AI's earnings gap is a people problem

A developer was fired after AI-generated code crashed a production system. The manager who approved the code had also used AI to review it ^[1]. The code was written by a machine and checked by a machine, and the only human in the loop was the one who got fired. A quality process with no human judgment in any link — and the oddest detail is that no one involved seems to have noticed the loop was empty. The layoffs meant to pay for all this are now the single largest reason companies give for cutting jobs. AI was cited in 87,714 US layoffs in the first five months of 2026 alone, more than all of 2024 and 2025 combined ^[2]. WPP cut 9,000 roles, Omnicom and IPG 8,200, and McKinsey 3,000 to 4,000 while deploying 20,000 AI agents ^[3]. PwC is trimming UK graduate hiring and cutting US entry-level roles by nearly a third over three years ^[4]. The costs being cut are the ones a spreadsheet can see. What the spreadsheet can't see is what those entry-level jobs were producing. Stanford researchers found AI hasn't caused economy-wide displacement, but it has widened the entry-level employment gap from 15% to 19% — and the widening came from companies hiring fewer people, not firing more ^[5]. The pipeline that turns early-career struggle into professional judgment is being cut at the source. Researchers at the Work AI Institute call what's being lost cognitive debt — an illusion of expertise in which workers feel more productive while the underlying skill atrophies, because automating first drafts removes the struggle that builds professional judgment ^[6]. The loss is being masked by a competency bubble. Nearly 20% of US LinkedIn users have retroactively edited past job titles to align with current demands, with a sixfold rise in AI-related terms since ChatGPT ^[7]. AI mentions on resumes jumped from 4% in 2023 to 13% ^[8]. And the evaluators can't tell the difference — AI assessment models used to screen resumes and papers systematically prefer AI-generated content, rewarding stylistic patterns that mirror their own training data ^[9]. The loop closes: AI writes the credential, AI reads the credential, and the human judgment that should sit between them is the very thing being cut. The earnings gap follows. Goldman Sachs found only 2% of S&P 500 companies have quantified AI's effect on productivity, and among those that did, there was no statistically significant difference in growth versus peers ^[10]. Datatonic puts the failure rate of AI pilots at 95%, attributing it to productivity leaking away when AI sits apart from the people who actually run the business ^[11]. Even the people selling the tools won't hand over their own work. Sam Altman still manages his email manually despite having Codex available, and he can't quite explain why.

And yet there's something in my mind that is encoded that doing this kind of stuff is what it means to work and what it means to be productive. — Sam Altman

The platform that hosts the competency bubble is now fighting its own output: LinkedIn's anti-slop campaign cut views of AI-generated content by 40% ^[12]. The seller won't delegate, and the host has to police what its own tools produce — two quiet admissions that the judgment problem is real, and that no one has a machine fix for it. The loop is recursive: AI generates the work, AI evaluates the work, and the humans who should catch the difference are losing the capacity to. Sixty-two percent of Gen Z workers admit to over-relying on AI, and 40% say they can't function without it ^[13]. But the diagnosis has a counter-example, and it's the quietest finding in the record. Publicis grew revenue 5.6% by training 85% of its staff on AI while cutting only 200 jobs ^[3]. When the investment goes into judgment instead of into cutting it, the gap closes. The problem was never the model.