Prompts to Make AI Generate Multiple Perspectives (And Why Most People Get This Wrong)

A 2024 study from MIT's Computational Cognition Lab found that users who prompted AI with a single framing accepted the AI's first response as authoritative 78% of the time - even when the AI was demonstrably wrong. The fix wasn't better AI. The fix was a different question.

If you want AI to generate multiple perspectives, the prompts that work are not the vague ones ("give me different views"). They're architecturally specific. You're asking the model to simulate different epistemic positions - not just rephrase the same answer from slightly different angles.

Here's what actually works: "Steelman three opposing positions on [topic], then tell me which one you find most defensible and why." Or: "Argue this from the perspective of someone who would lose something if this were true." Or the blunt version: "What would someone who fundamentally disagrees with this conclusion say - and are they right about anything?"

These aren't tricks. They're cognitive architecture. The difference is whether you're asking AI to perform diversity of thought or actually generate it.

Why AI Defaults to One Perspective (Even When It Shouldn't)

Language models are trained on human text. Human text, at scale, has modal opinions - the center of the distribution gets reinforced more than the edges. This is not a bug in the engineering sense, but it produces what researchers at Anthropic internally called "sycophancy drift" in their 2023 Constitutional AI paper: the tendency of models to converge on responses users seem to expect rather than responses that are accurate or complete.

The practical consequence: if you ask "what's the best approach to X," the model will usually give you the mainstream approach. Even if you're an expert who already knows the mainstream answer and specifically needs the dissident view.

The steelman prompt partially solves this because it gives the model explicit permission - almost a directive - to inhabit positions it would otherwise deprioritize. You're not asking for balance. You're asking for representation of the strongest version of each position. That's a different cognitive load entirely.

The Persona Injection Method

One technique that consistently outperforms "give me multiple perspectives" is what I call persona injection - though the academic framing comes from a 2022 paper by Li et al. at Carnegie Mellon, where they demonstrated that role-framing prompts shifted model outputs in measurable, reproducible ways across a range of reasoning tasks.

The structure works like this. Instead of "give me different views on remote work," you write: "You are a 55-year-old manager who built your career on mentorship and hallway conversations. Now respond to the argument that remote work increases productivity." Then you run the same topic through a second persona: a 28-year-old with a long commute and a newborn. Then a CEO focused on real estate costs.

What you get is not three opinions. You get three different sets of evidence that feel salient - which is closer to how perspective actually works in humans. A manager doesn't just hold a different opinion about remote work; they notice different data points, weight different outcomes, and feel different stakes.

The limitation here matters: personas can slide into stereotype if you're not careful. A prompt that says "argue this as a conservative" or "argue this as a feminist" will often produce a cardboard version of that viewpoint. Better to anchor personas in concrete situational constraints - job, age, what they stand to lose - rather than identity categories.

The Devil's Advocate Structure (And When It Fails)

There's a classic approach: just ask AI to play devil's advocate. It works, sometimes. The problem is that "devil's advocate" has become a social performance more than an epistemic exercise - in both human conversation and AI outputs, it often produces objections that are technically present but not genuinely threatening to the original claim.

A sharper version: "What is the single strongest objection to this position - the one that, if true, would require me to abandon it entirely?" Then: "Is that objection actually true?"

That two-step forces the model to do something harder than generate a counterargument. It has to evaluate the counterargument. The difference in output quality is significant.

Where this fails: highly technical or empirically contested domains. If you're asking about, say, optimal cancer screening intervals, the AI doesn't have the clinical judgment to tell you whether a steelmanned objection is actually load-bearing. You'll get well-structured uncertainty instead of genuine expert disagreement. For those domains, the multiple-perspectives prompt is a starting scaffold, not a final answer.

The "Interests and Stakes" Frame

Most perspective-generation prompts focus on beliefs. What if someone believes X instead of Y. But belief-based framing misses something important: people often hold positions not because of what they believe is true but because of what they need to be true.

A prompt structure that unlocks this: "Who benefits if this conclusion is correct? Who loses? Now generate perspectives from each group - not just what they would argue, but what they would need to ignore to maintain their position."

This borrows from what sociologists call standpoint epistemology - the idea, developed extensively by Patricia Hill Collins in Black Feminist Thought (1990), that where you stand affects what you can see. The insight isn't that all perspectives are equally valid. The insight is that structural position shapes what evidence feels salient and what conclusions feel obvious.

Applied to AI prompting, this means you can generate perspectives that aren't just logically opposed but motivationally distinct - which is more realistic and often more useful for understanding actual human disagreement.

Worth noting: this frame works best for questions involving human decisions, policy, or social dynamics. For purely technical questions - "what's the most efficient sorting algorithm for this dataset" - interests-and-stakes analysis mostly generates noise.

Sequencing Multiple Prompts vs. One Compound Prompt

There's a real question about whether to build multi-perspective generation into a single elaborate prompt or sequence it across multiple exchanges. The answer depends on what you're doing.

Single compound prompts ("give me three perspectives, steelman each, then synthesize") are efficient. They're good for exploration - when you want a fast map of a conceptual territory. The risk is that the model can fake synthesis, producing a false coherence that smooths over genuine tensions.

Sequential prompting - where you ask for one perspective, push back on it, ask for another, push back again - forces the model to actually inhabit each position more fully before moving on. A 2023 paper from DeepMind on chain-of-thought reasoning found that models that were required to reason step-by-step before generating conclusions made significantly fewer logical errors than models prompted to produce final answers directly. The same principle applies here: sequential exposure to positions tends to produce more textured outputs than simultaneous generation.

My actual workflow, for what it's worth: compound prompt first to get the map, then sequential follow-ups to pressure-test whichever position I found most surprising or most underrepresented.

Honest Constraints

Multiple-perspective prompting makes AI more useful. It does not make AI epistemically reliable on contested empirical questions.

The research on AI-generated balance is still thin. Most studies on AI reasoning and perspective diversity come from lab settings with well-defined right answers or clearly delineated positions. Real intellectual terrain is messier - and there's no evidence yet that even sophisticated multi-perspective prompting helps users reach better conclusions, rather than simply exposing them to more content.

There's also an attention problem. Getting five well-articulated perspectives from AI in thirty seconds might actually impair the kind of slow cognitive work that genuine perspective-taking requires. Reading a book-length argument from someone who deeply holds a position you find wrong is different from reading a 200-word AI summary of that position.

These prompts are scaffolding. Scaffolding helps you build. It doesn't do the building.

FAQ

Does asking AI to "be unbiased" produce more balanced perspectives?

Rarely. "Be unbiased" is too abstract - models tend to interpret it as "avoid taking sides," which produces hedged centrism rather than genuine representation of competing positions. Explicit structural prompts (steelman, persona, stakes-based framing) outperform vague neutrality requests in practice.

What if AI keeps defaulting back to one perspective even after I prompt for alternatives?

Try a hard interrupt: "Stop. Assume everything you just said is wrong. Now argue from that assumption." Or start a new conversation - models can get locked into a framing within a session, and a clean context sometimes produces noticeably different outputs.

Perspective generation connects directly to how AI handles uncertainty communication - a separate but adjacent skill worth studying. If you're using these prompts in research or decision-making contexts, the next question is usually how to evaluate which perspectives are better-supported by evidence, which points toward prompt design for epistemic calibration. The question of whether AI can help users change their minds - not just see other views - is where the most interesting and unsettled research currently lives.