UBA doesn't ban AI in exams: it forces AI to account for its bibliography

By La Gaceta · July 3, 2026.

The Faculty of Economic Sciences at the University of Buenos Aires has integrated an artificial intelligence agent into "Mi Econ," its digital platform, and approved a set of rules to regulate its use in the classroom. The tool already operates across 40 subjects, 184 courses and 44 departments, reaching some 2,000 students. Unlike a generic chatbot, the system works in a closed environment: it responds only with the bibliography and materials that each department uploads, without going out to search the internet. The rules also clarify that using AI does not change the authorship of a piece of work—responsibility still lies with the student, who must declare its use when appropriate—and that assessment will begin to value the process as well: drafts, work logs and brief oral defenses, not just the final result.

The interesting point is not that a university "allows" AI—almost all of them already do, by force of circumstance—but that it decides to build its own infrastructure to constrain it. Restricting the system to the official bibliography is a direct response to two real problems: the hallucination of sources (an open model can invent citations or mix theories from different schools without warning) and inequality of access, because if each student uses a different AI with its own quality criteria, assessment ceases to be comparable among peers. By fixing the corpus, the UBA turns AI into a curated tutor instead of an open oracle, and that is a pedagogical decision as much as a technical one.

The shift toward assessing the process—drafts, oral defenses, work logs—is, in fact, the most honest structural response to the problem every teacher has faced since cheap text generators appeared: if the final product can be manufactured in seconds, academic value has to move to the traceability of thought. This connects with something we have already seen in the education sector generally: the teacher (and here, the institution) who learns to orchestrate AI as a learning tool wins, not the one who tries to ban it or the one who ignores it. Universities in other countries have been testing variants of this—oral defenses, process portfolios—but few have gone so far as to build their own closed assistant with institutional funding; that marks a more ambitious bet than a simple usage policy.

Overall, this kind of initiative foreshadows a tension that will recur throughout higher education in the region: those with the resources to build their own controlled AI layer will be able to set their own academic-integrity rules; those who depend on external, free tools will be more exposed to easy cheating and to the quality gap between students. With this move, the UBA positions itself as an early benchmark, but the real challenge is not technological but one of scale: sustaining 2,000 students in a closed, curated system is far more manageable than doing so with hundreds of thousands at massive, lower-budget universities. That is where it will be tested whether this model is replicable or ends up being a niche experiment—well-intentioned but hard to generalize.

In the long term, initiatives like this are the seed of something more promising: an education in which AI does not replace critical thinking but frees up time to exercise it better, with tutors that know exactly the syllabus and can guide without shortcuts. If the short-term transition demands more pedagogical design work and more assessment friction—as the rules themselves acknowledge by requiring oral defenses and process logs—it is a reasonable price against the risk of a generation that delegates reasoning without ever developing it.