Researchers Clash Over Whether AI Has Already Automated Programming Jobs

On November 15, AI safety researchers David Manheim and jessicata engaged in a multi-round debate on X over whether programming has already been automated by AI. The core disagreement: is the absence of employment impact due to insufficient capability, or something else?

Confirmed

• The debate started from a paper on AI's employment impact, which argued that the current bottleneck for AI coding is code review, and that programmer employment could drop sharply in the coming years if this bottleneck is removed. jessicata argued from this that code review is a significant part of a programmer's job, and as long as it remains a bottleneck, the role hasn't been automated.
• Manheim countered that AI doesn't need to replace every task of every programmer—replacing an entire business unit wholesale is enough—and drew an analogy to the computerization of spreadsheets. He also noted that in approaches like artcraft, code review isn't a necessary step, and that AI coding can support large projects—most projects are no bigger than MS Office or Photoshop, and when a full rewrite takes less than a week, technical debt stops being fatal.
• jessicata offered a firsthand data point: for a recent research project, she needed to build a seemingly simple eval harness; implementing it to spec with Astra + Codex (Sol 6.1) took about two days—worse than doing it herself. She believes AI tools suit rapid prototyping and small projects, while large projects prioritize quality.
• Manheim held his ground: the capability exists today and will keep improving, and the missing employment impact stems from factors unrelated to capability. He also criticized the paper for assuming firms differ only in adoption timing, ignoring that early adopters were using weaker systems.
• On the data question, the two ended near a draw: jessicata pointed out that labor market statistics contradict the claim that programmers have been automated—a strong point for her side; Manheim partially agreed but argued that current industry statistics carry almost no information, because they don't measure what's expected to change.
• jessicata also expressed skepticism about exponential AGI narratives, suggesting automation will advance quickly in a few concentrated areas rather than across the board.

Why it matters

This debate illustrates the complexity of the causal chain between AI capability and employment impact: even when the tools are capable, organizational adoption, quality constraints, and measurement issues can delay or mask the disruption. Their discussion of the code review bottleneck, wholesale unit replacement, and the failure of existing statistics offers concrete arguments on both sides for assessing LLMs' real impact on programmer employment.

2026-11-15 ~ 2026-11-15 · 13 related posts

• Commenter argues AGI automation is domain-specific, not a hockey-stick takeoff — jessi_cata · 2026-11-15
• 'Programmers are automated' claims clash with labor market data, argues commenter — jessi_cata · 2026-11-15
• AI automation isn't showing in labor stats because economists are measuring the wrong thing — davidmanheim · 2026-11-15
• Labor market stats don't match what we'd see if LLMs had automated programmers — jessi_cata · 2026-11-15
• Researcher slams AI-adoption jobs paper for conflating learning curves with model progress — davidmanheim · 2026-11-15
• Manheim: code review bottleneck argument crumbles as tools shift review to QA — davidmanheim · 2026-11-15
• AI Coding Debate: Code Review Is the Real Bottleneck for Programmer Jobs — jessi_cata · 2026-11-15
• Code Review Isn't Automated Yet, So Programmer Jobs Aren't Either — jessi_cata · 2026-11-15
• Artcraft shows AI coding scales: when full reimplementation takes a week, tech debt matters less — davidmanheim · 2026-11-15
• [source] Debate: AI doesn't need to replace every task, just whole business units — davidmanheim · 2026-11-15
• [source] AI researcher: employment collapse is 'baked in' to today's available tech — davidmanheim · 2026-11-15
• [source] jessicata: Coding AI Took 2 Days for a Simple Eval Harness, No Real Speedup — jessi_cata · 2026-11-15
• Insurance researcher pushes back: coding AI can't yet replace whole programmer teams — davidmanheim · 2026-11-15