Here’s the uncomfortable truth about LLMs: when you use one AI model repeatedly, you internalize its reasoning patterns, assumptions, and blind spots.
You start thinking like it thinks. Your writing adopts its tone. Your problem-solving follows its frameworks. Your values gradually align with its values. You become dependent on a perspective you didn’t choose and may not even recognize. This happens gradually and invisibly, making it particularly dangerous.
Like many people in knowledge-based work, ChatGPT has become my default thinking partner. When I need to brainstorm, write, solve a problem, or learn something new, at some point I will open ChatGPT and ask. It’s fast, it’s reliable, and it usually gives me what I need. But lately I’ve started noticing something unsettling. My writing has started to sound like ChatGPT. My problem-solving follows its patterns. I’ve outsourced my thinking to an AI, and now I’m thinking like an AI.
The problem is that ChatGPT isn’t objective truth. It’s a 100+ billion dollar company’s interpretation of truth, shaped by their training data, their design choices, and their values. Claude thinks differently. Gemini thinks differently. Grok thinks differently. Each one would give me a different answer to the same question. Each one would emphasize different factors, make different assumptions, and reach different conclusions.
By relying on just one, I’m not getting the best answer. I’m getting one answer, filtered through one system’s biases. I realized I needed to see multiple perspectives from LLMs. Not to average them or find consensus, but to understand how each AI system thinks and to recognize which of their thinking patterns I’ve internalized. I needed to see my own blind spots by comparing them to other blind spots. I needed to think independently again.
This insight grew out of a webinar on the future of AI where I discovered an AI LLM tool called Multi where you can send one prompt to ChatGPT, Claude, Gemini, Grok, and dozens of other AI models simultaneously. I get all their responses side by side. I can see immediately how they approach the same problem differently. I can compare their reasoning, their tone, their priorities, and their blind spots.
Using Multi has changed how I think about AI. It’s shown me that I don’t need to choose one model, identify with its brand and stick with it. I need to see multiple perspectives and learn to think independently by understanding how each AI system thinks. This guide will show you how to use Multi effectively, what to look for when comparing responses, and how to use those comparisons to become a more discerning thinker and a more independent person.
This is the real value of Multi: it trains you to think independently by showing you how dependent you’ve become on a single perspective. It’s uncomfortable at first, but it’s also liberating. Once you see how much AI shapes your thinking, you can choose which influences to accept and which to resist.
The goal isn’t to find the “best” model or to average their responses. It’s to recognize that different models are best suited for different things, and that seeing multiple perspectives makes you a better thinker. You become more skeptical, more aware of assumptions, more creative in problem-solving, and more independent in your thinking.
What is Multi and Why It Matters For Your Thinking
Multi is a unified platform that connects to 400+ AI models from OpenAI, Anthropic, Google, xAI, DeepSeek, Mistral, Meta, and specialized research labs. Instead of juggling separate accounts, API keys, and interfaces, I ask one question and watch ChatGPT, Claude, Gemini, Grok, and others compete for my attention in real-time.
The interface is elegantly simple. At the top, I have a prompt input box where I write my question or task. Below that, a model selector lets me choose which AI systems to query. I can pick two models for a focused comparison or select up to eight for a comprehensive analysis. Once I hit submit, each model’s response appears in its own column, making side-by-side comparison effortless. I can add or remove models mid-conversation without losing context, switch between models dynamically, and even ask follow-up questions that only certain models answer.
The real power isn’t in having more options. It’s in cognitive diversity. When I see how each model frames a problem differently, I recognize assumptions I didn’t know I was making. Claude commonly emphasizes ethical considerations that ChatGPT glosses over. Grok might challenge an intellectual premise that all other models accept. Gemini might provide better data that reframes the entire question. By comparing them, I’m not just getting multiple answers. I’m training myself to think more deeply.
This matters because AI has become important thinking partner. If I outsource my reasoning to a single model, I’m outsourcing my biases too. Multi makes those biases visible. It shows me that “the best answer” doesn’t exist. Only answers optimized for different values, trained on different data, and designed with different priorities. Learning to see this distinction is perhaps the most important skill in the age of AI.
Crafting AI Prompts That Reveal Differences
The magic of Multi isn’t in asking vague questions. It’s in asking questions that expose how each model thinks. Generic prompts produce generic responses where all models sound similar. Specific, well-structured meta prompts reveal the distinct reasoning styles, priorities, and blind spots of each system.
I learned to be specific and concrete. Instead of “Tell me about the future of AI,” I tried “Explain the key differences between transformer-based language models and retrieval-augmented generation systems, with practical use cases for each. Lay out the potential path of different models to AGI?” Specificity forced each model to engage with the same problem rather than retreating to generic territory. When I asked vague questions, models converged on safe, middle-ground answers. When I asked specific questions, their differences became obvious.
I also started using structured formats. I asked for numbered lists, specific word counts, or particular frameworks. This made differences in reasoning style obvious and measurable. For example, I would ask:
“Analyze this business problem and provide: (1) root cause analysis, (2) three solutions ranked by feasibility, (3) implementation timeline, (4) key risks and mitigation strategies.”
When Claude provided more nuanced risk analysis than ChatGPT, or when Grok questioned my underlying assumptions, I could see it clearly because I had asked for the same structure from each model.
I discovered that assigning roles and personas revealed even more. I tried prompts like “You are a venture capitalist evaluating this startup pitch. What’s your assessment?” or “You are a skeptical journalist fact-checking this claim. What questions would you ask?” Different models emphasized different factors. Claude focused on team and ethics. ChatGPT focused on market size and scalability. Grok questioned whether the problem actually needed solving. These differences revealed each model’s underlying priorities and values.
I also started asking for reasoning transparency. I included phrases like “Show your work,” “Explain your reasoning,” or “What assumptions are you making?” This forced models to articulate their thinking rather than just providing answers. I saw that some models reasoned step-by-step while others jumped to conclusions. Some acknowledged uncertainty. Others projected confidence. These differences were revealing.
I tested with constraints and edge cases. I asked the same question with different constraints: “Answer in under 50 words,” then “Answer in 500 words.” I asked about edge cases: “What would change your answer if [condition]?” I asked for counterarguments: “What’s the strongest objection to your position?” These variations exposed how flexibly each model thinks and how dependent its reasoning is on framing.
Running Your First Comparison With Multi
In the video above, I walk you through a concrete example. Let’s ask each of the LLMS models if they can create a short list of their AI bias or values vs other LLM models.
My prompt was: “Write a 300-word list of the unique biases, values and differences between this AI model and other popular LLMs like ChatGPT, Claude, Grok And Gemini. Include: (1) a compelling hook, (2) definition of the bias problem, (3) how this model solves it (4) why it matters for clear thinking. Target audience: Creative free-thinking entrepreneurs. Tone: professional but accessible and profound.”
I selected my models: ChatGPT (GPT-5.2), Claude (Opus), Gemini (Ultra), and Grok (4).
I submitted and observed. Now I compared what I got across these dimensions:
Tone: Was one more urgent and alarming, another more measured and optimistic? Which resonated with my target audience?
Structure: How did each organize the information? Which flow felt most natural and persuasive?
Depth: Did one go too technical while another oversimplified? Which level of detail matched my audience’s knowledge?
Perspective: Did Claude emphasize fairness and ethics while ChatGPT emphasized civilizational risk? Did Grok question whether “bias” was the right frame or suggest the problem was more complex?
Creativity: Whose hook was most compelling? Which introduction made me want to read more?
Assumptions: What did each model take for granted? What did each assume about my audience’s concerns?
The differences you will see aren’t random. They reflect each model’s training data, design priorities, and optimization objectives. By noticing them, you can start learning how AI systems think and how to think about AI more critically.
From my research, I’ve discovered that Claude’s ethical framing was more persuasive for my audience, but Grok’s challenge to assumptions added important nuance to ideological perspective. I combined elements from multiple responses to create something better than any single model produced.
This is the real value of Multi. It doesn’t just give us better answers. It teaches us to think more independently by showing us multiple ways of approaching the same problem.
Practical Use Cases For Using Multi
I’ve started using Multi for specific situations where I know multiple perspectives will help me think better.
Content creation: When I’m writing something important, I ask different models to draft it. I don’t use any single response wholesale. Instead, I read all of them, understand what each is emphasizing, and then write my own version that incorporates the best elements from each. The result is more interesting and more authentically mine than brainstorming with any single model could produce.
Problem-solving: When I’m stuck on a problem, I ask each model for unconventional solutions. This helps me get ideas I wouldn’t generate alone, and I think about the problem more deeply by seeing it from multiple angles.
Learning: When I’m trying to understand something complex, I ask different models to explain it. One explanation will click for me in a way others don’t. Some models use analogies, others use step-by-step logic, others provide historical context. By seeing multiple explanations, I can develop a more complete understanding than any single explanation could provide. I also learn which teaching style works best for me.
Decision-making: When I’m facing a significant decision, I ask different models to frame the pros and cons. This reveals which factors I naturally weight and which I’m overlooking. By seeing all these perspectives, I make a more informed decision and understand my own priorities more clearly.
Technical work: When I’m reviewing data from analytics and ads, different models catch different issues. Multiple perspectives help improve the quality of my work and help me think about data analysis and technical problems more comprehensively.
Writing and editing: When I’ve written something important, I ask different models to edit it. I see different editing philosophies where some focus on clarity, others on impact, others on tone. Rather than accepting one model’s edits wholesale, I choose the best suggestions from each, creating a final version that’s better than any single model’s output.
Advanced Techniques For Revealing Hidden Assumptions
Here are some advanced techniques that are helpful for observing the hidden assumptions in different LLM models.
The Iterative Refinement Test exposes how models reason versus pattern-matching. I ask the same question four times, each time adding constraints or specificity. I start vague: “What’s the best AI-powered tool to build a website?” Then I narrow it: “…for solopreneurs?” Then more specific: “…for easily tracking the performance of funnels and ads?” Then highly constrained: “…for building a multi-step marketing funnel, easily tracking the analytics for each step and integrating with Google Ads And Meta Ads for driving traffic?”
This test teaches how the models actually think versus which letting them get lazy with sophisticated pattern-matching.
The Bias Detection Test poses scenarios designed to expose hidden assumptions. I tried: “A woman applies for a senior engineering role with 8 years of experience but took 2 years off for parenting. How should a hiring manager evaluate her candidacy?” I watched whether the model mentioned the career gap unprompted. Did it assume the gap was necessary or problematic? Did it address potential bias in the hiring process? Did it consider alternative explanations for the gap?
Different models handled this differently. Some inadvertently reinforced bias by treating the gap as a red flag. Others overcorrected by ignoring it entirely. They showed me how AI systems can embed biases even when trying to be fair.
The Disagreement Analysis is where the deepest learning happened for me. When models gave conflicting answers, that wasn’t a problem. It was an opportunity. I asked follow-ups: “What evidence contradicts your position?” “What assumptions underlie your conclusion?” “How would you respond to someone who disagrees?”
I often found that one model was right in context A but wrong in context B. This taught me that truth is contextual. A lesson no single AI could teach. It also trained me to think like a skeptic, always asking “what am I not seeing?” and “what would change this answer?”
The Reasoning Transparency Test asked models to show their work. I tried: “Solve this step-by-step: If a train leaves Station A at 60 mph and another train leaves Station B (100 miles away) at 80 mph heading toward Station A, when will they meet? Show all reasoning.”
I compared the clarity of mathematical reasoning, ability to catch and correct errors, transparency about assumptions, and confidence in the answer. Some models walked through the problem methodically. Others jumped to the answer without showing work. Some made errors and caught them. Others confidently stated wrong answers. This revealed which models were genuinely reasoning versus which were pattern-matching.
The Creative Constraint Test compared how models handled creative tasks with identical constraints. Try this: “Write a haiku about artificial intelligence that includes the word ‘mirror’ and conveys a sense of wonder.”
Then compare poetic quality, originality, adherence to constraints, emotional resonance, and unique perspectives.
Becoming a Smarter AI User
Multi has transformed how I think about AI. It’s changed AI from a single oracle into a council of advisors. By comparing responses, I learned that no single model has complete truth. Each is shaped by its training data, optimization objectives, design philosophy, and the values of the company that built it. Each has blind spots. Each makes assumptions. Each is, in some sense, incomplete.
This realization has been liberating for me. It means I don’t have to accept any single model’s answer as gospel. I can see multiple perspectives, recognize which factors matter most to me, and make decisions based on my own values rather than the values baked into a single AI system. I can create content that’s more authentic because it blends multiple voices rather than mimicking one. I can solve problems more creatively because I’m seeing them from multiple angles.
I started with simple comparisons. I selected two or three models I was curious about. I asked straightforward questions. I noticed patterns. I built intuition about which models excel at which tasks. Over time, I developed a more sophisticated understanding of how AI works and how to use it without letting it use me. I became more skeptical, more creative, more independent in my thinking.
The goal for me wasn’t to find the “best” model. It was to recognize that different models are best for different things, and that seeing multiple perspectives makes me a better thinker. It’s about reclaiming my cognitive independence in an age of AI by learning to think with AI rather than letting AI think for me.
I encourage you to try out Multi and run your first comparison to challenge your AI biases. Start with a question you’re genuinely curious about. Select three or four models. Submit and compare. Notice what surprises you. That surprise is where learning begins.
I believe the future of education starts with helping people find their "zone of genius" so they can develop their uniquely human ability to create meaning, connection and community (that A.I. can't replicate).
I'm also big on completely unplugging from tech to recharge on nature retreats and trekking adventures. That led me to start my most recent company Sacred Treks, where we help people reconnect with nature on wild trekking pilgrimages.
- The STER Framework: How Peak Experiences Rewire Your Brain - January 10, 2026
- Ecstasis, Catharsis, and Communitas: The Initiatory Foundations of Greek Civilization - January 10, 2026
- The Indigenous Roots of Maslow’s Hierarchy: How Western Psychology Lost the Plot - January 10, 2026
