not hearing you. It's about them not trusting what they heard will still be true next week. Check out the REWIRED. newsletter (link in comments).

Here’s my guilty pleasure. Something i’ve been working on for about a year already. Happily diving into every little aspect of it. Passionately asking the next questions. > The decision-making process under uncertainty. < I geek out on this so much! Now, time to start sharing the gems 👀 Every decision starts with how we define the problem, right? But what if that definition ultimately limits the solutions we’re able to come up with? Here's what we’ve been missing about decision-making. A decision is a choice based on how subjectively beneficial we deem each option. In reference to how we defined and understood the problem. Read that again. The challenge: Your brain automatically frames problems based on YOUR experience, YOUR unconscious biases and blindspots, looking through the lens of YOUR expertise. Ask me about a challenge in the organization and my first 10 thoughts are about people ;) Here's what happens: → You see: "We need to improve efficiency and cut costs." → Your operations manager sees: "We need better systems." → Your team lead sees: "We need more resources." → Your newest hire sees: "We need clearer processes." → Your HR person thinks: “We need more engaged people.” → Same meeting. Same data. Four different problems. Four different "right" solutions. Your brain filters information through existing patterns. Edits out data that doesn't fit your frame. You're not making decisions. You're making assumptions about what the decision should be. Before asking "What should we do?" Ask "What problem are we actually solving?" Before presenting solutions... share your problem definition. Before deciding.. check if everyone's solving the same thing. The most expensive decisions aren't wrong ones. They're right decisions to wrong problems. Next time you're in a decision-making meeting: Start with "Here's how I understand the problem." Then PAUSE. Give people space to think. Maybe you're not even in the same conversation. What's a recent decision where the problem definition would have changed everything? #DecisionMaking #CreativeBrainMethod #LeadershipDevelopment #CommunicationSkills #CognitiveBias

The first question positions you as someone managing a problem. Containing something undesirable – a posture of trying to control/limit something. The second positions you as someone building something; a mental model of expanding possibility.  Same topic yet, completely different solution space. I see this everywhere in corporate life. The question an organisation says it's asking, and the question it's actually asking, are often not the same. And the gap between the two is where strategy kind of dies a slow death. A bank says: "How do we drive financial inclusion?" But measures market share. Financial inclusion means making the market bigger than it currently is - reaching people the market has never served. Market share means competing for the people already in it. These are not the same question dressed differently. They produce fundamentally different behaviours, investments, and outcomes. The question you frame is the question you answer. And the question you actually answer is revealed not by what you say, but by what you measure, what you incentivise, and what you celebrate. So here's what I'd ask any leadership team: what is the question written on your walls - and what is the question running in your veins? Because only one of those is shaping your organisation. #SystemsLeadership #SystemsThinking #Strategy