Organization: Softmax ()
From sexology to the frontiers of artificial intelligence, Mahault Albarracin has spent her career following one stubborn question across every boundary it crosses. The field of AI is better for it.
About Mahault Albarracin
There are scientists who build careers around answers, and then there are those who build them around questions they cannot leave alone. Mahault Albarracin belongs firmly to the second kind. Her path into artificial intelligence did not begin with code or circuits. It began with a graduate thesis on gender and sexual fluidity and a deceptively simple question underneath it: how does a person construct a model of who they are, and what happens when the world forces them to rebuild it? That question, it turned out, was a question about inference. And once she found the formal language to ask it properly, through Karl Friston’s free energy principle and the framework of active inference, she never really stopped asking.
Today, as Machine Learning Engineer and Special Advisor in AI Science and Research at Softmax, Albarracin works at a genuinely rare intersection: formal cognitive science, applied machine learning, and the kind of ethical clarity that comes from having thought seriously about who gets left out of a room and what that costs everyone inside it. She has co-founded advocacy initiatives, contributed to strategic AI roadmaps, and published work that sits comfortably in the citations of neuroscientists, philosophers, and product engineers alike. She has also done graffiti, practiced yoga, and remained deliberately skeptical of the word visionary whenever it gets pointed at her.
In this exclusive Fempreneur conversation, she speaks about the path that brought her here, the science she finds most urgent, and why the most important thing AI can learn to do right now is admit what it does not know.
THE JOURNEY OF AN AI VISIONARY
Q1. Your work sits at the intersection of artificial intelligence, robotics, and cognitive science. What initially inspired your passion for AI and computational cognition?
I did not come to AI through computer science. I came through sexology. My master’s was on gender and sexual fluidity, and what kept pulling at me was a plainer question underneath it: how does a person build a model of who they are, and then rebuild it when the world pushes back? That is a question about inference. When I found Karl Friston’s work on the free energy principle, it handed me a formal way to ask it. Active inference let me treat minds, artificial ones included, as systems that stay alive by predicting and updating. I never really left that first question.
“Active inference let me treat minds, artificial ones included, as systems that stay alive by predicting and updating. I never really left that first question.”
Q2. From research affiliations to advisory leadership roles, your journey reflects both scientific depth and strategic vision. What key experiences have shaped the professional you are today?
A few things, and not in a straight line. Co-founding initiatives like Sexualis and helping run the Manifeste des femmes en tech in Montreal taught me that representation is not charity, it is epistemics: who is in the room decides what counts as a problem worth solving. At VERSES I learned how to take an idea from a whiteboard proof to something a product team could actually ship without lying about what it does. The PhD taught me patience. Real interdisciplinary work is slow, and you have to be willing to be the least expert person in three rooms at once.
Q3. As someone deeply involved in both academia and industry, how do you balance theoretical innovation with practical real-world AI applications?
Honestly, the tension is a bit overstated. The theory I work on, active inference, is about agents acting under uncertainty with limited resources. That is already a practical constraint baked into the math. Where it gets hard is timelines. A research question can stay open for years; a product cannot. So I keep two clocks running. One asks whether the math is right and will still be right in a decade. The other asks what we can responsibly put in someone’s hands this quarter. The mistake is letting either clock set the other’s pace. I have watched both failures up close.
Q4. Being recognized as The Most Iconic Female Personality to Watch in 2026 is a remarkable honor. What does this recognition mean to you personally and professionally?
I have a complicated relationship with the word iconic. I came up through women-in-tech advocacy, so I know recognition like this does real work: a young woman reads the name and thinks, that path exists. That part I take seriously. Personally, I am wary of the visionary framing. The work is collective. Almost everything with my name on it has five other names beside it, usually people you have never heard of doing the hard parts. So I will take the honor as a kind of pointer, aimed at the field and the people in it more than at me.
ADVANCING AI, SCIENCE AND INNOVATION
Q5. In your role at NEXT Integrative Minds Life Sciences, you contributed to developing advanced AI models using Partially Observable Markov Decision Processes. How do you see these technologies transforming the future of intelligent systems?
At NEXT we used partially observable Markov decision processes to model something deceptively ordinary: a learner whose internal state you cannot see directly. You only get noisy evidence, a wrong answer, a hesitation, a pause, and from that you have to infer what is going on underneath and act on it. That is the same structure as a good tutor, or a good clinician. The shift I care about is not bigger models. It is models that carry an explicit sense of their own uncertainty and change what they do because of it. Systems that know when they do not know are the ones I would trust near people.
“Systems that know when they do not know are the ones I would trust near people.”
Q6. You have helped shape strategic AI roadmaps and system architectures. What are the most important considerations organizations should prioritize when building responsible and future-ready AI solutions?
Build the governance before you need it, not after the incident. I think about it in layers. Who answers for the people building the system. Who answers for the system once it starts acting on its own. And who answers when many systems interact and produce something none of them intended. Most organizations only plan for the first layer. The other practical thing: give your ethics people real audit power from day one, not a veto they are too nervous to use. Ethics bolted on at the end is just public relations with a worse conscience.
Q7. Your work in active inference and cognitive modeling is highly interdisciplinary. How important is cross-disciplinary collaboration in accelerating breakthroughs in AI research today?
It is not optional for the problems I find interesting. Active inference only exists because physicists, neuroscientists, and philosophers agreed to argue in the same notation. My own background, sociology and sexology sitting next to machine learning, is exactly why I notice social dynamics that other modelers walk straight past. The cost is real though. You spend months just learning each other’s vocabulary, and you will be wrong in public quite a lot. People oversell collaboration as synergy. In practice it is closer to a long, slightly uncomfortable conversation where nobody gets to stay the expert.
Q8. AI is evolving at an extraordinary pace. Which emerging trends or innovations do you believe will redefine the industry over the next decade?
I will bet against the current consensus a little. I do not think the next decade belongs to ever-larger language models. They are remarkable, and also fundamentally unreliable, because they have no stake in being right. The interesting work is in agents that hold a model of their world and themselves over time, that can act, take feedback, and pay a price for being wrong. Energy will force the issue too. We cannot keep scaling compute the way we have. Constraint tends to produce better science than abundance does, so I am oddly optimistic about the coming squeeze.
“Constraint tends to produce better science than abundance does. I am oddly optimistic about the coming squeeze.”
Q9. As a board advisor and AI strategist, how do you approach decision-making in an environment where technology, ethics, and societal impact are increasingly interconnected?
I try to make the value judgment explicit instead of smuggling it inside a technical choice. Every architecture takes a position on who matters and what it is acceptable to get wrong. So the first question on any board or project is usually the unglamorous one: who carries the cost if this fails, and were they in the room when we decided? After that I want reversibility. In conditions you cannot fully predict, the smart move is keeping your options open and your decisions auditable, so that when you are wrong, and you will be, you can see exactly how and walk it back.
LEADERSHIP, REPRESENTATION AND INFLUENCE
Q10. Women continue to make significant strides in STEM and AI. What challenges have you encountered throughout your journey, and how have they shaped your perspective as a leader in this field?
The obvious ones, and I will not pretend otherwise. Being read as the note-taker in a technical meeting. Watching an idea get ignored until a man restates it. Coming into AI from sexology, I got the bonus of people deciding my background was soft before they had heard a single argument. What it taught me is that credibility is partly performed and partly granted, and the granting is uneven. So I stopped waiting to be granted it. I built the work, published it, and let the citations do the arguing. It also made me stubborn about holding the door open behind me.
“I stopped waiting to be granted credibility. I built the work, published it, and let the citations do the arguing.”
Q11. Your career reflects a strong commitment to bridging academic research with industry transformation. What drives your passion for creating meaningful impact through technology?
Stubbornness, partly. I find it hard to work on anything I cannot connect to someone’s actual life. The NEXT work pulls at me because it is about learning, about giving people sharper models of themselves, which is where I started two degrees ago. There is also something close to moral in the free energy principle for me. Living things persist by staying in contact with reality and updating. I would like the technology we build to do the same, to stay honest about the world instead of confidently hallucinating a nicer one. That gap, between confidence and contact, is what I keep circling back to.
Q12. For young women aspiring to pursue careers in artificial intelligence, robotics, and scientific research, what advice would you offer them today?
Do not wait until you feel qualified, because that feeling shows up late and unevenly, and for women it is taxed harder. Pick a real question and follow it across whatever disciplines it runs into, even the ones you were never trained in. My strangest credential, sexology, turned out to be my sharpest tool in AI. Find the people who will argue with you honestly instead of flattering you; you learn more from one good disagreement than from a hundred great-points. And keep something in your life that is not optimized. I do graffiti and yoga. It keeps me human enough to do the work.
VISION, IMPACT AND THE FUTURE
Q13. Looking ahead to 2026 and beyond, how do you envision the future relationship between AI, cognitive science, and human decision-making evolving?
I think the boundary gets blurrier, in a good way and a dangerous way at the same time. The good: cognitive science gives us AI that reasons about uncertainty more the way we do, so working with it feels less like operating a tool and more like thinking alongside a colleague who is strange but useful. The dangerous: systems that model us well can also steer us, quietly, and most people will never feel the nudge. So the question is not whether to let AI into human decisions. It already is in. It is whether we keep the steering legible. I want partners I can interrogate, not oracles I obey.
“I want partners I can interrogate, not oracles I obey.”
Q14. As a visionary shaping the future of AI and computational research, what legacy do you hope Mahault Albarracin will leave behind for future generations of innovators and scientists?
I am a little allergic to the word legacy, so let me answer sideways. If anything outlasts me, I hope it is a way of working rather than a result. That you can take a question seriously from the humanities and the formal sciences at once, without apologizing to either side. That you can build AI that is honest about what it does not know. And that the people I trained, or argued with, or made room for go on to do things I would never have thought of myself. The math will be superseded, and it should be. I would rather leave behind better arguers.
