Executive summary
Romanian
universities are deciding how to regulate generative AI, and the Ministry of
Education and Research is considering a national framework. This report asks at
which level the detailed rules should be written. It argues that common
principles are welcome but common detailed rules are not: the rules on what
students and researchers may do with AI should be set close to the discipline
and revised regularly to follow the international journals of each field.
Otherwise, Romanian universities will drift out of step with the communities in
which their students and researchers are meant to work.
Approach.
The AI policies of four large publishers (Elsevier, Springer Nature, the
Science journals, Wiley) and two philosophy journals were compared with the
policies of Oxford, Stanford and the Russell Group, and with three Romanian
university regulations, read in full (West University of Timișoara, Politehnica
Bucharest, UMF Carol Davila). Three studies on AI-text detectors and the
author's own experience with detectors complete the evidence.
What
journals require. A stable core is shared: AI cannot be an author, human
authors are accountable, substantive use must be declared, and manuscripts
under review must stay confidential. Beyond this core, policies diverge by
discipline. Science publishers allow AI drafting with disclosure. The British
Journal for the Philosophy of Science, since 7 October 2026, forbids AI even
for improving readability. Policies also change quickly; the major publishers
have revised theirs more than once since 2023.
What
universities ask. Leading universities set principles centrally and leave
the detailed rules to the assessment, the instructor or, for research, the
publisher. In Romania, West University of Timișoara follows this model,
requiring an AI section in every course syllabus (fișa disciplinei) and
faculty-level rules for theses. Politehnica Bucharest sets a single
university-wide rule for theses, and UMF Carol Davila sets university-wide
guidelines.
Alignment.
All institutions examined share the journals' core. Those that delegate
detailed rules can align with the journals of each field by construction. A
single university-wide rule coincides with one family of journals and differs
from others.
Consequences.
For a multidisciplinary university such as the University of Bucharest, one
rule cannot fit all faculties, or even all courses within a faculty. Thesis
rules should follow each programme's reference journals, doctoral dissertations
and the papers built from them should not fall under different standards, and
detection technology changes faster than any regulation.
Recommendations.
For universities: four layers of regulation (national principles, university
procedures, programme benchmarks indexed to international journals, course
rules in the syllabus), reviewed yearly; for theses, the principle of segmented
contribution, judging each component by the norms of the community that judges
that kind of work. For students: read the syllabus, declare all use, keep a
log, verify every reference, protect confidential data, read the target
journal's policy, and assume that undeclared use will be detected.
Limits.
This is a policy analysis, not a systematic review. The sample of journals and
universities is illustrative, policies were consulted as of October 2026 and
will change, and the author's experience with detectors is an informal
comparison, not a formal test.
Keywords:
generative
AI; academic integrity; higher education policy; journal policies; AI
disclosure; AI-text detection; subsidiarity; adaptive regulation; Romanian
universities; doctoral education
The full text can be downloaded from this link.
Note: text produced using several AI tools.
