BAISH (Buenos Aires AI Safety Hub)

BAISH (Buenos Aires AI Safety Hub) is Argentina's first AI-safety community organisation and the corpus's first Latin American entry for a civil-society AI-safety hub — a Buenos Aires-anchored group whose mission is to support students and researchers across the full talent pipeline from first contact with AI-safety ideas through to active research output. Founded in 2024 and based principally at the Facultad de Exactas y Naturales of the Universidad de Buenos Aires (UBA), BAISH runs courses, reading groups, and project programmes delivered in Spanish, directly addressing the language-access barrier that has limited engagement with the predominantly English-language global AI-safety literature across Spanish-speaking audiences. The AISafety.com global communities directory rates BAISH "Very Active." As of mid-2026 it has 120+ community members, a Telegram group of more than 190 participants, and a seven-person organising team its founding director describes as "almost self-running."

Founding and structure

BAISH was co-founded in 2024 by Eitan Sprejer and Luca De Leo. Sprejer serves as Founding Director and is a full-time AI-safety researcher specialising in LLM evaluation, chain-of-thought interpretability, and mechanistic interpretability; he holds a MARS Fellowship at the Cambridge AI Safety Hub, serves as a course facilitator for BlueDot Impact, and co-founded LANAIS (Latin American Network for AI Safety). De Leo led the operations side of the founding; he also co-founded ASTN (the AI Safety Talent Network, safetytalent.org), a career platform for the field, and joined 80,000 Hours as an engineer in 2026.

The seven-person organising team includes staff members Gaspar Labastie, Tobias Bersia, Gonzalo Heredia, and Rocio Monjes, and advisors Sergio Abriola (PhD) and Guido Bergman. Funding comes from Open Philanthropy ($14,500+ in grants to BAISH), Coefficient Giving (via its Navigating Transformative AI fund), and the Kairos Pathfinder programme for AI-safety field-building. The hub's primary venue — Facultad de Exactas y Naturales, UBA — places BAISH within Argentina's most concentrated pool of mathematics, physics, and computer science students and researchers.

Programme

BAISH's educational offering runs across three tracks, all in Spanish and free of charge.

Technical AI Safety Course. The centrepiece is a BlueDot Impact Technical AI Safety course — a 30-hour programme delivered through in-person weekly sessions at Exactas plus asynchronous preparation, covering current technical safety techniques and open problems in AI alignment. The course targets ML researchers, software engineers, and policy professionals. Multiple cohorts have run, with March 2026 and April–June 2026 editions both oversubscribed, indicating a quarterly cadence and steady demand. The inverted-classroom format with a final project mirrors BlueDot Impact's global delivery model but is localised entirely into Spanish — BAISH's most distinctive programme feature.

Technical AI Safety Project. A five-week, 30-hour programme in which participants replicate recent AI-safety research through programming, bridging coursework and independent research practice.

Frontier AI Governance Course. A third track, covering AI governance, with dates yet to be confirmed as of mid-2026.

Beyond the structured tracks, BAISH runs a paper-reading club, introductory socials, and open community channels (Telegram and WhatsApp) for connecting participants across technical and governance threads. The hub has also extended its reach beyond Argentina, serving as a teaching assistant at two ML4Good bootcamps in Colombia and Brazil in 2025.

Research output

BAISH describes its technical research focus as "Chain of Thought interpretability, LLM evaluations, Mechanistic interpretability of neural networks." Its members have produced a body of peer-reviewed and preprint work in its first two years:

Eitan Sprejer is first author of "Approximating Human Preferences Using a Multi-Judge Learned System," presented at the NeurIPS 2025 LatinX in AI and Reliable ML workshops. He is second author (with Austin Meek) of "Measuring Chain-of-Thought Monitorability Through Faithfulness and Verbosity," submitted to AAAI 2026. The BAISH FAIR research group contributed to "AI Debaters are More Persuasive when Arguing in Alignment with Their Own Beliefs" (NeurIPS 2024 workshop) and produced the arXiv preprint "Mind the Performance Gap: Capability-Behavior Trade-offs in Feature Steering," with Sprejer as first author. A formal research fellowships programme is in planning for 2026.

Regional connections

BAISH sits at the centre of an emerging LatAm AI-safety infrastructure. Eitan Sprejer co-founded LANAIS (Latin American Network for AI Safety), which launched with a kick-off event on 28 June 2025 and aims to connect researchers, students, and organisers across the region through a public directory, career resources, and virtual and in-person networking events. LANAIS is a distinct organisation from BAISH but shares a founder and addresses the same geographic gap at a network rather than hub scale. Regional sister groups include AI Safety Colombia (Bogotá), AI Safety Brazil (São Paulo), and AISMX (Mexico).

BAISH serves as one of five Latin American hub sites for the Apart Research Global South AI Safety Hackathon (19–21 June 2026), alongside EA Brazil (São Paulo), AI Safety Colombia (Bogotá), AISMX (Mérida and Guadalajara), and Santa Cruz Hub (Bolivia). The Latin America track carries a USD 3,000 prize pool across three winning teams, with projects eligible to address technical safety, governance, or locally tailored problems — including AI fairness for Spanish and Portuguese language models and regulatory analysis of emerging Latin American AI legislation.

Place in the movement

BAISH is the corpus's first civil-society AI-safety community group entry in Latin America — the regional equivalent of AI Safety 東京 in East Asia, Berlin AI Safety in continental Europe, AI Safety South Africa in sub-Saharan Africa, Singapore AI Safety Hub in Southeast Asia, and antoan.ai in Vietnam. Like those counterparts, BAISH performs the insider-community pattern — building the local practitioner pipeline through courses, reading groups, and research mentoring rather than public-facing campaign organising.

What distinguishes BAISH from many of those counterparts is its Spanish-first programming strategy. All courses are delivered in Spanish at a Spanish-language research university, directly addressing the barrier that has kept Latin America's substantial technical talent at arm's length from AI-safety work concentrated in English-language institutions and literature. The setting also sets it apart: UBA Exactas is one of Latin America's leading science faculties and has produced internationally recognised computer scientists; Buenos Aires hosts an active EA and rationalist community; the conditions for building a credible technical AI-safety hub are structurally closer to Berlin or Boston than to many of the Global South cities where AI-safety field-building starts from a thinner base. BAISH's research output — four papers across NeurIPS workshops, an AAAI submission, and arXiv in its first two years — and its 150+ member community reflect that foundation.

The Synthesizer's queuing note observed that Latin America "has data-center opposition but no AI-safety local-group." BAISH closes that gap in the corpus while also distinguishing the two organising cultures: the data-center and surveillance opposition campaigns already in the corpus sit in a civil-society-rights tradition, mobilising around concrete harms; BAISH sits in an EA-adjacent insider-community tradition, cultivating technical and governance researchers. Both are in scope; BAISH makes the latter visible for the first time in the LatAm layer of the graph.