AI is incredibly useful. It’s also, in a strange way, completely useless.

Those two things can be true at the same time.

There’s no real argument over whether AI can do useful work. It can complete an enormous number of tasks quickly and cheaply. It can modify a recipe, generate ideas for a brand, organize a pile of notes, write code, analyze information, or help build an entire piece of software. I’ve used it heavily almost every day since ChatGPT became something I could download as an app. I’ve paid for different tools, tried different models, and used them for tasks at nearly every level of complexity.

The reason I call AI “completely useless” isn’t that it does nothing well. It’s that the expectations surrounding it have become so extreme.

AI hasn’t merely been presented as a useful tool. We’ve been told that it’s revolutionary, that it will automate everything, replace half the workforce, and take everyone’s job. Those expectations are built into the valuations of the companies creating it and into the way people talk about it. If that’s the standard, then its mistakes—especially its basic mistakes—matter a lot more.

The gap between what AI can actually do and what people have been led to expect is where the conversation becomes interesting.

AI Is Very Good at Being Okay

My biggest takeaway from using AI is that it’s very good at being okay and terrible at being great.

It can produce a competent design, a plausible strategy, acceptable code, or a reasonable explanation in seconds. That’s genuinely valuable. But it’s much less reliable when the goal is to produce exceptional work—the best design, the cleanest architecture, the sharpest strategy, or the kind of original insight that changes how you see a problem.

That makes intuitive sense when you consider how large language models work. I’m not an AI researcher, so I’m not going to pretend to explain the machinery in technical detail. But these systems learn patterns from enormous amounts of existing material and generate likely responses from those patterns. They’re naturally excellent at producing something that resembles the center of what has already been done.

The problem is that average work presented with extraordinary confidence can look a lot like expertise.

Your Expertise Changes How Intelligent AI Appears

Ask AI for help in a field you know extremely well, and you’ll identify its weaknesses almost immediately. You’ll notice the assumptions it made, the context it missed, the options it failed to consider, and the details that sound right but aren’t.

Ask it about a subject you know nothing about, and the experience can be completely different. Compared with you, it appears to be a genius. It responds instantly, uses the right terminology, provides a detailed plan, and explains everything confidently. You walk away believing you’ve received expert advice.

But the quality of the answer may not have changed. Your ability to evaluate it did.

That’s one of the most dangerous parts of using AI. When you don’t know what you don’t know, you also don’t know which questions were left unasked. You can’t see the missing requirements, bad assumptions, or subtle errors. If the system doesn’t know what to do next, it may still produce something plausible rather than clearly signaling uncertainty.

Confidence isn’t the same thing as correctness, but AI makes the two very easy to confuse.

AI Works Best as a Worker, Not a Replacement for Judgment

If your goal is to outsource your thinking, planning, strategy, and skill set to AI, you’re likely to run into trouble.

You can ask it to act like a designer, developer, strategist, writer, and analyst all at once. It will happily accept every role. The problem is that if you don’t understand those disciplines, you can’t reliably tell whether the work is being done correctly or why one approach is better than another. AI can take you incredibly far in the wrong direction, and you may not know the difference until the consequences begin appearing.

On the other hand, AI can be extraordinarily effective when you already understand the work.

You don’t necessarily need to be the world’s leading expert. You need enough knowledge to define the outcome, give clear instructions, recognize bad decisions, and check the result. If you could perform the task yourself but it would be tedious, repetitive, or time-consuming, AI can make you dramatically more efficient.

The distinction is simple:

AI is excellent at accelerating execution. It’s much less dependable as a substitute for the judgment required to direct that execution.

A WordPress Example

Imagine I’m planning a large, content-rich WordPress website for a client. The site needs custom post types, fields, taxonomies, and relationships between different types of content.

Because I understand WordPress architecture, I can decide exactly what the system needs. I know which content should be a post type, which information belongs in a field, when a taxonomy makes sense, how different records should relate to one another, and how the data will eventually be queried in the templates.

I can give AI a detailed set of instructions and ask it to generate an importable configuration. In the right circumstances, the first result may get me 95 percent of the way there. What could have taken hours of repetitive dashboard work takes minutes instead.

That’s an excellent use of AI. It didn’t replace the architecture or the judgment. It executed a plan I had already made.

Now imagine someone attempting the same task without that experience. They may know roughly what they want the finished website to do, but they don’t know what a custom post type is, when to use a taxonomy, how a relationship should work, or which details need to be specified.

AI will still produce a file quickly. The file may even import successfully. That creates the feeling that an enormous amount of progress has been made.

Then the building begins.

A page doesn’t display the right records. A relationship works in one direction but not the other. A field stores the wrong type of value. A query returns nothing. The code solves one issue and creates two more. Each problem appears separately, so the user goes back, explains the latest symptom, applies another patch, and continues until the next failure appears.

Eventually, the time spent diagnosing and repairing the AI-generated system can exceed the time that would have been required to build it properly in the first place.

The user received exactly what they asked for. The problem was that they didn’t know everything they needed to ask for.

The Real Advantage Belongs to People With Skills

There’s a popular idea that AI makes expertise less important because anyone can now ask a machine to perform specialized work. My experience has led me toward almost the opposite conclusion.

AI makes expertise more powerful.

The person who already understands the field can use AI to remove enormous amounts of manual labor. They can test ideas faster, produce first drafts, automate repetitive work, and move through execution at a speed that would have been impossible a few years ago. They also know when to ignore the output, change direction, or do the work themselves.

The person without that foundation may still produce more than they could before, but they’re also more vulnerable to false confidence, hidden mistakes, and endless rework.

AI isn’t useless. It may be the most useful tool I’ve ever used. But it becomes useless surprisingly quickly when we ask it to replace the very knowledge required to use it well.

The best way to think about AI isn’t as a team of experts you can hire for a monthly subscription. It’s an extremely fast worker that needs direction, context, supervision, and quality control.

If you bring those things to the relationship, it can multiply what you’re capable of doing. If you don’t, it can multiply your mistakes just as efficiently.