Generative AI is not replacing jobs or hurting wages at all, say economists

Original Article ↗ Hacker News Discussion ↗

Study scope & limitations

  • Study covers Denmark, late 2023–24, 11 “exposed” occupations, and only chat-based tools; commenters argue this misses:
    • Other gen-AI (images, video, music) and agent-like workflows.
    • Workers already laid off or never hired.
    • Countries with weaker labor protections and more offshoring.
  • Many see it as “too early” data on deprecated models; analogous to judging the internet in 1995.

Visible job impacts (anecdotal but widespread)

  • Freelance copywriters, illustrators, translators, technical writers, VFX artists, stock-image–style designers report work “drying up” or rates collapsing as clients use image/text generators for “good enough” output.
  • Customer support: more full-AI front lines, fewer level‑1/junior roles; quality often worse, but cheaper than humans or offshore centers.
  • Some professionals (tax prep, diet advice, basic legal/tax questions, simple research) are skipped entirely when people use LLMs plus cited sources instead of paying experts.
  • Several report not hiring junior researchers/assistants specifically because ChatGPT/Gemini/Claude fill that gap.

Productivity vs wages and headcount

  • Individual engineers, writers, and knowledge workers report 5–10x subjective productivity boosts on certain tasks; others see ~5% real savings and lots of cleanup/verification.
  • Gains often go into “more work” (features, docs, meetings), not fewer hours or higher pay, so macro measures (wages, hours) stay flat.
  • Automation frequently creates new downstream tasks and even extra busywork (e.g., AI-generated resumes vs AI filters), so net labor savings are unclear.

Corporate incentives, hype, and data extraction

  • Boards and investors pressure executives to “do something with AI,” short‑circuiting normal cost–benefit checks; many AI features are seen as marketing theater.
  • Strong suspicion that aggressive “AI in every product” pushes are about collecting proprietary usage data and building “data moats” rather than immediate productivity.
  • Parallel drawn to dot‑com: massive capex may leave useful infrastructure—or may prove a misallocation if LLMs fail to generate commensurate returns.

Capabilities, limits, and appropriate roles

  • LLMs are praised as:
    • Better search/“bureaucracy navigators” (tax, DMV, forms).
    • Strong drafting and rewriting tools (emails, reports, code stubs, unit tests, summaries).
  • But they are criticized as:
    • Fundamentally non‑deterministic and unreliable for machine‑to‑machine or safety‑critical tasks.
    • Prone to “bullshit” rather than mere random error; still hallucinate citations and legal/tax details.
    • Unable to fully replace accountability, judgment, or domain-specific responsibility (e.g., lawyers, doctors, senior engineers).

Hiring dynamics and entry‑level erosion

  • Several commenters say they now:
    • Hire fewer interns/juniors, or delay new headcount, because AI plus seniors covers more work.
    • Use AI instead of contract developers, editors, or short‑term specialists.
  • This shows up as “jobs that never materialize,” which a wage/hours snapshot in Denmark won’t detect.

Macro outlook and historical analogies

  • Optimistic camp: history of automation (looms, PCs, internet) shows technology reallocates labor rather than causing mass unemployment; demand expands (more software, infrastructure, personalized services).
  • Skeptical camp: gen-AI may hollow out whole sectors (e.g., resume pipelines, low‑end creative industries) faster than new roles appear, especially for middlemen and entry-level “ladder” jobs; wealth and power may concentrate with AI capital owners.
  • Some argue a crash in AI valuations is likely (similar to NFTs/early crypto), even if the underlying tech remains and slowly transforms workflows over the next decade.

Customer experience, search, and “enshittification”

  • AI chat support is widely seen as degrading user experience while reducing costs; companies may accept this tradeoff if churn is manageable.
  • Strong debate over whether Google Search is “in decline” under SEO and AI slop vs still “working wonderfully” and earning record ad revenue.
  • Many expect LLM interfaces themselves to be enshittified by ads and sponsorship once growth slows, recreating search’s trajectory.