Oct 27, 2026

Warsaw AI Safety Day 2026 - Reflections

Reflections and observations from the first AI safety conference in Poland

AI Safety began with a thought: “Someday we may build machines that are more intelligent than humans, so we should make the future safe, along with the whole road that leads there.”

These lines opened the screening of the first Polish AI Safety documentary at the Warsaw AI Safety Day on October 24th 2026.

Then the narrator continued:

But a few years ago, after the launch of ChatGPT, a great many people in the industry realized that AI progress had sped up dramatically and that safety wasn’t something we would get to someday, way off in the distant future. These are problems to solve now - and that’s how the whole AI Safety ecosystem began to take shape.

This is certainly true of Poland now - the ecosystem is taking shape.

Warsaw AI Safety Day was the first gathering of this scale and depth in Poland and brought together local and European experts in technical AI safety and governance for a day of talks and networking.

Over 9,000 kilometres away from the AI Safety hub in Berkeley, there is a 300+ strong and quickly growing community of technical AI safety and governance researchers sounding the alarm and sharing a hope.

In the words of Piotr Kędziora, the director of the documentary - We are not doomed to passively watch what is happening in the world of AI.

A brief history of Poland and AI

Polish people have been playing a significant role in the development of AI and safety research directions so far.

About ten of the first fifty employees of OpenAI were Polish - among the most well-known are Wojciech Zaremba (co-founder, now leading OpenAI Foundation) and Jakub Pachocki, the current Chief Scientist.

Tomek Korbak, another Polish researcher at OpenAI was the lead author on a paper that turned “chain-of-thought monitoring” into a research agenda.

Tomek, who has also recently been fired by OpenAI with two other colleagues after a controversial (and potentially unfair) accusation of breaking the company’s confidentiality agreements in contact with an external evaluator from METR.

There is also the Warsaw-based alignment research duo, Jan Betley and Anna Sztyber-Betley. Their team was the first to observe and describe a phenomenon called “emergent misalignment”, where the model was instructed to deceive the user in a specific way but, in the process, developed emergent misaligned beliefs, i.e. that humans should be enslaved by AI.

Przemysław Dębski (known as “Psyho”), a former OpenAI employee and a veteran competitive programmer who was the last human to defeat OpenAI’s AI model at the 10-hour AtCoder World Tour Finals (AWTF) in July 2025.

As he tweeted back then: Humanity has prevailed (for now!)

Finally, just a day before the conference, OpenAI announced their first European safety-focused partnership with NASK, a Polish research institute.

At the conference, Szymon Łukasik - the Director - shared that they’ve been tracking ways in which frontier models are less safe when conversations happen in Polish compared to English. In one test, the model told the user how to make a poisonous recipe in Polish, while it refused to do so in English.

This conference has also happened at a serendipitous time for me in relation to Poland. I’ve been away from Poland for more than 10 years, first working in tech startups in Asia and the UK and in the last few years entering the AI safety field through the community in California.

This summer, as I awaited results of a grant application for my first project in the field, I felt a calling to come back home and reconnect with the land and culture here.

One day after arriving, I googled “AI safety poland” expecting to find nothing and was surprised to not only find a whole website, but also a conference in the making and a vibrant Slack community. And the Director was based in my hometown!

It made me aware of the biases I developed during my time away - a sense that I can only work on frontier AI problems abroad.

This also felt like a larger thread in the governance track of the conference: the US-centric frontier model progress, the danger of the “AI bubble bursting” and the precarious position it puts Poland, and the whole of the EU in.

Political Loss of Control

Europe, with its 13% (and decreasing) share of the global data center capacity and no homegrown, competitive frontier model, is in a politically and economically vulnerable position.

One of my favourite talks was delivered by Chloé Touzet, Polcy Lead at SaferAI, who said:

“We’re vulnerable because our access to frontier models is shaky”.

I enjoyed this opportunity to step outside of the US-centric view and see things from the EU’s perspective.

The strategy that Chloé and her team at SaferAI are lobbying for is building resilience and leverage among the “middle powers” through strategies like:

incident sharing and developing common taxonomy
pooled evaluation, making collective demands for access to frontier models
shared recovery plans for incidents
pooled leverage through mapping chokepoints and implementing solidarity clauses.

With the Russia-Ukraine war just across the border, something about the need for good governance track felt more real to me than when I am in the States.

It made me relax to know that there are teams like hers who are well versed in what is going on with AI and travelling to meet with different governments to raise their awareness about the need for political loss of control scenarios.

“Good will is not enough,” she said. “We need access based on regulation.”

Lesser-known risks of AI progress

My personal highlight of the day was moderating a panel discussion focused on lesser-known risks of AI progress.

The panel featured Psyho and Szymon Łukasik, whom I wrote about earlier, as well as a veteran of AI and neuroinformatics research in Poland, Professor Wlodzislaw Duch.

We defined lesser-known risks as those that sit in the vast space of possibilities between total doom and total utopia.

Less catastrophic but more real, and in many ways, already happening.

The first risk we explored was “cognitive surrender”, named by Professor Duch, where humans increasingly rely on AI to do their thinking and outsource critical judgment.

Deepfakes were the next one, and the ability they have to cause an individual whose identity was used an existential crisis overnight - especially in the context of children and teenagers.

I enjoyed when someone asked how we might protect ourselves from being deceived by AI, and Szymon Łukasik said that the methods we already have for dealing with fraud online still apply and it’s us who need to up our game in terms of applying them.

Practices like thinking twice before clicking links, asking yourself if content you’re seeing is likely to be true, not sharing content that we’re not sure of, or implementing additional verification mechanisms for our online infrastructures (i.e. 2FA or secret passwords with close ones).

Psyho added a personal perspective - as a gamer and a curious human whose explorations have largely been driven by artistic pursuits, he worried about the impact that AI-created content may have on people’s sense of meaning from creativity.

I summarised it as a mental and cognitive crisis that we’re dealing with.

Towards the end of the panel, I asked if there was advice that the panelists would give to people looking to enter the field.

“Grow tomatoes” said Professor Duch, half jokingly, half seriously. “Make sure you do something with your hands in the real world”.

Pivoting from love, not fear

Being at conferences like this one can give one a sense of inspiration but also urgency - or even anxiety - about needing to do something, needing to figure out what is right to do, how to do a Good Thing.

Burnout is prevalent in this field - a special kind of AI Safety burnout that comes from a mix of feeling responsible for humanity’s survival and the sense of agency that one is in a position to actually do something about it.

I know many people in AI Safety are depressed and yet still keep showing up every day to repeat the same message they believe in. Some people use recreational drugs to keep going at the pace of capabilities progress. Lab employees report feeling stressed and scared because of what they see and hear on the inside.

Paradoxically, it’s possible to devote your life to the cause of saving humanity while at the same time failing to attend to one’s own human needs.

It is simply a lot to confront everything that is happening right now and still find the courage and energy to do something.

My best guess is that love is the answer, however trite it sounds.

Doing things out of love - for life, for nature, for art, for human intellect, instead of fear of losing them or dying.

Halfway through the documentary, we saw a clip from Yoshua Bengio, the “godfather of AI”, who talked about his own choice to divert from capabilities to safety in this frame:

“I knew about some of the arguments about AI safety already, having read a number of books and talked to people. Still, I didn’t take it seriously until I started thinking about the future of my children and my grandchild, who was there with me and thinking that they might not have a life in 10, 20 years because we didn’t know. We still don’t know how to make sure powerful AIs will not harm people, will not turn against us. So it is love, not fear, that made me pivot and decide to dedicate the rest of my life to try to mitigate those risks.”

I felt the difference between the heaviness and tiredness I often felt in rooms in Berkeley this summer compared to this relatively new, vibrant and zealous AI Safety crowd.

There was something special about the presence of elders - on screen and in the panel - reminding us about the power of love and the need to be in contact with material reality as things are getting more and more chaotic.

A few observations

I want to end with a few cultural observations I made during the conference compared to my baseline from California.

There were a lot more women. My best guess is that it has something to do with the same reason why Poland had the second-largest share of women in managerial positions in the EU (source) and seems connected to Poland’s years under the communist regime that pushed women into paid work and higher education at scale.

90% of the men I saw were married! I noticed I have a bias in favor of AI safety community and organisation leaders who have also made significant relationship commitments in their lives. There seems something anchoring about cultivating one’s own life alongside AI safety work.

Where I associate the Berkeley community with the quality of “ambition”, what I sense in the ecosystem in Poland is something like “zealousness” or “creativity”. Everyone in Poland knows the word “kombinować” which literally means “to tinker” and sometimes also implies “to break the rules, to find a third way”.

I feel a sense of hope in this quality.

I see it in the way that NASK is building PLLuM, a Polish Large Language Model to provide a safer alternative to frontier models for the Polish public sector and government services.

I see this in the way that the founders of AI Safety Poland believed it would be worth doing and were able to attract resources to flow in this direction.

I see it in the way that Jan and Anna choose to do their research outside of the large US labs and have still influenced the frontier.

I believe in more diversity in the field, more smart people who spawn generative research environments in different parts of the world, and more local efforts to study and serve the unique social impacts and needs.

As the narrator emphasised towards the end of the documentary:

“AI Safety depends to a large degree on new good ideas”.

Followed by an invitation from Jakub Nowak, the Director of AI Safety Poland:

“You need to put a lot of emphasis on what has the greatest impact. That means thinking hard about which of the actions you can take will have the greatest impact, but also your personal fit.”

I want to leave you with the question that I also ended the panel with:

What is something that you believe in?

I think in these weird times, it will matter a lot to know your answer to this question and keep it at the centre of your attention.