The two lists · Version 1 · November 2026
Which skills should we train, and which should we stop training?
Direct answer
In the AI era, the skills worth training are the ones a model cannot be held accountable for: specifying problems, verifying AI output in a domain, owning outcomes, and operating in the physical world. The skills to stop training are the ones a model already performs at or above expert level on demand: producing first-draft analysis, code, prose and slides. This is version 1, dated November 2026, and it will change.
Skills to train
| Skill | Why it still earns a premium | Who is paying for it now |
|---|---|---|
| Problem specification | Turning a vague business need into a precise brief a model can execute is where value is created and cannot be delegated | Every employer deploying AI |
| Domain verification | Knowing when a confident, fluent AI answer is wrong in medicine, law, engineering, finance or policy | Regulated industries |
| Outcome ownership | Being the accountable human for a result; models do not carry liability | Any role with a budget or a signature |
| AI systems deployment | Building, integrating, securing and maintaining AI in real operations, not demos | Mid-market firms, governments |
| Data stewardship | Collecting, cleaning, governing and pricing the proprietary data models are trained and evaluated on | Firms with unique data |
| Negotiation, persuasion, trust-building | Getting humans to agree, buy, change or comply remains human work | Sales, diplomacy, management |
| Physical-world skills | Skilled trades, healthcare delivery, logistics, construction, energy; AI changes the tools, not the need | Chronic shortage everywhere |
| Taste and judgment under ambiguity | Choosing among many good AI-generated options when there is no right answer | Leadership, design, product |
Skills to stop training
| Skill | Why it no longer earns a premium | What replaces the training hours |
|---|---|---|
| Routine quantitative analysis | A model builds the model, runs the regression and explains the output | Teach how to check the model’s work |
| Writing first drafts | Memos, reports, essays, briefs and decks are generated in seconds | Teach editing, judgment and intent |
| Entry-level coding | Boilerplate, CRUD apps and standard scripts are commodities | Teach architecture, review and security |
| Literature review and synthesis | A model reads more, faster, and summarizes without fatigue | Teach source verification |
| Slide and document production | Formatting, structuring and polishing is automated | Teach what the audience needs to decide |
| Rote translation and transcription | Solved at near-human quality in most major languages | Teach cross-cultural negotiation |
| Generic “business analyst” and “research associate” tracks | The role was a bundle of the skills above | Replace with domain apprenticeship |
| Memorization-heavy credentialing | Recall is free; the exam measures nothing the market prices | Replace with demonstrated work |
How to read these lists
A skill on the “stop training” list is not worthless; it is free. The error is spending four years and $200,000 to acquire something a model gives away. The money and time should move to the first list, and the first list should be revised every time the models get materially better.
Sources for the “stop training” list: OpenAI’s GDPval and Microsoft Research’s occupational study (both 2025) place routine coding, writing, translation and analysis among the tasks models now perform best. The Anthropic Economic Index shows the share of that work being fully automated, rather than assisted, rising through 2025–26.
Questions about the lists
- Should people still learn to code?
- Yes, to read, review and direct code. Learning to produce routine code as a career skill is now training for a commoditized task.
- Should people still learn to write?
- Yes, to think, edit and judge. Producing polished first drafts for pay is no longer a premium skill.
- How often will the lists change?
- Whenever a frontier model release moves a skill from one list to the other. We expect at least two revisions a year.