The SuperSkills Canon

This is a map of the field, not a reading list and not a defence. 84 works on what increasingly capable AI does to human capability, judgement, learning, expertise, work and agency, each classified by the role it plays, each with a one-line finding, an honest note on how strong the evidence actually is, and a short reading of why it matters. Some of it supports the SuperSkills argument. Some of it complicates the argument. At least one entry contradicts it directly, and is marked as doing so.

How this is classified

Every entry carries one of five classifications. They describe the role a work plays, which is a different question from how strong its evidence is, and both are stated separately because collapsing them is how bad citations happen.

• Canonical (24). Sources anyone entering this field should know. Roughly twenty-five of them.
• Strong evidence (39). Peer-reviewed work, large empirical datasets, or institutional analysis credible enough to build on.
• Important perspective (5). A serious argument that shapes the debate. Not empirical proof, and not treated as such here.
• Foundation (14). Older work on cognition, expertise, automation, judgement and skill acquisition. Most of this predates AI and explains it better than most writing about AI.
• Watch (2). A living programme or dataset whose conclusions will change. Track it rather than cite it once.

Every external link has been fetched and confirmed. Where a statistic is quoted, its base is given: the BCG deskilling finding rests on seventy executives, and saying so is more useful than the headline. Where a work has been superseded, the earlier edition is kept as a longitudinal comparator rather than replaced, because what the field predicted and what happened is itself evidence.

Browse

Filter by classification or by theme. Everything is on this page, so a search within it will find anything the filters do not.

What the Canon is for

There is no shortage of AI reading lists. What is missing is a map that separates the four kinds of authority this field actually runs on, and says which is which. Hard data, telling you what has been measured. Experimental evidence, telling you what happens under controlled conditions. Serious interpretation, telling you what thoughtful people make of it. And intellectual foundations, mostly written long before AI, explaining cognition, expertise, automation and skill in ways that current writing rediscovers without attribution.

That last category is the one most often missing, and the most valuable. Lisanne Bainbridge described the core problem in 1983, writing about process control: automation removes the routine work through which operators stayed practised, and then depends on those operators for the hard cases. Michael Polanyi explained in 1966 why expertise cannot be fully written down, which is why it cannot be fully written into a prompt. Almost everything published about AI and deskilling in the last three years is a restatement of one of those two arguments.

How to use it

If you are writing about this field, the twenty-five canonical entries are the ones to have read. If you are checking a claim, find the entry and read the evidence-strength line before quoting the finding. If you are looking for the counter-argument, the important perspectives are where it lives, and Acemoglu and Johnson challenge an assumption this site relies on rather than one it opposes.

Every entry has a permanent anchor and can be cited on its own. The whole Canon is published as structured data at canon.json under a Creative Commons Attribution licence, so it can be read by a machine rather than scraped from prose. Where an entry also appears in the graded evidence base, its entry there is linked.

What is missing

A great deal. This is a first edition and it is not yet a hundred works. It is under-weighted outside the English-speaking world, thin on education systems beyond higher education, and lighter than it should be on the sociology of work. Anything included has been read and verified rather than listed, which is the constraint on how fast it grows. Suggestions for what belongs here, particularly work that challenges the argument, are genuinely welcome.

Related SuperSkills research

The graded, method-first version of the empirical material is the evidence base. The interpretation lives across the research: AI and human judgement, how humans learn with AI, human and AI decision making, capability debt, what stays human and staying valuable in the age of AI. What the evidence establishes, and how strongly, is in what we actually know about AI and human capability, and the full map of the territory is in 500 Questions About Humans and AI.

About this reference

Compiled and maintained by Rahim Hirji, author of SuperSkills: The Seven Human Skills for the Age of AI (Kogan Page, 2026) and founder of The SuperSkills Intelligence Company. Every URL was fetched and confirmed before publication; where a publisher page renders empty to automated checks, a DOI or an institutional copy is used instead. Classifications and readings are the author's and are open to argument. Entries are revised when a work is corrected, superseded or shown to have been mischaracterised here.