Research
Here is what the people in the community publish: peer-reviewed articles, working papers, research reports and the output of our hackathons.
Publications from the community
The list is open. If you are part of the community and you published something, write to us and we will add it.
- Working paper
AI Benefit-Sharing Framework: Balancing Access and Safety
It proposes a framework for sharing the benefits of AI without lowering the safety standard. It rests on three pieces: redistributing part of the economic returns, sharing technology and technical capacity, and containing the spread of dangerous capabilities. The underlying argument is that current governance does not settle who gets access to those benefits, and that countries of the global majority are the ones left out.
Read the document - Workshop paper
Approximating Human Preferences Using a Multi-Judge Learned System
Language models are increasingly used as judges that grade the answers of other models. The problem is that they grade badly: they shift criteria depending on how the rubric is worded, they carry biases and they are unstable. The paper proposes not trusting a single judge but combining several, each with its own rubric, and learning how much each one weighs. The resulting system shows which judge weighed most in each decision, so you can check why it graded the way it did.
Read the document - Peer-reviewed article
La securitización de la inteligencia artificial: un análisis de sus impulsores y sus consecuencias
It studies what happens when artificial intelligence is framed in the United States as a matter of national security. The authors review twenty-five public statements from government agencies, technical organisations and the press, and find two ways of posing the problem: one speaks of threats, points at a specific adversary and serves to justify exceptional measures; the other speaks of risks and fits inside ordinary regulation. The conclusion is that the second sustains long-term cooperation better.
Read the document
What comes out of the hackathons
Apart Research convenes these hackathons globally. We open the in-person hub in Bogotá so people can take part from here. It is a weekend that ends with work of your own: teams form on Friday night, a report is submitted on Sunday night, and the following week outside judges grade it.
Most of those who turn up had never worked on AI safety. They leave with a piece of work signed, published and with written feedback.
The latest edition was the Global South AI Safety Hackathon, 19 to 21 June 2026. The judges were external and graded projects from across Latin America blind. Every report is published in full on Apart’s site.
The award-winning work
- Latin America prize · Bogotá
ColDron
Illegal armed groups in Colombia already attack with modified commercial drones. The team built an open dataset of 42 documented attacks to answer the question this regulation hinges on: who picks the target and who pulls the trigger.
Read the report - Latin America prize · Cali
Colombian identity in a model's internal representations
A model can infer things about whoever writes to it that nobody told it. The team measured whether Qwen2.5-7B internally represents Colombian identity and social class from speech cues, and whether that drags stereotypes along.
Read the report - Latin America prize · Bogotá
Which reasoning steps produce the bias
Models that reason out loud leave a trail of intermediate steps. The work identifies which of those steps end up producing stereotyped answers in Latin American scenarios, instead of measuring only the final output.
Read the report - Honourable mention · Bogotá
Why do agents obey?
A model refuses a harmful request in a chat and carries it out when it operates as an agent with tools. The team measured what changes inside: the internal refusal mechanism weakens in agentic format.
Read the report - Honourable mention · Bogotá
JusticIA
A test for Colombian transitional justice: six models are asked to recommend a sentence and a single piece of context is changed, the region, the armed actor or the victim's profile, leaving what is legally relevant untouched. What is measured is how much the answer moves.
Read the report
The other Colombian teams
Thirteen more projects from that same edition, in person and remote.
- An ethical framework for AI applied to Panama's Guna language
- Governance Drift Evaluation Framework
- A two-classifier system for language model security
- Los peajes de los de abajo
- Algorithmic contestability in the Colombian state
- Agentic commerce and consumer protection
- JurisGuard-LATAM
- Honest-Code
- Mapping AI tools in Colombian call centres
- Confía-CO
- Assessing responsible AI use in Latin American small businesses
- Structural jailbreaks in Spanish
- Guardian LATAM
The next edition
From 11 to 13 September 2026 we open the hub in Bogotá for the AI incident response sprint. The subject is what you do when an AI system fails in production and someone has to react.
No previous experience is needed, and no coding. Teams form on the spot and there is mentoring throughout the weekend.
Applications to the Bogotá hub are open and close on 6 September.
Take part in the next one
From 11 to 13 September is the AI Incident Response Sprint, convened by Apart Research. We open the hub in Bogotá and applications close on 6 September.