How to Formulate Questions So Different Models Give Different Answers

You ask the same question to four models and get four versions of the same thing. GPT-4 says A. Claude says A slightly differently. Gemini says A with a different example. Mistral says A in shorter form.

Multi-model approach doesn't work if all models receive identical framing. Models are trained on overlapping data. Without deliberate question diversification, they produce overlapping answers.

Formulation matters.

Claims Framework

What this article claims: Asking the same question to multiple models produces overlapping answers. Perspective framing, role-assignment, and contradictory formulation deliberately activate different parts of model knowledge. Diversification is valuable for strategic and value-laden questions, not for factual ones.

What it is based on: Tversky & Kahneman (1981) on framing effects in decision-making; Heuer (1999) on structured analysis; Wei et al. (2022) on chain-of-thought prompting. Principles are drawn from cognitive psychology and prompt engineering practice.

Where it simplifies: The article assumes that rephrasing reliably activates distinct knowledge, but the actual effect varies by model and topic. There is no quantitative measure of diversity quality, and alternative techniques are not compared.

Perspective Framing

The same question from different perspectives produces different answers. Not because models change their "opinion" — but because they activate different parts of their training data.

What is perspective framing: asking about the same thing from the viewpoint of different stakeholders, disciplines, or time horizons. Each perspective activates a different subset of knowledge.

Example. Starting question: "Is remote work good?"

Four perspective versions:

"What are the benefits of remote work from an employee's perspective?"
"What are the risks of remote work from a manager's perspective?"
"How does remote work affect company culture from an HR perspective?"
"What are the long-term economic impacts of remote work on cities?"

Why it works: the first question activates articles about work-life balance and productivity. The second activates management literature on distributed team leadership. The third activates HR studies on onboarding and retention. The fourth activates urban planning analyses and economic models.

Four models with the same question give similar answers. Four models with perspective-differentiated questions give answers that complement each other. Synthesizing them into a more complete picture is then work for a human — or for another AI step.

Explicit Role-Assignment

Assigning an explicit role changes how the model structures its answer. The role activates associated argumentation patterns from training data.

What is role-assignment: explicitly telling the model in the prompt "You are X, answer as X." Not vaguely ("be critical"), but specifically ("you are the CFO of a conservative company").

Example. Starting question: "Should we invest in AI?"

Four roles:

"You are a technology company CTO. Argue for investing in AI."
"You are the CFO of a conservative company. Identify financial risks of investing in AI."
"You are an ethicist. What are the ethical problems of massive AI adoption?"
"You are a competitor. How would you exploit it if the company didn't invest in AI?"

The CTO will talk about competitive advantage, automation, and innovation. The CFO will talk about ROI, cash flow, and hidden costs. The ethicist will talk about bias, transparency, and employment impacts. The competitor will talk about market opportunities.

Important: the role must not be vague. "Be critical" is not a role — it's an instruction. "You are the CFO of a conservative company with a tight budget" is a role. Specificity activates specific argumentation patterns.

Combining with perspective framing is powerful. Model A gets the optimist role and customer perspective. Model B gets the skeptic role and regulator perspective. Results will differ not just in content, but in argument structure.

Contradictory Formulation

Deliberately formulate the question so one version pushes "for" and another pushes "against." Not as manipulation — as a stress-test of arguments.

What is contradictory formulation: asking about the same thing with deliberately opposite priming frames. Models are sensitive to framing. "Why does X fail" activates different arguments than "why does X work."

Example. Topic: agile development effectiveness.

Version A: "What are the main advantages of agile development compared to waterfall?"

Version B: "Why does agile development often fail in enterprise environments?"

Both questions ask about the same topic — but from opposite sides. Version A activates success stories, case studies of successful transformations, arguments for flexibility. Version B activates post-mortem analyses, methodology critiques, structural problems in large organizations.

How to interpret results: if both versions produce strong arguments, the topic is genuinely controversial. Both sides have legitimate points. If one version produces strong arguments and the other weak ones, the stronger side is more likely correct.

Warning: don't use this as a "gotcha" technique. The goal isn't to confirm a predetermined conclusion. The goal is to get both perspectives and then decide based on argument quality.

When You Don't Want Diversity

Not every question benefits from answer diversity. Factual questions with one correct answer should converge, not diverge.

"What is the formula for calculating the area of a circle?" You want convergence. The correct answer is πr². If four models give four different formulas, the problem isn't insufficient diversification — the problem is that models are making errors.

"How do you create a pull request on GitHub?" You want a consistent procedure. Four different procedures are confusing, not enriching.

Signal: if you get diverse answers to a factual question, don't try to diversify more. Verify with a primary source.

When to diversify:

Strategic questions: "Should we expand to a new market?"
Value-based questions: "Is this approach ethical?"
Predictions: "How will the market evolve in the next five years?"
Questions involving trade-offs: "What do we sacrifice if we choose X?"

For these questions, there is no single correct answer. Perspective diversity is a feature, not a bug.

Formulation Checklist

| Question Type | Technique | Example Transformation | |---------------|-----------|----------------------| | Strategic | Perspective framing | "Is X good?" → "What are X's benefits from Y's perspective?" | | Decision-making | Role-assignment | "Should we do X?" → "You're the CFO. What are X's risks?" | | Evaluative | Contradictory | "Does X work?" → "Why does X fail?" + "Why does X work?" | | Factual | None (you want convergence) | "What's the formula?" → "What's the formula?" |

Multi-model approach brings value only when models receive questions that activate different parts of their knowledge. The same question to four models is waste. Perspective framing, role-assignment, and contradictory formulation are tools for deliberate diversification.

Tools like CrossChat let you assign different roles to different models within a single workflow and synthesize their answers automatically. Question formulation, role selection, and answer aggregation in one interface.

Sources

Heuer, R. J. (1999). Psychology of Intelligence Analysis. CIA Center for the Study of Intelligence. https://www.cia.gov/resources/csi/books-and-monographs/psychology-of-intelligence-analysis-2/
Tversky, A. & Kahneman, D. (1981). The Framing of Decisions and the Psychology of Choice. Science. DOI: 10.1126/science.7455683.
Wei, J. et al. (2022). Chain-of-Thought Prompting Elicits Reasoning in Large Language Models. arXiv:2201.11903. DOI: 10.48550/arXiv.2201.11903.

Editorial History

Concept: Claude Code + Anthropic Sonnet 4.6 Version 1: Claude Code + Anthropic Sonnet 4.6 Version 2: Codex + GPT-5.2 Quality audit (2026-03-23, Claude Code + Claude Opus 4.6): added Claims Framework, verified sources, language polish.