A Matter of Perspective: Why AI Answers Are Often "Right" and Still Wrong
What I learned treating ChatGPT as a room full of experts instead of a single authority.
In my last post, I shared how AI helped me build a fantasy-league-style app called OctaPicks by dividing the work across three AI assistants: ChatGPT (the “Brain”), Claude (the “Architect”) and Cursor (the “Worker”).
That setup solved one big problem I had: drift. When I used Cursor alone, it kept slowly wandering away from the original requirements: it added duplicate code, pulled in unnecessary libraries and changed existing behaviors unexpectedly. Locking requirements and designs in markdown specifications, and handing them over to Claude and Cursor made the implementation much more stable.
However, as I kept building, I bumped into a very different problem upstream. While ChatGPT’s answers were detailed and technically “correct”, they were not what I needed. The problem wasn’t that AI was hallucinating or misunderstanding the spec. It was answering from a perspective I hadn’t asked for.
For individual contributors, this is just annoying. However, for people who lead teams, it’s dangerous: AI can produce answers that seem reasonable, but are optimizing for the wrong goal. At organization scale, this can result in shipped work that is misaligned with the strategy, even if it feels strangely “correct”.
What I eventually realized was that the problem wasn’t accuracy. It was perspective.
Two Experts, One Question, Two Incompatible Truths
If you ask the same question to two experts in different fields, you’ll often get answers that sound detailed and confident, but are incompatible.
I ran a very simple experiment with ChatGPT:
Prompt — “As an Economist, tell me how many students should we have in school classrooms”
Response:
“There isn’t one magic number, because the “right” class size is a cost–benefit decision: you’re trading off more teachers/rooms (cost) against better learning + fewer disruptions + higher long-run earnings (benefit). An economist answer is: pick the size where the next student you add costs more (in lost learning) than it saves (in staffing):
- 15–18 for ages 4–7
- 18–22 for ages 8–11, and
- 22–28 for ages 12–18”
Prompt — “As a pedagoguy students should we have in school classrooms?”
Notice how the wording implies that fewer students is better. It already hides a value judgment.
Response:
“Learning is relational. If the teacher cannot see, know, and respond to each student regularly, the class is too large. Class size is not about silence or control. It’s about attention, feedback, and belonging:
- 10–15 for ages 3–6
- 12–18 for ages 6–12, and
- 15–22 for ages 12–18”
Even if we pretend the age ranges provided are the same, the economist suggests between 20% and 80% more students per classroom than the pedagogue.
Put another way: both answers are valid, but they optimize for different things. The economist is optimizing for affordability and cost-effectiveness. The pedagogue is optimizing for learning quality and relationships.
They are both right, yet, they cannot both be satisfied at the same time.
This isn’t disagreement caused by uncertainty, but one caused by different definitions of success.
Fro to Product Screens
Once I saw this in the classroom example, I started noticing the same pattern in my own work on OctaPicks.
- When you optimize for UX / UI, you sometimes compromise branding.
- When you optimize gamification, you compromise UX / UI.
- When you optimize performance, you compromise expressiveness.
Different “experts” inside your head — or inside your AI prompts — are pushing for different definitions of success.
When you ask ChatGPT, it will usually default to the role that’s most common in its training data or the one you implicitly set in the last few turns. Unless you tell it otherwise, it will happily answer from whichever perspective feels “natural” to the model, not necessarily the one you care about.
Here’s a concrete example. This is the home page for a logged-in user in OctaPicks. I asked two versions of the same question:
- As a UX / UI expert, can you give feedback about this screen in one sentence?
“…but the heavy glow/contrast and dens text hierarchy slightly hurt readability and scannability — dialing back effects and clarifying primary vs. secondary actions would make the screen faster to understand at a glance.”
- As a gamification expert, can you give feedback about this screen in one sentence?
“…the screen could improve engagement by surfacing progress, stakes, or social comparison (e.g., rank, points, or urgency cues) more prominently to turn ‘View Picks’ from a passive action into a compelling motivation.”
These answers are not aligned. One answer optimizes for clarity and speed. The other optimizes for emotional pull and habit formation.
The more complex and in-depth these questions are the bigger the differences will grow. If we let the UX / UI expert lead the design, we might completely miss an opportunity to add gamification cues that can improve the engagement, but if we follow the gamification recommendations blindly, we might end up with a very dense page that feels heavy and hard to read.
This doesn’t only affect look and feel. It also affects the data we surface on each screen, the page performance (the more data the slower the page), the database usage and cost, our caching strategy, and so on.
And that’s exactly why we need to put on our product manager hat, and start deciding what do we want to prioritize in this app, on this particular page, for this user state and the context state.
What finally clicked for me is this: AI answers are not objective truths. They are local optimizations.
When using ChatGPT as an assistant to build an app, every response it gives is an answer to an implicit question:
“Optimize this decision according to what?”
If you don’t specify the “according to what”, the model will guess. And it will usually guess based on:
- The most common framing in its training data
- Or the role it has implicitly adopted in the last few turns
That’s why the same question can yield answers that feel technically correct and yet deeply unsatisfying.
Once you see AI answers as local optimizations, you start noticing the same pattern all the time.
Ask for “the best onboarding flow”, and you might optimize conversion at the expense of brand perception. Ask for “safest architecture” and you might get a very robust design that hinders flexibility. Ask for the “quickest way to hit this sales target”, and you might get a strategy that benefits short-term sales but damages retention.
None of these are right or wrong per se, but if you don’t make the objective explicit, the model will silently choose one for you.
At that point, I quit trying to ask better questions, and I started deciding who got the microphone.
Roles as Lenses, Not Commands
This is where “roles” stop being a prompt trick and start being a design tool.
When I ask ChatGPT to “answer as a UX / UI expert”, I’m not asking it to “cosplay” a job title. I’m telling it:
- What to optimize for
- What to deprioritize
- What kinds of trade-offs are acceptable
Each role is a lens:
- UX / UI optimizes for clarity, speed, and cognitive load
- Gamification optimizes for motivation, reward loops, and emotion
- Retention optimizes for habits and long-term return visits
- Engineering optimizes for maintainability, performance, and cost
None of these lenses are “right” in isolation. They are incompatible by design, and that’s exactly why they’re useful: they surface trade-offs you would otherwise skip.
The Product Manager Problem
(and why AI makes it sharper)
In a traditional team, the tensions between these roles surface during planning, design reviews, and architecture discussions. People argue, defend their priorities, and negotiate trade-offs openly.
With AI, something subtle happens: because all the perspectives are compressed into a single voice, those debates disappear unless you force them back in.
If you let a single role dominate (or don’t ask explicitly at all), you can easily end up with:
- Beautifully clean but disengaging screens
- Highly engaging screens that feel overwhelming, or
- Elegant architectures that ship too slowly.
AI doesn’t resolve trade-offs for you. It hides them behind whatever answer is more common or more readily available to the model.
What at small scale looks like a product management problem becomes a leadership problem at larger scales: someone has to decide what is the definition of success for the product, for the team, for the company.
That’s why the most important role in this setup isn’t “The Architect”, “The Worker” or any of the roles “The Brain” can play. It’s still a human making a call. Deciding where the company is going, and how.
What This Means If You Are a Leader
If you are in charge of a product delivery or an engineering team, this isn’t just about better prompts, it has several material implications:
- In design and architecture reviews, treat AI-generated artifacts (whether they are specs, flows or recommendations) as single-perspective proposals instead of neutral truth. Ask: “What is this optimizing for?”
- Surface hidden perspectives by explicitly asking. Notice if any angles are missing and ask for them: “What would a UX / optimization / finance / branding lens say about this?”
- Encourage the whole team to think in terms of roles and lenses. Get them to surface trade-offs so you can choose the option that best aligns with your strategy.
You don’t need a huge process change for this. You just need one or two moments in your existing rituals where someone asks, “Which lens is this answer coming from, and is that actually the one we want right now?”
How I Use This in Practice
My workflow eventually stabilized around a simple rule:
- Ask the same question to multiple roles
- Let them disagree
- Make the trade-off explicit
- Decide intentionally
As a leader, you can turn this into a simple ritual instead of a personal habit.
For any important decision that’s touched by AI (e.g. new features, architecture updates, or price tweaks), try to run a couple of lenses on it (for example, UX and engineering, or growth and retention) and let them disagree on paper. Only then commit.
It adds a bit of time, but it drastically reduces the risk of shipping a “correct” answer that was optimizing for the wrong goal.
Sometimes the decision is:
“Yes, this hurts UX slightly, but engagement matters more here.”
Other times it’s:
“No, this gamification cue doesn’t justify the additional complexity on this screen.”
The key is that the decision is mine, not an emergent property of the last prompt.
I use the different roles to surface the trade-offs, then choose which “definition of success” wins for this specific feature, page, or user state.
Closing Thought
AI didn’t remove judgment from product building. It amplified the cost of skipping it.
It also hid the need to apply any judgment at all, which makes it much easier to miss.
When you treat ChatGPT as a single, neutral authority, you’ll get answers that feel polished but misaligned.
If you treat it as a set of specialists with competing incentives, you get something far more valuable: clarity about what you’re choosing.