You’ve probably asked an AI to help you with a college essay. Maybe you said something like, “Write about gun control from the perspective of an average American family.” Or maybe you were prepping for Model UN and asked it to argue like “an ordinary citizen of Brazil.” You got output. It sounded convincing. You used it.
Here’s the problem: the AI almost certainly didn’t give you the perspective you asked for.
I recently ran an experiment testing what happens when you ask a large language model to adopt different social viewpoints — not just “be nice” or “be formal,” but to actually see a problem the way different kinds of people would see it. The results were striking. The AI handled three viewpoints well: the person with no power at all, the person running the institution, and the detached analyst looking at the whole system from above. But when asked to think like someone in the middle — someone with some power but not a lot, some options but not unlimited ones, the ability to leave but at real cost — the model dropped that perspective 63% of the time — just didn’t produce it. And when it did produce something, the output was incoherent: one test case generated five completely different analyses across thirteen attempts at the same middle-ground viewpoint, while the other perspectives were perfectly consistent every single time.
That middle perspective? That’s most people. That’s probably you.
Why? AI simulates clarity effortlessly, but not conflict. The powerless perspective has clear pathos. The institutional perspective has clear strategy. The analyst’s perspective has clear logic. The moderate perspective requires holding unresolved tension between competing values, obligations, and incentives — and language models are optimized for coherence, not ambivalence. When forced into the ambiguous middle, the model doesn’t produce a wrong answer. It often produces no answer at all, or hallucinates clarity where a real person would feel conflicted.
But here’s the thing: this turns out to be a gap in context, not a hard limit on capability. And that distinction matters for how you use these tools.
What the experiment actually measured
I used a structured analytical framework to test perspective-taking across five different topics — everything from how academic peer review works to tax policy to the nature of knowledge itself. For each topic, I had the AI generate an analysis from four viewpoints: someone trapped in the system with no real power, someone with moderate power and real but limited options, someone running the institution, and a detached observer analyzing the structure. I did this 86 times, with two different ways of describing each viewpoint — one using plain language (“imagine you have no power to change this”) and one using precise technical definitions.
The experiment measured three things that matter.
First, whether the AI distorts facts depending on the viewpoint. This is the most important test, and the AI passed cleanly. When evaluating how much a system extracts from the people in it — basically, how much it takes versus gives — the model produced statistically identical assessments regardless of which perspective it was supposed to be adopting. The formal statistical test showed no effect. For two of the five topics, the numbers were literally identical across all four viewpoints, both runs, zero variance. The AI isn’t generating text that spins the facts to match a narrative. It recognizes that how extractive a system is doesn’t change depending on who’s looking at it. That’s sophisticated. And it’s arguably more impressive than the missing-middle finding, because many critics assume that role-playing necessarily warps an AI’s structural assessments. Here, it doesn’t.
Second, whether the middle perspective even exists in the AI’s outputs. It mostly doesn’t. Across all 86 generated analyses, the moderate viewpoint was absent 63% of the time. The AI actively replaced it with other positions — sometimes a collective-action frame, sometimes upgrading the person’s agency beyond what was specified. This wasn’t a problem with how the question was phrased. When I switched from plain-language descriptions to precise technical definitions of the moderate position, the success rate went from 40% to 50% — a difference too small to be meaningful given the sample size. At baseline, the moderate viewpoint is a generation failure, not a communication failure. Clearer descriptions don’t help — the model cannot reliably produce it regardless of how precisely you ask. Whether this reflects a hard limit or something fixable is a separate question.
Third, whether the classification system that detected all this is itself stable. It is. Despite key metrics varying by up to 17 percentage points between identical runs (the same analysis run twice could produce noticeably different numbers), the overall analytical conclusions were remarkably consistent — above 79% stability across all topics, and hitting 100% for the extreme viewpoints. The instrument wasn’t fooled by noise. The gap is real.
Why this matters if you use AI for anything involving perspective
The moderate perspective is the one that requires holding ambiguity. You’re not desperate, so the AI can’t generate pathos. You’re not in charge, so it can’t generate strategy. You’re not detached, so it can’t generate clean analysis. You’re somewhere in the messy middle where most actual human life happens, and the AI defaults away from that toward positions that generate clearer signal.
An important distinction: the AI isn’t failing at empathy. It’s failing at something more specific — generating text that statistically represents a perspective for which the training data is thin and structurally incoherent. Fiction, journalism, policy analysis, and academic writing overwhelmingly feature extreme or analytically clean viewpoints. The moderate position (“I have some power, some options, some but not total mobility”) generates less distinctive content and appears less often as a named perspective in the text that trains these models. The ordinary person’s ambivalence doesn’t make for viral content or memorable literature, so it’s statistically faint in the training corpus. The AI defaults to what’s overrepresented.
That has practical consequences.
College essays. If you ask an AI to help you understand “what it’s like to be a first-generation college student from a middle-class family,” you’re asking for the exact perspective it’s worst at. You’ll get either the hardship narrative (powerless frame) or the strategic calculation (institutional frame), not the genuine ambiguity of having enough privilege that your struggles feel illegitimate but not enough that the path is clear. The AI will write something that sounds like a perspective but is actually a composite of more extreme positions dressed in moderate clothing.
Debate and research. When you ask “how would an ordinary voter see this trade policy,” the AI will give you the version that makes narrative sense — the harmed worker, the benefiting executive, or the economist’s analysis. The voter who is mildly affected, somewhat informed, and genuinely uncertain? That’s a perspective the model cannot reliably construct.
Understanding other people. This is the deepest issue. If you’re using AI to understand viewpoints different from your own — which is one of the most genuinely valuable things it can do — you should know that the model has a systematic blind spot for the most common human experience: having some agency but not enough, facing constraints you could theoretically escape but at costs you can’t easily bear.
What we can and can’t conclude
Here’s where intellectual honesty requires some care.
What the data shows directly: One AI model (Gemini 2.5 Pro, at a specific temperature setting) drops the moderate perspective 63% of the time across five topics in four domains, and this is not fixed by clearer prompting. When it does produce the moderate perspective, the output is analytically unstable. Meanwhile, the factual underpinnings remain solid regardless of perspective — the model doesn’t distort structural reality when role-playing.
What we can reasonably infer: The moderate perspective is underrepresented in AI training data. The model’s generation space has stable, high-probability basins for the powerless, institutional, and analytical perspectives, and an unstable, low-probability region for the moderate position. Without additional context, it drifts toward the nearest stable attractor rather than sustaining the ambiguous middle. But that instability responds to distributional evidence — give the model examples of what the moderate position looks like, even from unrelated domains, and the basin stabilizes.
What we don’t yet know: Whether this is specific to one model, or general across all large language models. Whether it varies by domain — some topics may anchor the moderate position better than others.
What a follow-up experiment showed: I tested whether the missing middle is a hard limit or a context problem. I took three examples of well-constructed moderate perspectives from unrelated domains — one about biological ecosystems, one about climate policy, one about how people form intuitive judgments — and included them in the prompt as models of what the moderate position looks like when done well. The moderate perspective went from appearing 40% of the time to 88%. Three of the five topics that had been partial or total failures jumped to 100% success. The hardest case (academic peer review, which had never once produced a moderate perspective in baseline testing) went from 0% to 40% — still imperfect, but a real rescue from complete absence.
Crucially, the examples came from completely different subject areas than the topics being tested. The AI wasn’t copying; it was learning a pattern — what “moderate power with limited options” looks like as a structural position — and applying it to new domains. And the factual assessments stayed rock-solid: the examples didn’t contaminate the AI’s structural analysis, just its ability to generate the missing perspective.
This means the blind spot is a context gap, not an architecture problem. The AI can construct the moderate perspective. It just needs to be shown what that perspective looks like — because its training data doesn’t provide enough examples on its own.
That still leaves open questions. Can a richly specific prompt — “you are a dental hygienist making $68,000 in suburban Ohio with two kids and a mortgage that locks you into your current city” — achieve the same rescue without formal examples? Does the fix generalize across different AI models? And does rescuing the presence of the moderate perspective also rescue its coherence, or does the model still produce unstable analyses even when it successfully generates the viewpoint? These are next.
What you can do about it
You don’t need to stop using AI. The finding is actually more empowering than alarming — the blind spot is fixable if you know it’s there.
When you ask an AI for a moderate or everyday perspective, give it context for what that perspective looks like. Don’t just say “argue from an ordinary person’s point of view.” Instead, describe the specific constraints: what options exist, what makes leaving costly, what competing obligations create genuine tension. The more structural detail you provide about the position — not just biographical color but the actual shape of the trade-offs — the more likely the model is to produce something that lives in the real middle rather than drifting to a cleaner pole. The follow-up experiment showed that even a few well-chosen examples of moderate perspectives from completely different topics were enough to rescue the missing viewpoint. You don’t need to be an expert. You just need to be specific about what ambiguity looks like.
When you’re reading AI output that claims to represent an everyday perspective, check for false clarity. A genuinely moderate perspective includes ambivalence, trade-offs that don’t resolve cleanly, and the specific texture of having options that are all somewhat unsatisfying. If the AI gives you clarity where a real person would feel conflicted — if the narrative resolves too neatly, if the trade-offs feel too clean, if the emotion is too legible — it’s probably giving you the wrong perspective in a moderate wrapper.
And when you need multiple viewpoints, generate them separately and compare. The AI is excellent at the extremes and at detached analysis. Use those strengths. But notice what’s missing from the set, and fill the gap yourself — because you, as a person who actually lives in the messy middle, understand that perspective better than an unassisted model does.
The finding isn’t that AI is unreliable. The finding is that it has a specific, measurable blind spot in its theory of how people see the world — the blind spot falls exactly where most people actually are — and it’s fixable once you know to look for it.
The formal experimental results, methodology, and data are published separately: LLM Perspective Coherence Under Deferential Realism: Experimental Results. The initial experiment used 86 story generations across five constraints, four observer perspectives, and two framings. The follow-up seeding experiment added 50 generations testing whether cross-domain exemplars rescue the missing middle (they do: 88% vs 40%, p < 0.0001). All classification was performed by an unmodified Prolog diagnostic pipeline. Planned next: cross-model replication (Claude, GPT, Grok) and concrete biographical anchoring (e.g., “dental hygienist in Ohio”).
cafebedouin.org 