AI Commoditizes Expertise

Dan Shipper's mechanism for why automation increases expert human work: LLMs cheapen yesterday's competence, abundance breeds sameness, and sameness creates demand for the one thing models can't supply — human difference.

The mechanism (the arc)

1. AI makes yesterday's competence cheap. Models train on "the visible residue of human competence" — code, prose, images, support tickets, product specs: the exhaust of successfully completed tasks — and package it cheaply for anyone.
2. Cheap competence gets adopted fast. When a once-rare skill (a pull request, a YouTube thumbnail, a newsletter) drops in cost, supply explodes. At Every: ops and support people issue PRs; marketers make thumbnails; engineers draft articles.
3. Abundance creates sameness. Because everyone uses the same models trained on the same corpus, default output converges → Slop ("visible sameness repeated ad nauseam").
4. Sameness creates demand for difference. Humans spot the sameness and want better. Demand for difference is new demand for experts.
5. Only humans supply difference. "The current generation of models only knows about work that has been done. Humans know about what needs to be done right now."

The payoff line: "Making expert work cheaper does not simply replace experts. It creates more situations where expert judgment is needed." Experts then move in two directions: build systems that absorb the flood of new work, or do bigger/more interesting work than was previously possible.

The "human premium" (related framing)

NLW ties the "demand for difference" to his own framework, the human premium — seven categories of value "that don't transfer when you remove the human," introduced in AIDB — The New Jobs AI Will Create (#to-read — the seven categories aren't enumerated in this source).

What "expertise" is made of

Willingham's research sharpens the "residue vs. expert" distinction. Experts aren't people holding generic thinking skills plus facts. They have practised a domain until its deep structure is perceived directly, and their critical thinking is bound to that knowledge. That is why the residue (finished artifacts) commoditizes while the expert (knowledge that sees what needs doing next) doesn't. It also means the supply of experts depends on people still building that knowledge by doing the work AI now makes optional.

If the junior work that used to build expertise is the first thing AI absorbs, where does the next generation of "difference"-supplying experts come from? Willingham's account makes this a real risk; Shipper doesn't address it.

Borrowed expertise: the terms of the bargain

Anthropic — Investigating Vault Particles with Claude is the clearest case in the wiki of AI supplying expertise its users don't have. A team with "limited experience in structural biology" solved a novel 9.5-megadalton structure in under four months, "because Claude could make up for what we didn't know."

That is Shipper's arc running in reverse of the usual reading: not an expert using AI to go faster, but non-experts reaching expert output. The bill came due as error cost — errors ran for two weeks in the areas they didn't know and were caught in minutes where they did (Verification Requires Domain Knowledge). So borrowed expertise is real but not free, and the discount depends on how much of your own you keep. It also sharpens the open question above: if non-experts can produce expert artifacts, the residue commoditizes faster than ever, and the "difference" that stays scarce is specifically the knowledge needed to tell good residue from flawed residue.

Connections

• Critical Thinking Is Domain-Bound — the cognitive account of why expert judgment can't be copied cheaply.
• Slop — the failure mode this mechanism produces by default.
• The Human Sandwich — the "judge & extend" step is where humans inject difference.
• AI and Employment — Replacement vs Expansion — this is the engine of the expansion case.
• Token Shortage — commoditized demand at planetary scale is part of why AI demand outruns supply.
• The Bottleneck Shift — Ethan Mollick reaches the same conclusion independently ("humans know what needs to be done" ≈ "the bottleneck is figuring out what to do").
• Clichés Carry Less Information — Norbert Wiener's information theory underwrites why difference is the scarce, valuable thing (cliché = low information).

Sources

• The AI Daily Brief — Why Agents Still Need Humans — quoting Dan Shipper — After Automation.
• Making AI Work — Leadership, Lab, and Crowd — Mollick's convergent "bottleneck shift" framing.
• The Human Use of Human Beings — Wiener (excerpt) — the information-theoretic root of "demand for difference."
• Willingham — Critical Thinking, Why Is It So Hard to Teach — what expertise is cognitively; why it resists commoditization.
• Anthropic — Investigating Vault Particles with Claude — borrowed expertise in practice, and what it cost.