Why this exists, what it is, and how to read it.
Why I built it
The literature of artificial intelligence is now larger than any one person can hold, and the maps we have are mostly selling something. Affiliate lists, vendor reading guides, threads ranked by who shouted loudest. Each asks you to trust the curator. Almost none let you check the curator's work.
I build governance for a living. My one rule is that governance is not what you claim, it is what you can prove. A field this consequential deserves a reference work held to the same standard: a library whose every judgment can be inspected, questioned, and overturned with evidence. So I built the thing I wanted to exist and could not find. Not to own it. To give it.
This is a public good. It is free, it always will be, and nothing in it is for sale.
What I built
A reference library of the texts that define artificial intelligence: the books, the papers, and the standards that shaped how we think, build, and govern it. Alongside the texts, a curated record of the people, organizations, and platforms that form the field's context.
The rankings get the attention, but they are not the product. The library is the product. A ranking is one view of it, produced by a published method at a fixed date. The method is the part that matters. Anyone can assert that a book is important. The Canon shows its reasoning, names its sources, and invites you to prove it wrong.
So the Canon is a reference work first and a ranking second. It answers why a text matters, what underpins it, and how it relates to the rest, before it answers which one is first.
How it is constructed
Two kinds of thing live here, and they are treated differently on purpose.
Texts are scored. Books, papers, reports, and standards are ranked by evidence: citations, library holdings, syllabus adoption, sustained readership over time. The scoring is deterministic, which means the same inputs always produce the same result, and anyone holding the audit file can rebuild a ranking exactly. Every number carries its source, its date, and its confidence. Where evidence is missing, that absence is recorded and counted, never guessed. A standard is never ranked against a novel; each kind of text is judged within its own domain, because comparing them would be meaningless.
People are not scored. The voices, organizations, and platforms are described and categorized, never ranked. There is no score field for a person anywhere in the system. This is built into the foundation, not added as a courtesy. I will rank texts. I will not rank human beings, and the structure makes it impossible to start.
The method is open. The corpus, the ontology, the weights, and the audit files are all public. The strength of this project is not a secret formula. It is the labour of careful curation, the discipline of showing the work, and the growing record of challenges met in the open. Those cannot be copied by reading the method. They can only be earned by doing the work.
A word on honesty about coverage. The Canon is strong in English. Its multilingual layer is still in development, and its Chinese-language spine is a known gap. Until that gap is closed, this is not a worldwide canon, and it will not call itself one. I would rather state the limit plainly than claim a completeness I have not earned.
How it is maintained
By a person, with help from machines that are not allowed to decide anything.
I adjudicate every challenge myself, against the evidence, and publish the resolution. I write the descriptions. Automated tools harvest the data, watch for changes, and propose updates, but no ranking ever moves without a human reviewing and approving it. The tools observe, assess, and escalate. They do not act. That distinction is the whole point of a canon you can trust rather than an algorithm you must accept.
Every change is logged and nothing is ever quietly altered. Corrections become new versions; the old ones remain visible. I do not promise a schedule, because a schedule I would resent is a promise I would break. I promise the discipline instead: every update logged, every correction traceable, every ranking open to challenge, and no commercial influence of any kind, ever. The changelog is the proof. It cannot lie about whether the work is being done.
Sources and acknowledgments
The candidate corpus was seeded, with gratitude, from curated reading lists: Jurgen Appelo's AI reading list, the Monett Critical-AI reading list (14th edition), and the project's own candidate hunt. Seeding from a list is candidacy for scoring, not an endorsement by those curators, and it carries no weight in any ranking. The full per-work source is kept in the downloadable corpus data for transparency.
Who it is for
One library, many doors. A student opens it to find what to learn and in what order. A professor opens it to find what to teach. A journalist opens it to find who and what to research, and the record of who got things wrong. An investor opens it to read the landscape. An author or editor opens it to find the works that endure. A practitioner opens it to find the standards that govern their work.
It is not built to flatter any of them, and not built to convert anyone into a customer, because there is nothing to buy. It is built to be useful and honest to all of them at once, the way a good library is.
How it should be interpreted
A rank is not a verdict on intrinsic worth. It is a transparent output of declared evidence, weights, and missing-data rules, at a specific release date. Read it as exactly that, and no more.
A high rank means a text scored well on the evidence the method measured, under one set of weights, on one day. Change the weights and the order changes, which is why three different weightings are published rather than one. A low rank, or an absence, is not a judgment that a work is bad. It may mean the evidence is thin, the work is recent, or the method does not yet see it well. Inclusion in the seed corpus means a work is a candidate, not that it is canonical.
If you disagree, you are not a nuisance. You are the mechanism. The Canon improves by being contested, and a challenge you can file is worth more than a ranking you are asked to believe.
What you can do with it
If you are here to learn AI, start with the library, not the ranking. Pick the domain you care about, governance or deep learning or the economics of automation, and read the texts that have endured there, in the order their descriptions suggest. Use the rankings as a second opinion, not a syllabus. Follow the connections: a book leads to the papers under it, the papers to the people who wrote them, the people to the institutions they work in. The structure is a map of how the field thinks, and you can walk it. Download the whole corpus as open data and keep it. It is yours.
If you find yourself here as a subject, an author whose book is listed or a voice in the field, two things are true. You are not being ranked against your peers. The people on this list are described, never scored, and never placed above or below one another. And your entry is built from public, professional, bibliographic facts, held to the same standard as everything else. If something about your entry is wrong, a misattributed work, an outdated affiliation, a description that misses the mark, tell me, and it will be corrected in the open through the same challenge process as any other claim. If you believe a text's rank or its absence is unjust, contest it with evidence, and the resolution will be published whether it goes your way or not. What you cannot do is buy your way up, or have a rival quietly removed, because no one can, and that is precisely what makes your presence here mean something.
Nothing is for sale. Nothing is hidden. Nothing is final. Challenge anything: office@apparens.nl