AI Doesn't Level the Playing Field. It Multiplies What You Already Have.

When ChatGPT launched, the take I kept hearing was: juniors will code like midlevels, midlevels like seniors. The gap closes. Everyone levels up.

I've spent the last 6+ months using Claude Code on production tasks. That's not what I see.

A weak developer with AI writes bad code fast. A strong developer writes good code significantly faster. AI doesn't raise the floor. It stretches the gap.

The myth that AI levels the playing field

The narrative isn't wrong about everything. AI tools genuinely reduce the cost of entry. A junior can look things up less, get boilerplate generated faster, understand unfamiliar code in 30 seconds. That part is real.

But leveling assumes the gap came from information access. It didn't.

A senior engineer isn't slower because they take longer to look up syntax. They're slower because they're thinking about what happens in edge cases three months from now. About how this falls over under load. About why the previous team made that specific choice and not another.

Those questions don't come from prompts. And AI doesn't help you ask them.

What I observed using Claude Code for 6 months

I started using Claude Code heavily in early 2024. By November I had 200+ production tasks done with AI assistance: PHP refactoring, test writing, migration generation, PR reviews.

The pattern became clear around month three.

Tasks where I already understood the domain clearly — AI cut my time by 2-3x. I knew what I needed. I checked what Claude wrote. I told it where it was wrong. Output was fast and correct.

Tasks where I was fuzzy on the domain — AI created an illusion of progress. The code compiled. Tests passed. A week later in production, a bug appeared that I would have caught if I'd read more carefully. Because I would have gone slower and noticed something felt off.

Concrete example: cursor-based pagination with an off-by-one on exact multiples. The code looked right. Tests passed. It sat in production for seven days.

That's not Claude's fault. That's what happens when I didn't understand the problem well enough before handing it over.

Multiplier, not elevator

The math goes like this.

If AI is an additive tool — say +5 units to everyone — a 10-unit developer becomes 15 and a 3-unit developer becomes 8. Gap unchanged in absolute terms, ratio improves. Sounds like leveling.

But AI isn't additive. It's multiplicative. Something like 2-3x depending on the task.

10-unit developer at 2x = 20. 3-unit developer at 2x = 6. The absolute gap just doubled.

This matches what I see. The tool works exactly as well as you can set it up, verify its output, and recognize where it's likely to be wrong. Those are skills that multiply. They're not skills AI teaches you.

This is why spec-first discipline matters: before asking Claude to write anything, answer five questions about what you actually need. That only works if you already understand the problem.

What this means for hiring

I'm not hiring someone who "knows how to work with AI." I'm hiring someone who thinks clearly about code — and also uses AI.

"Knows how to work with AI" describes a prompting skill. What I want to know: can the candidate explain why their solution is correct? Can they predict what breaks first when traffic spikes? Can they see that Claude just wrote something technically valid but semantically wrong?

I've been using paid trial projects instead of interviews for a while. The trial task used to be a standard PHP case. Now I've added one sentence: use Claude Code or Cursor however you like.

The results are revealing. Strong candidates deliver better work than before — faster, cleaner, with clear reasoning. Average candidates deliver code that works but is unreadable. When I ask "why this approach specifically?" — silence.

AI helped the weaker candidate finish the task. It didn't help them understand what they'd built.

Red flags: the "AI-native developer" candidate

Things I've learned to watch for.

Can't explain choices — ask "why can't we have N+1 here?" and there's a pause. The prompt existed. The understanding didn't.

Tests at 80% coverage, bugs still exist — Claude writes tests for what it already wrote. Happy path passes. Edge cases weren't mentioned, so they weren't tested.

Done in 2 hours, review takes 4 — generation speed isn't the same as thinking speed. Fast delivery with slow review is a flag, not a green light.

Boilerplate without adaptation — AI generates pattern-book solutions. A developer adapts them to context. If the PR reads like a docs example with variable names swapped, ask if they understood the constraint they were solving for.

These aren't reasons to reject. They're reasons to ask a follow-up question.

The bar is going up, not down

I hear the fear that AI will replace developers. I don't share it.

What I see instead: AI will make visible what used to hide behind typing speed and surface productivity. The skill that survives is thinking. Staying useful means thinking well.

The hiring bar is going up. Not because you need to know how to prompt. Because AI removes the buffer that hid insufficient depth. Now it shows up earlier.

For me, that's good news. I'll hire fewer people. But each one — much better.