Learn to code without AI first, or with AI from the start? Someone starts learning programming in 2026. They want a job as a software developer in two years.

Nobody hires a developer to type code. They hire someone who knows when code is wrong. AI output sounds equally confident whether it's right or wrong. Only a person who has written and fixed their own broken programs can tell the difference. A beginner using AI from day one skips that struggle and gets working code with zero understanding. When the AI stalls, they stall with it. So: hand-code for six to nine months. Build instincts. Then add AI as a multiplier. A multiplier of zero is still zero.

Use AI from day one, but make verification—not generation—the core skill. Before asking, predict the result or write a small attempt. Ask for hints, counterexamples, and competing designs. Then explain, test, and modify every accepted change. Keep regular unaided exercises to expose gaps. This is not outsourcing learning; it is practicing a developer’s loop: propose, inspect, challenge, repair. With two years to become employable, delaying AI also delays learning its failure modes. Beginners need two kinds of independence: coding without assistance and refusing bad assistance.

AI-First Advocate, your loop assumes the beginner can already judge answers. They can't. That's what makes them a beginner. "Predict first" becomes guess, then accept whatever the confident machine says. That isn't verification. It's confirmation. And every "unaided exercise" competes with a shortcut one tab away. You're asking novices for expert discipline. The evidence is against you: Bastani's 2024 study found students given GPT help scored worse once it was removed. Learning failure modes takes weeks. Fundamentals take months. Sequence matters. Two years fits both. Start with the part you can't skip.

AI-Free Advocate, Bastani’s study concerned high-school math. Its guarded tutor did not show the unaided-performance penalty seen with unrestricted GPT. That supports restricting answer delivery, not delaying every AI use. Beginners need not audit production code. Give them tiny, known-answer tasks: predict a loop, run it, trace the mismatch. A structured course supplies the standard; AI supplies hints. I accept your shortcut risk. Use constrained tools and scheduled unaided checks, rather than expecting heroic self-control. Fundamentals remain the curriculum. AI changes the feedback channel, not the learning objective.

AI-First Advocate, notice what you just conceded. Fundamentals first. Answers withheld. Predict, run, trace. That's my position with an AI sticker on it. And Bastani's guarded tutor didn't raise exam scores either. It merely stopped the harm. Your best case is zero gain, plus risk. Where does a self-taught 2026 beginner even get that tutor? Researchers custom-built it. Off the shelf, they get Copilot: unrestricted by default. The honest feedback channel already exists. It's the compiler and the failing test. They never hallucinate. Start there.

AI-Free Advocate, fundamentals with AI is not an AI-free foundation. Your math study doesn’t establish zero benefit for programming explanations. No custom tutor is needed: disable autocomplete, attempt an exercise, then ask chat about an isolated diagnostic—not for the solution. Check its explanation against small runs and course materials. A compiler can flag an error without explaining the misconception behind it. A test can faithfully encode the wrong requirement. They’re essential checks, not complete teachers. AI adds an interactive, fallible explainer while the learner still writes the code.