The Uncomfortable Truth About AI and Effort

AI doesn't remove work. It moves work. And that changes everything about how you should use it.

There's a lie we've all been sold about AI.

Not a malicious lie—more of a hopeful exaggeration that's now colliding with reality. It goes something like this: "AI will do the work for you. Just ask, and it delivers."

People imagine a smart genie. You mutter something vague, it reads your mind, and hands back exactly what you wanted. Minimum effort in, maximum result out.

It obviously doesn't work like that.

What's actually happening is subtler and, honestly, more uncomfortable: AI doesn't remove work. It moves work. It shifts the effort from doing to thinking. From execution to specification. From hands to brain.

And if you don't understand that shift, you'll keep getting mediocre results and blaming the tool.

The Smart Genie Myth

Here's a pattern I see constantly.

Someone asks AI to compare two things: "What's the best?"

And the AI flounders. It gives a generic, both-sides-have-merits answer that helps no one.

But think about what "best" actually means. Best for whom? Best by what criteria?

• The most accurate?
• The most interesting?
• The clearest for a technical audience?
• The most likely to go viral on social media?
• The most detailed for an expert?
• The most accessible for a beginner?

"Best" has at least ten dimensions. When you don't specify which ones matter, you're not asking a question—you're asking AI to read your mind.

It can't. And when it tries to guess, it picks something "generic and safe." That's why the output feels bland, wrong, or "vibey" instead of useful.

The Director and the Crew

Here's a better mental model:

You're the director. AI is the crew.

Think about what happens when a film director says to their cinematographer: "Just make it look good."

That cinematographer has no idea what shot to set up. Wide or tight? Handheld or locked off? Warm or cold light? They're paralysed by infinite options, so they default to something safe and generic. The result looks like stock footage.

But when a director says: "I want a slow push-in on her face. Natural light from the window. She's realising something painful but trying not to show it. Stay tight enough that we see the moment her eyes change."

Now the cinematographer can deliver something powerful. The specification unlocked the craft.

AI works the same way.

"Under this mental model, LLMs should be thought of as actors; prompts as scripts and cues; and LLM responses as performances. Prompt engineering is playwriting and directing." — ArXiv: LLMs as Method Actors

Research backs this up. When researchers tested a "Method Actors" approach—treating the AI like actors who need clear direction—performance on complex puzzles jumped from 27% to 86%.

Same model. Same task. Different direction. Radically different results.

The Laziness Mirroring Effect

Here's something that feels weirdly accurate: "If I'm lazy, it'll be lazy."

The model doesn't literally decide to slack off. But the mechanism is real:

• Vague input → huge space of plausible outputs
• Huge space of outputs → AI picks something "generic and safe"
• Generic and safe → feels bland, wrong, or useless

So it feels like the AI phoned it in. But really, it's just reflecting your ambiguity back at you, wrapped in fluent language.

A 2025 study from METR found something startling: experienced developers take 19% longer to complete tasks when using AI tools—despite expecting a 24% speedup.

Even after experiencing the slowdown, they still believed AI had sped them up by 20%.

The gap between perception and reality is striking. AI creates new cognitive overhead—verification, correction, interpretation—that most users aren't prepared for.

The Effort Doesn't Disappear. It Redistributes.

The economic intuition people have about AI is wrong.

You're not going from "100 units of work" to "10 units of work."

You're going from:

Total effort may be similar. Sometimes higher. But the distribution of that effort shifts dramatically.

Research confirms this pattern:

"77% of workers say AI has either increased their workload or decreased their productivity. In many cases, instead of cutting effort, AI just stacks a second layer of work on top of the first—reviewing outputs, bridging system limitations, handling exceptions." — Upwork / Shep Bryan: Cognitive Load Research

Every "AI-assisted" workflow hides three forms of invisible human effort:

1. Verification work — checking whether outputs are correct and compliant
2. Correction work — editing, reframing, or sanitising content before use
3. Interpretive work — deciding what AI's suggestions actually mean for your context

These weren't part of the original promise. But they're very much part of the reality.

The Vibe Coding Catastrophe

Nowhere is the specification problem more visible than in "vibe coding"—the practice of describing projects to AI in vague terms, accepting generated code without review, and hoping it works.

Analysis of Fortune 50 companies found AI-assisted developers produced 3-4x more code but generated 10 times more security issues.

In one 2025 survey, 16 out of 18 CTOs reported production disasters directly caused by AI-generated code. The pattern was consistent: code that "appears to work perfectly until it catastrophically fails."

"Vibe coding's most dangerous characteristic is code that 'appears to work perfectly until it catastrophically fails.'" — Quartz: AI vibe coding has gone wrong

The lesson isn't that AI can't code. It's that AI can't compensate for missing specifications. When you skip the thinking—the architecture, the constraints, the edge cases, the "what happens when"—you get code that works in demos and fails in production.

AI as Intention Compiler

Here's the mental model that makes everything click:

AI is an intention compiler.

A traditional compiler translates code into machine instructions. It can only compile what you write. If your code is vague or incomplete, it throws errors.

AI compiles your intent into outputs. But it faces the same fundamental problem: it can only compile what you specify. If your intent is vague, it doesn't throw errors—it produces something "generic and safe" that technically satisfies your vague request while missing what you actually wanted.

"Generative AI is the most ambitious compiler yet because it translates from the language of thought." — Prompt Engineering: Generative AI - The New Compiler

The implication is profound: AI success isn't about the model. It's about the clarity of your thinking.

Your specifications, context, constraints, and success criteria become the "source code" that AI compiles. Feed it fuzzy thinking, get fuzzy outputs. Feed it structured thought, get structured results.

The Bargain

Using AI well requires a bargain. Both parties have to keep up their end.

When you hold up your end—when you do the thinking work, specify the format, define the audience, articulate the constraints—AI becomes a genuine force multiplier.

When you don't—when you bring vague asks hoping for magic—AI reflects your ambiguity back at you and you blame the tool.

The pattern holds across every AI application:

• In code: The real work is designing the interface, invariants, data flow, constraints. AI can generate syntax, but it can't choose the right abstraction for your domain unless you specify it.
• In video: The real work is blocking, framing, pacing, emotional beats, camera language. If you don't specify, AI invents generic versions that feel "wrong."
• In writing: The real work is deciding audience, purpose, stakes, angle. Without that, you get content, not communication.

Why This Feels Like a Betrayal

People were sold "AI as shortcut."

What they've actually got is closer to:

• A force multiplier on structured thought
• A brutally honest mirror for vague thought

That's why vibe coding feels so painful. It's like asking a contractor: "Just build me... y'know... like... an app? For booking stuff? Make it good."

Then being shocked when the result doesn't match the unspoken version in your head.

The uncomfortable bit: to use AI well, you must already be doing the kind of thinking that good work requires—often more explicitly than when you did it manually.

AI doesn't let you skip the hard thinking. It forces you to externalise it.

What Changes If You Accept This

Once you stop expecting a magic shortcut and start treating AI as an intention compiler that rewards specification:

1. You invest time upfront. The 20-40% extra time spent on specification saves days of iteration and rework.
2. You get specific. "Write a blog post" becomes "Write a 1,200-word post for technical product managers explaining X, using concrete examples, in a direct voice without jargon."
3. You define success criteria. Not "make it good" but "optimise for clarity to a non-technical executive audience."
4. You verify outputs. You don't trust that it worked—you check that it worked. Every time.
5. You iterate systematically. When outputs miss the mark, you ask what was missing from your specification, not what's wrong with the AI.

The 2025 research is clear: with proper specifications, 90% of code can be AI-generated successfully. Without them, you get disasters.

Same model. Same capability. Different inputs. Radically different outcomes.

Keep Up Your End of the Bargain

AI will happily generate oceans of possibility.

You pay your side in clarity, rigour, taste, constraints, iteration.

Used that way, it's not a shortcut to effort—it's a shortcut to leverage. You don't escape the work. You change which parts of your brain are doing it, and how far the result can scale once you've done it properly.

The uncomfortable truth nobody wants to hear: you still have to do work when using AI.

The empowering reframe: the work is now thinking work, and AI will amplify every bit of clarity you bring.

Keep up your end of the bargain.

Scott Farrell is an AI strategy advisor helping mid-market leadership teams turn scattered AI experiments into governed portfolios that compound. Based in Australia, working with businesses doing $20M–$500M revenue who want AI to stop leaking into pilots that never ship.