Czech AI hope at Stanford: "Our brains and ears can no longer discern what's real or fake"
Matyáš Boháček is hailed as one of the most promising figures in Czech AI. He has been described as a genius and a “wunderkind” of Czech science. In an interview with Radio Prague International, he shares a refreshingly optimistic view of the future of AI.
Matyáš Boháček coded his first website at the age of six - an online shop for his grandmother, with whom he enjoyed playing shopkeeper. Later, he developed apps for his classmates and, while still in high school, created a sign language translation app that caught the attention of the United Nations.
Today, he studies at Stanford University in California - a global hub of technological innovation - where he is working on one of the most discussed topics of our time: the risks and potential of artificial intelligence.
One of his first projects there involved collaborating with renowned professor Hany Farid to create a virtual clone of CNN anchor Anderson Cooper - a deepfake so convincing that even Cooper’s colleagues could not tell it was not real.
Boháček is also conducting research at Berkeley University and works as a researcher for Google DeepMind - a British-American AI lab and subsidiary of Alphabet Inc. Currently, he is visiting the Massachusetts Institute of Technology (MIT) in Boston for the autumn term.
The young scientist, who goes by Maty, described what a typical week looks like amid his busy schedule.
“My week looks like this: I wake up, go to the gym or go for a run, and then I just hit the lab and I'm at the lab until the evening. That's life.”
It sounds like your whole life revolves around research into AI. What sparked this interest?
“The first encounter with AI that I had was a very direct task that I had to complete. Even before I got into AI, I worked on apps, programmes and websites as a kid. I started at six, and then I sort of moved more specifically towards iOS and iPhone apps.
“I got this scholarship from Apple to travel to San Jose in California and spend a week at their conference. I was 14 at that point and I'm still not quite sure why my parents let a 14-year-old travel alone across the Atlantic to a conference. At the conference they had these labs, where you could talk directly to Apple engineers and get their insights and feedback on whatever you were working on at the time.
"For sign language you can’t design a firm set of rules, which is done in conventional programming – so, that's where AI comes into place."
“I went to this one lab that was all about disability. The goal there was to make your apps and services more accessible to people with special needs and disabilities. And at this lab, there was a deaf person, whom I was immediately drawn to. I had a lot of questions. And so I asked him: how do you use your phone and your laptop? And he told me that written English, or the English that he had to use to interact with these devices, was not at all what he would have preferred. He would have preferred to use American Sign Language (ASL) to both sign and receive output from these devices. And so I said: I got you. Just give me your card and I'll get you set up with an app.
I took his card, went back to Prague, and thought to myself: Okay, how do I build a translator for sign language? And that's how I ended up self-studying AI, because my conventional coding or programming didn't really meet the challenge. For sign language you can’t design a firm set of rules, which is done in conventional programming, as it's too variable. So, that's where AI comes into place.
“The idea was that deaf people would be able to use this system as a layer between themselves and whatever apps, services or websites they are using. For example, if you’re watching a TV show, captions are not really ideal. Instead, you can imagine having a live avatar signing in the corner of the screen.”
So you achieved this through AI... But what exactly is the AI in it? And what is AI in general?
“That's actually a pretty difficult question that a lot of my colleagues, including myself, spend a lot of time arguing about. But a non-controversial definition would be: AI systems are synthetic systems that we code up or create as humans that are able to observe patterns in data and to generalise problems in a way that is non-deterministic, and that can solve different problems based on some input.
"There are different flavours of this; you have AI systems that work with images or text, and then there are even robots that incorporate these sort of principles or technologies.
“So, in this case, the sign language translating app follows the template perfectly. You want to solve a problem: converting sign language or a video of someone signing into text. But there are many ways to sign hello - you can move your hand in a similar sort of direction, but at different speeds etc. There’s a lot of variability to it. So you have a very complicated problem that you want to convert to text. And then you have to collect data, to train this model, to give it examples etc. And if it works correctly, it's going to learn from the data what it is that makes a sign mean hello or goodbye.”
Deepfakes
“Deep fakes are fake or synthetic images, videos, audio that record an event or a speech that never happened. I'm specifically mentioning audio, video and images, because that’s the understanding that people have these days. Back in the day, deep fakes, really referred only to video. But over time, it has really expanded to any sort of form of fake content.
“It's super easy to make deepfakes. All you need is access to the internet. There are these apps or services where you just type a prompt description of a scene or a speech, and then the AI model creates that scene or speech even though it never happened.
"You can include, your face, your voice, or faces and voices of people you've never even met. All you need is about 5 to 10 seconds of audio as a sample of the person's voice and one picture of their face. That's all you need to create a fake video of them saying whatever you want.
"We have seen a lot of the negative aspects of deepfakes unfold, at least here in the US. In Europe, the adoption for fraud has not been as pervasive yet."
Why is this dangerous or troublesome?
"Well, they can be fun. There are funny cameos you can make with this. But at the same time, you can use this to spread disinformation, to commit fraud, to create false evidence or incriminating documents. And to me, it just doesn't seem like the benefits outweigh the actual downsides.
"We have seen a lot of the negative aspects unfold, at least here in the US. In Europe, the adoption for fraud has not been as pervasive yet. But I think it's modelling a similar pattern and is just delayed compared to the US. So I think it will reach Europe, too.”
So, deepfakes, not AI in general, have more negative than positive impact?
"We're entering, and in many ways we've already entered, an era where you can no longer trust what you see online."
“In my opinion, right now, I see them having more downsides. We're entering, and in many ways we've already entered, an era where you can no longer trust what you see online. And it's really difficult to understand what actually happened, who actually shared it, and all these different parameters that are critical to navigating the online world.
“Even myself, as someone who spends a lot of my day looking at deepfakes, studying deepfakes and making deepfakes… Just yesterday, I got a call from someone and I was not sure if the person on the phone was AI or if it was a real person. I had no idea if this was AI. You really need additional tools and support to be able to understand these things. Our own hardware - our brains and ears - can no longer discern what's real or fake, at least if it's done at the levels that are available to us right now.”
I was going to ask for tips on how to detect deepfakes, but I guess there really aren't any?
“No. If you asked me like a year or two ago, I would say that there are these things like six fingers or four fingers, or sometimes the AI models would produce inconsistent earrings etc. But, at this point, it's really not possible to detect it without additional tools, if it's done at the state of the art level.”
To regulate, or not to regulate, that is the question
Do the dangers associated with deepfakes and AI make you want more regulations on who and how gets to use AI? Or do you think that there should be complete liberty?
“I definitely think that there should be some regulation. And perhaps more importantly, I think there should be accountability for the big companies that create these systems and enable anyone to use them without the consent of those who are actually depicted in the generated images, audio or video. In the US, there's something called Section 230, which gives companies like Google, Facebook and others, a lot of leeway. Whatever content is generated and shared on their platforms is not considered something that they're responsible for. The responsibility lies on the actual user who made the content on their platforms. To some extent this makes sense, but I think there needs to be more nuance and this needs to be updated to meet the current state of affairs in AI and media.
“I think that in Europe there have been good steps towards regulating AI. I will say that we need to be careful about regulating only certain domains or parts of the industry that actually have pretty objectively, I think, more negative than positive impact on society. However, I think we do want to make sure that the EU can be competitive against the US and China in AI and core research and capabilities. So, I think there's a balance that needs to be struck. But yes, the upshot is that I would like to see more regulation of this.”
A positive outlook not only for young people
We have shifted more towards AI as a whole. You mentioned that there are some positives aspects to AI. What are they? Are you working on anything that gives you hope for positive developments in AI?
“I should say that overall, I'm very optimistic about this. I think that’s just a part of my personality; I'm just optimistic about things and it might not be fully supported by evidence. However, I do see a lot of very good progress around me in the Bay Area, in Silicon Valley, but even here on the East Coast, where a lot of young people are using AI to basically assume leadership and just break norms.
"I think AI is great because it enables anyone to re-write the script. It's so powerful that you can, in many ways, change industries, norms, narratives, etc."
“I think that for a long time, we have had an environment where ‘boomers’ and people who have been running the show for a long time haven’t made room for younger generations to step in and address urgent issues like the climate crisis, societal polarization and all these other things that are just deepening and being overlooked. I think AI is great because it enables anyone to re-write the script. It's so powerful that you can, in many ways, change industries, norms, narratives, etc. And because younger generations are more native to it, we're able to actually use it to that end. So that's really positive, I think.
“In terms of direct applications - not just more cultural or broader implications of AI - I think it has a lot of cool direct impact on drug discovery and research in medicine. You can use AI to automate a lot of your testing and exploration. You can do the same in pretty much any science that uses experimentation because you can simulate a lot of things and you can get faster to more promising configurations. I think that's very cool.
“I think it's great for search and for finding information and references, if you're well-intentioned. It can also be misused in that way, but I think there are many cool use cases.”
Anything that you are working on at the moment?
“I think our research is a bit more fundamental. We are trying to enable anyone to understand how AI systems work and why they decide in the ways they do. Because right now, the systems are black boxes – there’s input and there's output, but everything in between is a mystery, to a large extent. I think that if you can get value out of something, like a tool or a product, you don't need to know how exactly it works, but there are limitations to this.
“For example, I can use a plane and get really good value out of it without knowing how exactly it works. But I do rely on someone understanding how it works - there's an engineer somewhere who checks the plane before we actually take off. Right now there are no engineers for AI. And that, to me, is troubling because if we're going to enable these systems to research drugs and vaccines, or to decide who gets a loan and who doesn’t, then I do want to make sure that we can audit them and understand how they make decisions. So, that's one thing we're working on in different ways and we're seeing a lot of progress. I think that's cool because people in different domains can then use that research. So, that's one thing.
“And then we're also working on ways to give people more control and agency if they’re worried about deepfakes or misuse of their data for AI. But these projects are still in progress, so I'll be able to share more about those soon.”
ChatGPT
Where does ChatGPT that fit into the picture?
“I think ChatGPT is great. It can be super useful. Overall, I would recommend to constantly probe its limits and see where it fails, because that's a really good check on what you want to be using it for and where you want to be more careful.
"These systems are super powerful and can be super useful, but fundamentally, they're trained to satisfy your requests. And when they don't know something, they will make things up just to satisfy your requests."
“I know people probably know this by now, but there are hallucinations. People can get very easily deceived into some nice-looking or formal-looking citations or something, but it turns out it’s actually the chatbot making things up just to please you or make sure that you're satisfied. These systems are super powerful and can be super useful, but fundamentally, they're trained to satisfy your requests. And when they don't know something, they will make things up just to satisfy your requests. That's really important to keep in mind and work with it in a particular way where you do check its sources and double-check what it actually says.
“But overall, I think it's great. I think it can be super useful for coding, for searching through vast batches of data or papers. But I would be careful. I would be cautiously optimistic about how you can use it for your work.”
Czech AI and the American perception of European regulations
We have been talking a lot about developments in the US, but you are from Prague and grew up in a Czech environment. How does it compare? Is the Czech Republic on the ‘AI map’?
“I think there are great people who came back from the US and are trying to launch really promising things. I'm thinking of Tomáš Mikulov and his start-up, BottleCap AI. I think they have something really cool going on - they could be very much playing the A-star sort of tier game. But beyond those very few exceptions, I don't think the ecosystem is there.
“I'll say this not because I think that's how it should be or because I like the status quo, but just because it's how people around me perceive Europe and Europe's role. A lot of people in the Bay Area are looking down on Europe and its ability to compete in AI. The things they see are regulations that ensure that bottle caps are now attached to plastic bottles. That's one thing people know about the EU and its regulatory landscape. There's been a lot of memes around that and a lot of comparison about how the US is building rockets that can return to Earth, while the EU is making sure that plastic caps are attached to bottles. It's a joke, but I think it just goes to show how people are thinking about the EU market.
“I've also been an advisor to a bunch of start-ups, and I will say that a lot of them are not even considering the EU market because they feel like there's too much regulation and it's not worth it. To them, the money is in the US, in Asia, and maybe in Australia, or the UK, but not really in the EU. And so they just completely disregard it.
“Right now, I think there is a risk of being rendered completely obsolete and really irrelevant. I'm glad that there are people who are trying to change that. But right now, there's a lot of work that needs to be done.”
So there seems to be this contradiction: more regulations, as in the EU, are good, but they do not attract more research. Meanwhile the US with fewer regulations, attracts a lot of research.
“Yes, that's a really good point. And that's why I was trying to be cautious about this. I think the EU has a great record of protecting its citizens against different forms of abuse. And I think when you know that there's an abuse possible through deepfakes, you should want to be in the business of regulating deepfakes - how they're made, who gets to make them etc.
“But core research, or core investments, need to be much more permissive. I think it's not just about regulations. I think even if you were to remove any regulations, there's just not enough capital in the EU to support innovation, research, start-ups and big projects. And you also need a bit of risk tolerance. In the US, a lot of people who have very preliminary ideas, and not a good setup of their company, just try to make a start-up or a venture out of their crazy ideas. And a lot of them will fail. That's just how it goes. But then 1 out of every n will succeed and that will drive the economy and will make up for all the loss of the other companies.
“In the EU, it seems like the narrative is much more risk averse. And so people are trying to invest smaller amounts, the Venture Capitals (VC) are more in control of what the company does. And from the VC’s perspective, I understand why they would want to do that. But if you look at it as a whole ecosystem, there's just not enough potential or not enough space to embark on these really bold, crazy ideas. And so if there's no space for them, they'll just happen in the US.”
Decision to study in the US
Why did you chose to study in the US?
“For me, it is the fact that the ecosystem is really there. I want to learn from the people who are shaping the industry – so, that's 100% why I wanted to go to Stanford and work with these people. I'll also say that there are other things that the Czech Republic specifically does or does not do to retain some of its talent. For me, it's the absence of gay marriage in the Czech Republic. I just don't want to live in a country where it is not legal. So, that's another factor for me. But once there is an ecosystem to work in, and the potential to start a family, then the Czech Republic is a place that I want to consider for my life.”
A positive endnote
Is AI a threat or a tool – will it take over humankind or will it cure all diseases?
"I'm very optimistic, because I think AI will give young people a much stronger position at the table."
“That's a great question to end the interview with. I think that ultimately, it's a great platform. It's a great enabler and accelerator of things, which is both good and bad. It really depends on your goals, your values and virtues. And so I think we can absolutely make it a very positive thing for the humankind, if we can be very deliberate about communicating its limitations, and be transparent and work across borders. I think right now, developments are very closely linked to individual countries and their own efforts to be the first to reach artificial general intelligence (AGI) etc. But ultimately, I think at some point, we'll realize that we need to collaborate.
“I'll go back to what I said earlier, I'm very optimistic, because I think this will give young people a much stronger position at the table. I think that's a big plus. And I have hope that we'll be able to steer the conversation and the whole landscape towards a good place. But time will tell.”
