# Esmond AI policy register: methods

Version 2026.11.8 (2026-11-20). Licence: CC BY 4.0, except the column `quote` (section 1). This file describes how the records in the download are made. The columns are defined in `DATA_DICTIONARY.md`; the denominator and the count per status for each country are in `coverage.csv`, the figures per country in `summary.csv`. Column names and values are English from release 2026.11.5; the Dutch names of earlier releases are listed in section 14.

## 1. What a record is

One record per higher education institution. A record says whether the institution publishes a rule on generative AI for the thesis or dissertation, what kind of rule it is, where it applies, and where to read it. The summary in `rule_summary` is ours, in one sentence, in the language of the register (`summary_language`).

For `checked` records, the column `quote` gives the passage of the source that carries the rule, word for word and in the original language, at most 300 characters. It is the passage that the check confirms in the fetched source (section 5). Where the quotation on file is longer, the column stays empty. The column quote contains short quotations from the institutions' own published rules, included under the quotation exception (Art. 5(3)(d) Directive 2001/29/EC; Belgian Code of Economic Law, Art. XI.189). They are not licensed under CC BY 4.0; rights remain with the institutions. An institution can ask for the removal of its quotation (section 13).

The register is made for students and for people who study policy. It is not legal advice, and it does not replace the regulations of the institution, the programme or the supervisor.

## 2. Which institutions

Per country we start from an official list of recognised institutions, published by a public authority or a sector body. `coverage.csv` gives the list, its source and its date for each country. Every institution on the list is in the download, also when we found nothing: the denominator of every figure is the list, not the set of institutions we happened to find.

A few records concern institutions that are not on the official list of their country (for example because they have their own page on our site). They are in the download, `on_official_list` is `false`, and are not counted in the figures per country.

## 3. Sources

We use only sources published by the institution itself on its own site: regulations, a guideline or policy, a decision of the rectorate, board or senate, a graduate school handbook. News items, library guides, course pages, forms and third-party sites are not sources for a record.

Every source is fetched through a gate that follows `robots.txt`, including rules for AI crawlers. A source that the gate refuses is never fetched by another route, never quoted and never summarised. We do not use theses or repositories of student work as a source.

## 4. Search

Each institution is searched at least once, in the language of its country. In the search round of 15 November 2026, the institutions in twelve countries that were still without an outcome were searched again with two searches each: one for an AI rule under the names used locally (guideline, declaration of originality, decision of the rector, charter), and one for the current regulations of the whole institution (examination regulations, general regulations, thesis regulations, graduate catalogue). A third search followed when there was a concrete lead. Where the search engine gave nothing, we navigated from the institution's own site to its page with regulations.

## 5. Outcomes

Every institution gets one of five outcomes in the column `status`.

| status | meaning | condition |
|---|---|---|
| `checked` | checked | The source was fetched through the gate, and the quotation that supports our summary is in the fetched text word for word. We keep the SHA-256 hash of the fetched source, or the quotation in a manifest that is checked against the source at every check. |
| `link_only` | published, not checked | The institution publishes the rule at an official URL of its own, but the text cannot be checked word for word. The record carries the link only: no summary, no quotation, no hash. |
| `no_rule_found` | no rule found | We read the current regulations of the whole institution, or of one level of it (all bachelor's programmes, the graduate school), dated 2024 or later, and the search terms for AI are not in the text. The summary names the document and its year. |
| `not_found` | not found | Researched, and no public source of the institution found that meets the conditions above. |
| `not_researched` | not yet researched | On the official list, not yet searched. |

For `link_only`, the column `reason` says why the text could not be checked:

- `robots_txt`: `robots.txt` excludes automated access, so the gate refuses;
- `http_403`: the server answers 403;
- `refused`: another refusal by the server, such as a different error status or a bot check;
- `no_text_layer`: the document is a PDF without a usable text layer (a scan).

In each case the URL must be the rule itself, on the institution's own site, and there must be evidence of that outside the unread text: the title in a search result, the name in an official list of decisions, or the file name. A record never gets the status `checked` without the word-for-word check. A `link_only` record never carries a summary or a quotation.

For `not_found`, `reason` is `no_public_source` (nothing found), `gate_refused` (the site refuses the gate and no URL is known that is itself the rule) or `not_retrievable` (a source was named but could not be fetched).

`not_found` is a researched outcome, not a gap. Every search step for such an institution is logged (the page fetched or the query run, and what it gave); `search_steps` gives the number of logged steps, and `checked_on` the date of the last search. Every `not_found` record has at least one logged step. An institution that has not been searched has the status `not_researched`. From release 2026.11.5 a new search round is recorded in `search_round`.

`not_found` and `no_rule_found` do not mean that the institution has no rule. They mean that we did not find one in a public source that meets our conditions. A faculty, programme or supervisor can have a rule that is not public.

## 6. Kind of rule (`rule_type`)

Our label for the main rule of the source. It is not the institution's own term.

| rule_type | meaning |
|---|---|
| `permitted` | The institution allows some uses of AI, with conditions. |
| `disclosure` | The student states which tools were used and what for, in a declaration or acknowledgement. |
| `consult_supervisor` | The rule asks the student to talk to the supervisor before using a tool. |
| `prior_approval` | Use in the thesis needs approval in advance. |
| `left_to_programme` | The institution leaves the decision to the programme, department or examiner. |
| `misconduct_only` | The regulations mention AI only in the clause on misconduct or plagiarism and do not say which uses are permitted. |
| `link_only` | Status `link_only`: the kind of rule was not read. |
| `none` | Status `no_rule_found`. |

Where a source fits more than one label, we choose the one that tells the student what to do first.

## 7. Scope of a rule (`scope`)

A record counts as a rule of the institution only if its source applies to the whole institution. Every record with status `checked` or `link_only` carries the column `scope`, with one of two values.

- `institution`: the source applies to the whole institution. This includes a rule of the graduate school, a regulation for one level of the whole institution (`applies_to` then names the level), the same provision in the regulations of every faculty or department when all of them were checked (`applies_to` gives the number checked), and an institution that consists of a single faculty or programme.
- `unit`: the source is a document of one faculty, department, institute or programme. The record stays in the dataset, `applies_to` names the unit, and the summary says that no rule for the whole institution was found. A graduate school for one field next to other schools counts as a unit.

Records with status `no_rule_found` do not carry the column: the regulations that were read are those of the whole institution, or of one level of it.

How the column was assigned (15 November 2026). All 153 records whose source named a single unit were examined again. For each record we first searched for a rule of the whole institution, and then checked whether the same provision appears in the documents of the other units (three first; if it appeared in each, all of them). A second reader decided per record on the fetched text. For 3 records the same provision was found in the regulations of every unit; for 18 a rule of the whole institution was found and replaced the unit rule; 12 turned out to apply to the whole institution; 120 remained `unit`. The second reader rejected 39 proposals to move a record to `institution`: a framework or position paper without a rule for students, a page of a teaching centre, a source without a readable text, a draft, or a standard sentence that appears in many but not all documents.

Where the same text was found in several but not all unit regulations, the record stays `unit`. A small number of records that support a page on our own site kept their original source although a rule of the whole institution was found later; they count as `unit`.

## 8. Second reading

The machine check confirms that the quotation of a `checked` record is in the fetched source, and that the search terms of a `no_rule_found` record are not. It does not show whether the source is the right kind of document, or whether our summary says what the source says. Every new or changed record of the search round of 15 November 2026 was therefore read by a second reader: not a sample, but all 216 records.

The second reader checked the quotation against the summary (never stricter or broader than the source), the kind of rule, the scope, and for `no_rule_found` the year in the document and whether the regulations cover the whole institution or one level of it.

Of the 216 records, 134 were confirmed without change, 44 were adjusted (mostly the scope or the fixed advice at the end of the summary), and 38 were withdrawn and returned to `not_found`. Reasons for withdrawal: a guide, FAQ or tips page without an adopted rule; a draft; an expired document or one from before 2024; a handbook of a teaching group; a link whose title and path do not name an AI rule.

### Measured accuracy

In a random sample of 300 checked records (seed 20261202, drawn 2026-11-18), 295 summaries matched the source (98.3%; 95% Wilson interval 96.2% to 99.3%). The sample was drawn at random from all 1,268 checked records, stratified by country. Each record was judged against the source text fetched the same day by two independent readers (Claude Opus); where they disagreed or one skipped a record, a third reader decided. A record counts as correct only if every claim in the summary is in the source (no invented, stricter or looser obligation, scope as in the source), no obligation that would change the kind of rule is missing, and the kind of rule fits the source. The two readers agreed on all 300 records. The 5 records counted as wrong: 4 with a claim the readers could not find in the source text shown to them (one of these also with the wrong kind of rule) and 1 with a missing obligation. In all 4, the passage quoted as evidence lay outside the reading window shown to the readers (a long source is cut to 30,000 characters around its AI passages); they still count as errors. A first measurement on the same day (seed 20261201) gave 291 of 300 (97.0%; 95% interval 94.4% to 98.4%) and was not published: its lower bound was under the 95% bar. After it, every record was read again for the most frequent error type (wrong kind of rule, together with a missing obligation that changes the kind) before the second sample was drawn.

## 9. The role of AI models

Searching and fetching were done by AI research agents (Claude Sonnet, Anthropic) under written instructions. Decisions on scope and the second reading were done by a different, larger model (Claude Opus). A person sets the rules, decides open cases and approves every release. The word-for-word check and the check on absent search terms are done by a script, not by a model.

## 10. Figures

"Institutions that publish a rule" is the number of records with status `checked` or `link_only` and `scope = institution`, with `on_official_list = true`. The column `counts_as_institution_rule` is `true` for exactly these records (in `legacy/`: `telt_als_regel_instelling = ja`): counting it gives the figure on the register pages, `institution_rule` in `coverage.csv` and `summary.csv`, and `regel_instelling` in `dekking.csv`. A record with a rule that is not on an official list is in the download but not counted. Records with `scope = unit` are reported separately (`unit_only` in `coverage.csv` and `summary.csv`) and are not part of the share that asks for a declaration. The share that asks for a declaration counts the records with `rule_type = disclosure`. `summary.csv` is computed from `alle.csv` by the export script and can be recomputed from it.

## 11. Updates and versions

Each record has the date of its last check in `checked_on`; for `link_only` this is the date on which the link was seen or the source refused. The sources are fetched again on a regular schedule and at least once a month. When a quotation is no longer in the source, the record gets the date of that finding in `source_changed_on` and the page says that the source changed and the record is being re-checked. A person reads the source before the record changes; a proposed new summary is prepared automatically but is never published without that reading.

A release has a version number of the form year.month.sequence and a line in `CHANGELOG.md` with the totals at that moment. An `id` is stable and is never reused, also after a record is withdrawn.

## 12. Limits

- The register depends on what a search engine and the institution's own site make findable. Rules behind a login, in a learning platform or in an internal handbook are outside its reach.
- Coverage differs by country. In some countries almost no institution has regulations on a public site; `coverage.csv` shows this per country.
- A regulation for one level of the institution can be silent on AI while another level has a rule.
- The kind of rule is a simplification of a text that often says several things.
- Sources change. A record is true on the date in `checked_on`.

## 13. Corrections

An institution or reader who finds an error can report it to support@esmond.be, with the `id` of the record. A correction is checked against the official source in the same way as any other record and is listed in the changelog of the next release. An institution that does not want its quotation in the column `quote` can ask for its removal through the same address; the quotation is then removed in the next release.

Verification with the institutions. Next to the checks against the published sources, we ask institutions directly whether their record is right: since November 2026, starting with Flanders and the Netherlands, and on an ongoing basis. A reply from an institution is not a source in itself. A summary still has to be supported by the published text; a reply can lead us to reread the source, to change the wording, or to add the reference to an internal guideline that the institution names. Changes after such a reply are listed in the changelog of the next release.

## 14. Names in the legacy files

Up to release 2026.11.4 the download used Dutch column names and values. The files in `legacy/` keep them for one release. `DATA_DICTIONARY.md` gives the mapping per column; the values map as follows.

| column | legacy value | value |
|---|---|---|
| status | `getoetst` | `checked` |
| status | `gepubliceerd_niet_getoetst` | `link_only` |
| status | `geen_regel_gevonden` | `no_rule_found` |
| status | `onvindbaar` | `not_found` |
| status | `niet_onderzocht` | `not_researched` |
| rule_type | `toegestaan` | `permitted` |
| rule_type | `melding` | `disclosure` |
| rule_type | `overleg` | `consult_supervisor` |
| rule_type | `goedkeuring` | `prior_approval` |
| rule_type | `opleiding` | `left_to_programme` |
| rule_type | `fraude` | `misconduct_only` |
| rule_type | `gepubliceerd` | `link_only` |
| rule_type | `geen` | `none` |
| scope | `instelling` | `institution` |
| scope | `onderdeel` | `unit` |
| counts_as_institution_rule | `ja` / `nee` | `true` / `false` |
| reason | `robots` | `robots_txt` |
| reason | `403` | `http_403` |
| reason | `geweigerd` | `refused` |
| reason | `geen_tekst` | `no_text_layer` |
| reason | `geen_publieke_bron` | `no_public_source` |
| reason | `poort_weigert` | `gate_refused` |
| reason | `niet_op_te_halen` | `not_retrievable` |
| reason | `niet_onderzocht` | (empty: the status says it) |

The legacy files carry no `quote`, and for `link_only` the fixed sentence in `regel`; the English files leave `rule_summary` empty there.

## How to cite

Esmond. (2026). Esmond AI policy register: what higher education institutions say about AI in the thesis (Version 2026.11.8) [Data set]. Zenodo. https://doi.org/10.5281/zenodo.24406864

To cite all versions, use the concept DOI https://doi.org/10.5281/zenodo.24258917. BibTeX and RIS: `README.md` and `CITATION.cff`.
