Mario Brcic
AI Alignment · Associate Professor, University of Zagreb
I work on making AI do what we intend, safely.
Alignment research in four parts: stating exactly what AI must do, building the tools that check it (formal verification, interpretability), finding where those tools hit hard limits, and asking who gets to set AI's goals.
Open projects
Tools for alignment, and where they are taught.
A research project
VALOR
Value-aligned and interpretable optimization and reasoning.
The FER research project Mario Brcic leads (2025–2029) on value alignment and interpretability of large language models: behavioural audits so far (AIES 2026, code public), an interpretability benchmark under construction, circuit discovery planned.
Visit the project site →
An open Lean library
AI Safety Formalization Atlas
The open Lean library for formal AI-safety results.
An open Lean 4 library of machine-checked AI-safety mathematics, from impossibility results to control and oversight. Each formal statement is checked against the published result it claims, with source and scope on record.
See the project →
A textbook
Modern Algorithm Design
Design algorithms from primitives, and know what to demand of them.
A textbook that teaches algorithms as structures chosen to solve stated problems, and how to check them, including when AI writes the code to pass the test. AI agents write the code and low-level text to the author's designs; machine checks enforce them.
See the project →The Cognitive Sovereignty Trilogy
Three essays that make one argument about power, AI, and who is in charge.
The society part of the work: who sets the goals AI pursues. Cognitive sovereignty is the capacity to keep your own judgment and decisions when AI systems mediate what you see and frame your choices.
01
The Memory Wars
Why AI memory is the new contested terrain.
arXiv:2508.05867
Read essay →02
The Power Gambit · Pt 1
How AI comes to run things without anyone deciding it should.
Shadow sovereignty: AI holding real decision power without formal authority
Read essay →03
The Power Gambit · Pt 2
A working approach to keeping AI aligned.
Seal · Re-aim · Tune: a three-step alignment audit
Read essay →The research
How do we make AI do what we intend, and what can never be guaranteed? The work has four parts:
- Goals: stating exactly what AI must do (Atlas, textbook)
- Tools: machine-checked safety mathematics (Atlas), and behavioural audits and interpretability of language models (VALOR)
- Limits: what no alignment method can guarantee (Impossibility Results in AI, ACM Computing Surveys)
- Society: who sets AI's goals, and whether others can see and contest them (Cognitive Sovereignty)
Background
Associate Professor at FER, University of Zagreb. Principal Investigator of VALOR. Over 3,700 citations (Google Scholar, November 2026).
Background, papers, and full bio on the about page.
Get in touch.
For research collaboration, students or press, the contact page reaches me directly.
Get in touch