Where AI is used, and where it is not
The record is made without a language model: its links come from the sources' own records, read by written rules. Anthropic's Claude is used in bounded places beside it: the assistant that answers questions about the sheet in front of you, the reading of a supervisor from a thesis's own text, a few likely teachers read with its help, and the search for leads above the top of a line. What a model reads is kept as a kind of its own, comes in through a check, or stays outside as a lead that is never drawn.
What a language model is
A large language model is a program trained on a very large body of text to continue text: given the words so far, it gives the words most likely to follow. Trained at that scale, it learns to read, summarise, translate and answer questions in fluent prose. Claude, made by Anthropic, is the family of models this site uses; OpenAI's GPT models and Google's Gemini are others of the kind.
What it does well follows from that. It reads a page and says who is named on it and in what role, it puts a long answer into a few sentences, and it understands a question asked in a reader's own words. What it does badly follows too. It does not look anything up unless it is given it: what it knows of the world is what its training left in it, unevenly and without sources. It does not count reliably, and it does not remember exactly. And asked about something it was not shown, it can still answer, fluently and wrongly. This is called hallucination; more plainly, invention.
For a record whose every link must say where it came from, that second list settles where a model may be used. A model may read, and it may answer a question about what it was given. It is never the source of a link.
Retrieval-augmented generation
The common way to hold a model's answers to what is known is retrieval-augmented generation, RAG for short. Before the model is asked, the documents that hold the answer are found and put in front of it, and it is told to answer from them.
In its usual form it works in four steps. First, ahead of time, a collection of documents is cut into passages, and each passage is turned into an embedding: a long list of numbers, made by a model of its own, that places passages of like meaning near one another. The embeddings are kept in a vector index, a database built to find quickly the lists of numbers nearest a given one. Second, when a question comes, it is turned into an embedding the same way, and the index gives back the passages nearest it. Third, those passages are put into the model's context, the text it is given to read with the question. Fourth, the model writes its answer from them, often naming the passage each statement rests on.
It reduces invention because the model has the facts in front of it rather than half remembered, and because a reader can follow an answer back to its passage. It still fails in known ways. The nearest passages are those nearest in meaning, not those that hold the answer: two people of one name, or two theses on one subject, sit close together, and the wrong one may come back. The passage that holds the answer may not be among the few retrieved at all. A question that needs every passage of a kind, such as how many students someone supervised, is answered from the handful that were retrieved. And the model may still write beyond what it was given.
The assistant beside the sheet
Every sheet on this site, the drawing of a person's advisors, students and the people beside them, has an ask tab where a reader can ask about what is drawn: who taught this person, which student carries the deepest line, where the record stops. The answers come from Claude Opus 5, Anthropic's model, asked through Anthropic's own service. The idea is RAG's, a model answering from what it is given rather than from memory, but the retrieval is of another kind.
Nothing is searched for by meaning. The sheet in front of the reader was drawn from the record by an exact walk: from the person's id, their advisors and students are followed link by link in the record's PostgreSQL database, and beside them is whatever the reader asked to see, such as coauthors or examiners. When a question is asked, that sheet is what is sent: the people drawn, nearest the person first, each with the counts the record holds for them; the doctoral links among them; and every other kind of link (a master's thesis, a coauthor, an examiner, a teacher) in a list of its own, so that none can be read as supervision. Where a large sheet was folded, the folds are said, so the model knows a line goes on past what is listed.
With the sheet go the rules the model is bound to. They say what each list means and, among much else: use only the sheet; never invent a person, a year or a count; if the sheet does not hold the answer, say what is missing; a coauthor, a teacher or an examiner is never an advisor or a student, and never counted as one; and where someone has no recorded students, the open ring that marks them is the end of the record, not evidence of no students. The answer is a few plain sentences, then a line naming the people it spoke of.
The page reads that line against the sheet. Only people on the sheet are named under the answer, by the names the record gives them, each opening their card, and the first is lit on the figure and brought to the middle. A name the model wrote that is not on the sheet lights nothing.
A conversation keeps its recent turns, which go back with each new question. The sheet, the same from one question to the next, is held in Anthropic's prompt cache: the part of a request that repeats is kept by Anthropic for a few minutes, so a second question on the same sheet is read from the cache rather than taken in afresh, which is quicker and costs less. The question is sent to Anthropic to be answered and is not kept by the site, as the privacy notice says. Each question costs the project money, so the assistant answers 50 questions a day across the whole site, counted to midnight UTC; past that it says so, and it is back the next day.
Why an exact walk suits this record
A search by meaning is the right tool where the answer lies somewhere in a heap of prose and nobody knows where. The record is not prose. It is a graph: people and theses are its things, supervision and the other roles are the links between them, and every person has an id. A question about a sheet is a question about known people, and the way to find what the record holds of them is to follow their links exactly. That is what a database does well, and what a vector index does not do at all.
Three things follow. Nothing near is taken for the thing itself: a namesake one department over is another person with another id, and is on the sheet only if a link puts them there (who is who says how people are told apart). Nothing is left out by chance: every link the record holds among the people drawn is sent, not the few that happened to come back from a search. And the counts are the record's own, carried on each person, so the model reads a count rather than makes one.
What it cannot do follows as plainly. The assistant sees the sheet and nothing else: not the rest of the record, not the web, not a source the site does not read. It cannot change the record or add to it. And it is still a language model, which can misread what is in front of it, so an answer is a reading of the sheet, and the sheet, with the marks on every link, is what an answer is to be held against.
Reading a supervisor from a thesis's own text
Many thesis records name no supervisor, though the thesis itself does, on its title page or in its acknowledgements. Where a thesis's file is open to read, its text can be given to a model to read for the people it names. The site does this in one bounded way.
The text of a thesis's first pages, where the title page, any approval page and the acknowledgements are, is given to Claude with one task: list the people named, the role each holds, and the line the role was read from. That line must appear in the text word for word. A role whose line is not there is dropped, whatever the model said: this test is what stands between a model that invents a supervisor and the record. Only a supervisor or a co-supervisor becomes a link. The model is asked through Anthropic's own service or through Amazon Bedrock, AWS's service for hosted models; the links of this kind the record holds now were read by Claude Sonnet 5.
Before any link was kept, the reading was tried on theses whose repository already names the supervisor, those of DiVA, the repository Swedish universities share, and its answers were compared with the record's own; only once it passed was it used on theses whose record names none (measured, not asserted says what passing means). A link read this way is a kind of evidence of its own, extracted, and is never folded into the links a repository's record cites (where every link comes from). A person's page says such a link was found in the text of the thesis, and the card shows the quoted line it rests on.
So far the record holds links of this kind for master's and bachelor's theses of one university, LUT University in Finland. Elsewhere a supervisor named on a title page is read by a rule written for that repository's pages, not by a model: the words that head the supervisor's name on a Czech title page, say, are read the same way every time.
Likely teachers, read with a model's help
Some people, physicians among them, trained under someone who supervised no thesis of theirs, and no record names that person. A rule can propose one: the person written last on the first papers someone wrote, often the head of the laboratory they trained in. On its own that rule is right too seldom to be drawn.
So each pair it proposes is read one at a time by the project, with Claude's help, against the papers' bylines and affiliations and the grants behind them, and called likely or not. A pair called likely is drawn as a likely teacher, not confirmed, with what the reading rests on beside it. It is counted nowhere and never stands as an advisor, as the privacy notice also says.
Leads above the top of a line
Every line in the record has a top: a person who supervised a doctorate, and of whom neither a repository's record nor Wikidata names an advisor. Who trained them is often written down somewhere: in a thesis no source the site reads holds, a CV, an obituary, a biography, the Mathematics Genealogy Project.
To find such places the project runs a search with two models in two roles. Claude Haiku, Anthropic's smallest and quickest model, works through the people at the tops of lines, those with the most doctoral students first: what the record already holds of them, then Wikidata, then the web, reading any page under the site's own name and at the pace its host asks, as every read the site makes is. Each claim it makes is kept with the page and the words on that page it rests on, and whether those words are really there. Then Claude Sonnet, a larger model, reads each quoted claim and tries to refute it.
What survives is a lead, kept for a person to read. It is never drawn, never counted and never used to join two people, and it has no kind of evidence of its own: until the project names one, a lead stays outside the record.
Where a model reads in a person's place
A reading of a source, or a rule that joins two people, enters the record only once it has passed a check: a sample read against the source's own pages, a score, and a bar to clear (measured, not asserted). Some samples are read by a script, some by hand, and many, on the owner's word, by Claude standing in for a person, with the source's records open beside it. Each check says which, and a sample a model read is kept, so a person can read it again.
One join was let in by such readings pair by pair. The rules that put an OpenAlex author record beside a Wikidata person did not pass their check, so none of their joins was taken on the strength of a sample. Instead every pair they proposed was read by Claude, each record's works against the person's own entry. A pair it called the same person is shown as a checked match; one it was unsure of, as a likely match, not confirmed; one it called wrong is left out.
Models also help with work that writes nothing to the record. Claude Haiku looks for the address at which a repository the site does not yet read serves its records, which the site then asks itself, under its own name, before the repository is added to its list of sources. And models grade a reader still being written, as a measure of whether a change to it made it better.
What models are not used for
The record itself is made without one. Every repository's records are read by code written for their format and their source's habits, the same way every night (from a repository record to a link). People are told apart and joined by written rules and shared identifiers (who is who); a model has read pairs a rule proposed, as above, but has never proposed one. A thesis's level is read from its record's own words. Every count the site shows is counted from the record by code. And no link is written on a model's word alone: a supervision is cited because a record names it, stated because Wikidata states it, and extracted only through the reading above, its line found word for word in the thesis.
Nor does a model stand between the record and what a reader sees. The search, the sheet, a person's page and the routes between two people are drawn from the record by the same exact walks, and no model chooses what is shown or in what order.