In March 2023, Anthropic posted a job listing for a “prompt engineer and librarian” with a salary range of $175,000 to $335,000. The role required no traditional engineering degree. The candidate needed a “hacker spirit” and a talent for coaxing useful answers from large language models.
It didn’t last. Job searches for the title on Indeed spiked from 2 per million U.S. searches in January 2023 to 144 per million by April 2023, then fell back to 20 to 30 per million, according the Wall Street Journal. That full cycle — from gold rush to near-obsolescence — took about 24 months.
The story of prompt engineering is also the story of AI slop, vibe coding, and tokenmaxxing. Each arrived as a phenomenon, peaked, and left behind a smaller, more durable practice. The pattern is worth understanding, because the next wave is already here.
The democratization of AI and its side effects
Prompt engineering entered public consciousness as a genuine career category in early 2023, weeks after ChatGPT’s release made large language models a household curiosity. Companies treated the ability to write effective instructions for AI as a scarce and valuable skill. Median salaries at top firms reached $296,000 at Meta and $279,000 at Google, according to People in AI. A 2023 McKinsey survey found that only 7% of organizations adopting AI had hired prompt engineers.
The decline was just as rapid. A Microsoft-commissioned survey of 31,000 workers across 31 countries ranked prompt engineer second to last among new roles companies planned to add. Job searches on Indeed told the same story. As models got better at interpreting plain instructions, the specialized skill lost its value. It got absorbed into existing roles, becoming a single bullet point among 20 on a job description.
As prompting became democratized, the volume of AI-generated content became its own problem. Merriam-Webster named “slop” its 2025 Word of the Year, defining it as digital content of low quality produced in quantity by AI. And the scale was measurable. A Graphite study of more than 65,000 English-language articles found that AI-generated content exceeded human-written articles on the web by November 2024. A separate study estimated that by mid-2025, about 35% of newly published websites were AI-generated or AI-assisted.
The cultural backlash was swift. An em dash in casual writing could get someone accused of using ChatGPT. But the economic correction was more telling. Despite the flood, Graphite found that 86% of articles ranking in Google Search were still human-written. The slop was voluminous but invisible to the distribution channels that mattered.
The corporate reckoning with AI’s real limits
In February 2025, Andrej Karpathy, a co-founder of OpenAI and former AI leader at Tesla, posted on X that he had discovered a new way of working he called “vibe coding” — fully surrendering to AI suggestions, accepting generated code without reviewing it, and working around bugs by asking for random changes until they disappeared. He acknowledged it was fine for throwaway weekend projects.
The corporate world didn’t treat it as a weekend hobby. Security problems followed. A Veracode report testing more than 100 AI models across 80 coding tasks found that 45% of AI-generated code samples failed basic security tests. Java had a 72% failure rate. The Cloud Security Alliance reported that 62% of AI-generated code contained design flaws or known vulnerabilities. A Georgia Tech tracker confirmed 74 AI-linked security vulnerabilities through March 2026, with researchers estimating the true number at five to 10 times higher.
By early 2026, the metric of choice had shifted from what AI could produce to how much of it was consumed. Meta built an internal leaderboard ranking employees by token consumption, with titles such as “Token Legend.” Amazon had its own informal version. Sequoia Capital partner Sonya Huang told the Wall Street Journal that every founder she advises should adopt the same mindset.
The correction came fast. Uber burned through its annual AI budget in three months and capped use at $1,500 per developer per month. GitHub Copilot switched to usage-based billing on June 1, and users reported bills jumping from about $50 to $3,000. An NBER working paper tracking more than 100,000 GitHub developers found that AI coding tools increased lines of code by 741% but actual software releases rose by only 20%.
Each fad traced here fits the same arc. A new capability produced a wave of inflated expectations, followed by a correction that left behind a smaller, more durable practice. Gartner’s 2025 Hype Cycle for AI has a name for the stage where that correction happens: the trough of disillusionment.
Prompt engineering became a skill, not a career. AI slop forced platforms to filter harder. Vibe coding exposed the gap between generating code and shipping software. Tokenmaxxing revealed that consumption metrics measure activity, not output.
Anthropic’s projected $10.9 billion in second-quarter revenue — a 130% increase over the prior quarter and its first operating profit — shows who benefits from the frenzy while it lasts. The company has signaled to investors that profitability may not last the full year as infrastructure costs rise. The party’s economic footprint is real. The question each wave leaves behind is the same: what remains when the leaderboard comes down?
