Get your work evaluated at esmondreview.com
EsmondAcademic review · Made by educators How Esmond works Why Esmond?

AI policy register: an open dataset of thesis AI rules

For each higher education institution on an official national list: does it publish a rule on generative AI for the thesis or dissertation, what kind of rule, the official source and the date we checked it. Free to download, reuse and cite.

Open datasetCC BY 4.0
3,903Records
44Countries with an official list
1,287Checked against the source
1,268With a rule for the whole institution

One record per institution: whether it publishes a rule on generative AI for the thesis or dissertation, the kind of rule, a one-line summary, the official source and the date of the check. In 44 countries the denominator is the official list of institutions from the public authority or the sector body: 3,756 institutions, of which 3,618 researched so far.

Key figures

Figure 1. Status of the 3,903 records in release 2026.11.8.
  • Checked 1,287 (33%): the source was fetched and the quoted passage is in it word for word.
  • Published, link only 163 (4.2%): the official rule exists but the site keeps automated tools out or the document is a scan; link, no summary.
  • No rule found 264 (6.8%): no rule on AI in the current regulations for the whole institution.
  • Not found 2,051 (52.5%): no public rule found on the institution's site.
  • Not yet researched 138 (3.5%): on the official list, not yet searched.
Figure 2. Kind of rule in the 1,287 checked records.
  • Disclosure required432 (33.6%)
  • Set by program301 (23.4%)
  • Approval first224 (17.4%)
  • Permitted138 (10.7%)
  • Misconduct clause121 (9.4%)
  • Discuss first71 (5.5%)
Figure 3. Share of the institutions on the official list that publish a rule on AI for the whole institution, in the 32 countries with at least 30 institutions on the list. A country with many institutions not yet researched scores low until the search is done.
  • Australia78.9% (30 of 38)
  • Finland76.3% (29 of 38)
  • United Kingdom74.5% (149 of 200)
  • Canada74.2% (72 of 97)
  • Belgium73.8% (48 of 65)
  • France66.7% (48 of 72)
  • Switzerland64.9% (24 of 37)
  • Norway64.6% (31 of 48)
  • United States64.4% (210 of 326)
  • Netherlands63% (34 of 54)
  • Austria57.1% (44 of 77)
  • Spain49.5% (47 of 95)
  • Sweden44% (22 of 50)
  • Germany42.5% (166 of 391)
  • Denmark42.4% (14 of 33)
  • Italy39% (39 of 100)
  • Poland33.1% (45 of 136)
  • Colombia25.8% (23 of 89)
  • Portugal24.7% (23 of 93)
  • Peru20% (21 of 105)
  • Brazil19.4% (40 of 206)
  • Mexico14.1% (11 of 78)
  • Ecuador12.5% (8 of 64)
  • Tanzania12.5% (7 of 56)
  • Argentina10.7% (13 of 122)
  • Ghana7.8% (4 of 51)
  • Nigeria3% (10 of 328)
  • Kenya2.6% (2 of 77)
  • Uganda1.6% (1 of 64)
  • Algeria0% (0 of 58)
  • Congo - Kinshasa0% (0 of 189)
  • Côte d’Ivoire0% (0 of 155)

All figures are computed from release 2026.11.8 by the scripts that build the download; the country figures are in coverage.csv. A record with status Not found is information too: it means no public rule turned up on the institution's site.

Download

Licence: CC BY 4.0, reuse allowed with attribution. 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. Each register page also offers its own country file.

The dataset is also on Zenodo (every release, with a DOI), Hugging Face, Kaggle.

How to cite

Cite the version you used. The version DOI points to that release; the concept DOI always resolves to the latest release.

APA 7

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

APA 7, latest release (concept DOI)

Esmond. (2026). Esmond AI policy register: what higher education institutions say about AI in the thesis [Data set]. Zenodo. https://doi.org/10.5281/zenodo.24258917

BibTeX

@dataset{esmond_ai_policy_register_2026_11_8,
  author    = {{Esmond}},
  title     = {Esmond AI policy register: what higher education institutions say about AI in the thesis},
  year      = {2026},
  version   = {2026.11.8},
  publisher = {Zenodo},
  doi       = {10.5281/zenodo.24406864},
  url       = {https://doi.org/10.5281/zenodo.24406864},
  note      = {All versions: https://doi.org/10.5281/zenodo.24258917}
}

RIS

TY  - DATA
AU  - Esmond
PY  - 2026
DA  - 2026/11/20
TI  - Esmond AI policy register: what higher education institutions say about AI in the thesis
ET  - 2026.11.8
PB  - Zenodo
DO  - 10.5281/zenodo.24406864
UR  - https://doi.org/10.5281/zenodo.24406864
N1  - All versions: https://doi.org/10.5281/zenodo.24258917
LA  - en
ER  - 

Accuracy

In a random sample of 300 checked records (seed 20261202, 18 November 2026), 295 summaries matched the source.

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.

What the status checked means: the source was fetched through a gate that follows robots.txt and AI-bot rules, at most one request per two seconds per site, and the quoted passage was found in it word for word. That confirms the passage, not that a one-line summary is complete, so read the source before you rely on a record. Every checked source is fetched again on a regular cycle; a changed source is marked on the row until a person has re-read it.

Corrections

Every row on the register pages has a link "Report a correction", which opens an email to [email protected] with the record id filled in. Institutions can also ask us to remove a quotation of their rule; it is removed in the next release. Each correction is logged in corrections.json: date, record id, what changed, and whether an institution or a reader reported it. No names.

  • , be-kdg: Summary corrected after the institution's reply: students may use AI unless the lecturer decides otherwise; the earlier claim that the regulations are not public was wrong and is removed. (reported by the institution)
  • , be-luca: Summary now says, at the institution's request, that LUCA follows the AI guidelines of KU Leuven; the regulations themselves still contain no AI rule. (reported by the institution)
  • , be-etf: Summary set to the institution's own wording, with a reference to its internal guidelines for responsible use of generative AI. (reported by the institution)
  • , be-ugent: Summary reworded after the institution's reply: responsible use of generative AI is allowed, also in the master's thesis; no syllabi or articles in AI tools without the author's consent; fabricated sources count as an irregularity. (reported by the institution)

Embed a record in a library guide

Each row on the register pages has a button "Embed this record". It copies one line of plain HTML: the institution, the rule in one line, the official source, the date of our check, a link to the row and the DOI. No script and no tracking: paste it into a LibGuide or any web page.

Releases

  • , version 2026.11.8: 3,618 of 3,756 institutions researched in 44 countries, 1,268 with a rule on AI.
  • , version 2026.11.7: 3,618 of 3,756 institutions researched in 44 countries, 1,243 with a rule on AI.
  • , version 2026.11.6: 3,618 of 3,756 institutions researched in 44 countries, 1,243 with a rule on AI.
  • , version 2026.11.5: 3,618 of 3,756 institutions researched in 44 countries, 1,255 with a rule on AI.
  • , version 2026.11.4: 3,618 of 3,756 institutions researched in 44 countries, 1,245 with a rule on AI.
  • , version 2026.11.3: 2,952 of 2,952 institutions researched in 37 countries, 1,039 with a rule on AI.

Browse by country

The other countries are in the download and in the tracker for the United States, Canada, Africa, Asia and Oceania.

What a record holds

A record is one institution. It gives the status of our search, the kind of rule, the rule in one sentence in our own words, whether the rule covers the whole institution or one faculty or program, the title and link of the official source, and the date of our check. Each record has a stable id, so a link, a citation or a correction points to one institution across releases.

Per country the starting point is the official list of recognized institutions from the public authority or the sector body, named in coverage.csv. An institution on that list without a public rule is in the file too, with its status. That makes shares comparable: the denominator is the list, not the institutions that happen to publish something.

What the register does not do

A record sums up a published rule and links to it. It is not legal advice and not the institution's own wording. A program, a department or an instructor can set a stricter rule than the institution, so read the source and ask your supervisor or advisor before you rely on a record.

Where a site keeps automated tools out, we respect that: the record then carries the link only, with no summary. We do not guess a rule from a news item or a secondary source.

Who maintains it

The register is maintained by Esmond, a Belgian company; it is a separate, free resource with a monthly release cycle.

Frequently asked questions

Can I reuse the data?

Yes. The data is published under CC BY 4.0: you may copy, adapt and redistribute it, also commercially, as long as you cite the register. Cite the version you used. One column is the exception: the short quotations in the column quote stay with the institutions and are not under CC BY 4.0.

How often is the register updated?

Once a month every checked source is fetched again, and each round of new searches becomes a numbered release on Zenodo. The Releases list on this page shows the totals of each release.

I found an error, or I represent an institution. What do I do?

Use Report a correction on the row, or email [email protected] with the record id. Institutions can also ask us to remove a quotation of their rule; it is removed in the next release. Every correction is logged in corrections.json.

Last revised

  1. Report page 1
  2. Report page 2
  3. Report page 3
  4. Report page 4
  5. Report page 5
  6. Report page 6
  7. Report page 7
  8. Report page 8
  9. Report page 9
  10. Report page 10
  11. Report page 11
  12. Report page 12
  13. Report page 13
  14. Report page 14
  15. Report page 15
  16. Report page 16
  17. Report page 17
  18. Report page 18
  19. Report page 19
  20. Report page 20
  21. Report page 21
  22. Report page 22
  23. Report page 23
  24. Report page 24
  25. Report page 25
  26. Report page 26
  27. Report page 27
  28. Report page 28
  29. Report page 29
  30. Report page 30
  31. Report page 31
  32. Report page 32
  33. Report page 33
Figure 1. A real Esmond report on a fictional thesis (33 pages).