Are You Blaming the AI When the Real Problem Is the Question?

If AI gives you a disappointing answer, what’s the first thing you blame?

The AI?

The prompt?

Or could the real problem sometimes be that you haven’t clearly defined what you actually need?

I started looking into this because it connects to a bigger question we’re researching at The Income Edge: when someone becomes dissatisfied with an AI tool, how often do they actually need a different or more specialized AI—and how often is something else going wrong?

I asked AI users a simple question: “Do people actually know how to ask AI for the result they need?”

The discussion has now reached thousands of people and produced more than 20 comments across two AI communities. The answers are more complicated than “learn how to write better prompts.”

People described problems with unclear requirements, missing context, lack of subject knowledge, AI limitations, verification, correction and rework. Several also described something I hadn’t considered strongly enough when I asked the question:

Sometimes AI can help you figure out what you should have asked in the first place.

This article is still being researched. I’m documenting the findings now and will update it as people respond to the follow-up questions.

Finding #1: A Better Prompt Won’t Fix an Undefined Problem

The strongest pattern in the discussion wasn’t about clever prompt formulas.

It was about requirements.

Several experienced AI users described situations where the real problem appeared before the prompt was ever written: they hadn’t completely defined what they wanted the finished result to be.

One commenter compared it to giving instructions to a new employee. If you handed a new hire a few vague sentences without explaining the objective, audience, constraints, examples or what success looks like, you probably wouldn’t expect consistently excellent work. Why should we expect AI to fill in all of those blanks correctly?

Another user described repeatedly tweaking prompts and still receiving mediocre answers before realizing the underlying requirements weren’t clear enough.

And another put the problem particularly well: sometimes arguing with AI about a “bad” result exposes that you never actually defined what “correct” looked like.

That’s an important distinction.

If you don’t know what a successful result looks like, changing a few words in the prompt may not solve anything. And switching to another AI tool may simply move the same undefined problem to a different piece of software.

Finding #2: Sometimes You Don’t Know Enough to Ask the Right Question

This is where the discussion became more interesting.

Telling someone to “be more specific” assumes they already know enough about the subject to know what information matters.

Sometimes they don’t.

One commenter described needing help understanding tax laws outside their area of expertise. Instead of pretending to know the correct terminology, they explained the problem using the language they understood. The AI responded with follow-up questions and choices that gradually helped narrow down what they actually needed.

Another commenter raised the other side of this problem: domain knowledge helps you tell the AI what is right, what is wrong, what it can decide and what constraints it should follow.

That creates an uncomfortable question for anyone using AI to learn something unfamiliar:

If you don’t know enough about the subject to formulate the right question, do you know enough to recognize a wrong answer?

This doesn’t mean AI can’t help with unfamiliar subjects. It suggests a different way to use it.

Instead of immediately asking AI for the finished answer, you can first ask it to help you understand the problem: What information is missing? What assumptions am I making? What questions should you ask me before answering?

In other words, AI may sometimes be more useful as an interviewer before it becomes the answer machine.

Finding #3: Let the AI Ask You Questions Before It Answers

Several commenters independently described a simple technique that may solve part of this problem:

Don’t make the AI guess what’s missing. Ask it to question you first.

One user said they ask AI to identify what it needs to know before answering. Another noted that this can expose assumptions the user didn’t even realize they were making.

A more structured version is to ask AI to create a plan first, break that plan into milestones, and look for places where one “task” actually contains several different jobs. Problems can surface during that planning process before the AI spends time producing the wrong finished result.

This suggests a practical workflow:

1. Explain what you’re trying to accomplish in your own words.
2. Tell the AI not to solve it yet.
3. Ask what information, requirements or constraints are missing.
4. Have it ask you the questions it needs answered.
5. Define what a successful result should look like.
6. Only then ask it to perform the task.

That’s very different from searching for a magical “perfect prompt.”

You’re using the conversation to build the requirements before building the answer.

And that may be especially valuable when you’re considering paying for another AI tool. Before assuming your current AI isn’t capable enough, it gives you a way to test whether the real bottleneck is the software—or whether the job simply hasn’t been defined clearly enough yet.

Finding #4: A Clear Request Still Doesn’t Guarantee a Correct Answer

There’s an important warning on the other side of this argument.

It would be a mistake to conclude that disappointing AI results are always the user’s fault.

One commenter pushed back on that idea directly. Even when a request is discrete and well defined, the AI may simply lack the capability, information or reliability required to produce a correct answer.

Another described how accumulated context can sometimes cause a model to become fixated on something and perform worse rather than better.

And an experienced user who already knows how to perform their work manually said that even after giving AI good initial instructions, a significant amount of steering and correction can still be required.

They still considered AI a substantial productivity improvement—but that’s an important distinction.

The real question isn’t simply:

“Can AI do this?”

It’s:

“After I include the time required to instruct it, verify it, correct it and sometimes redo its work, is AI still saving me enough time or money to make it worthwhile?”

That’s also why simply buying another AI subscription may not solve the problem. Before switching tools, you need to know which problem you’re actually trying to fix: unclear requirements, insufficient subject knowledge, poor instructions, model limitations, unreliable output—or simply too much correction and rework.