Taiwan, whom Beijing claims as one of his provinces, faces permanent pressure from China. In addition to the almost daily incursions of Chinese aircraft in its area of air identification or boats near its coast, the island, which is de facto an autonomous policy, is also one of the most targeted countries by cyber attacks.
Passing through Paris on the occasion of the AI summit, Audrey Tang, traveling cyber-ambassador of Taiwan and former digital minister, is on the front line of this hybrid conflict. His experience on how to combat disinformation accelerated by AI is precious. L’Express met her.
What do you think is the most dangerous AI for the future: American AI or Chinese AI?
AUDREY TANG The most dangerous would be an AI guided by his ego, who considers himself a separate species, with a survival instinct. It looks like science fiction, but many people think that there is a significant possibility for something like this to happen within ten years. Of course, consensus is not a very high probability. But taking this risk amounts to playing Russian roulette. It is very dangerous.
There are actually two different races at current AI. The first is vertical: we try to increase capacities, to obtain more data and to develop larger, more efficient language models. But there is also another race, horizontal, that of the distribution and distribution of the Open Source [ou source ouverte].
It is time that instead of considering AI as an overpowered brain, capable of doing everything, we conceived it so that it is more oriented towards precise tasks (such as the translation or the summary of the document). For me, the answer requires the creation of language models [NDLR : les systèmes permettant de faire fonctionner les applications d’intelligence artificielle] compartmentalized and smaller, which can provide very intelligent answers, but without ego. In short, you have to imagine an AI with a Buddhist spirit [Rires.]. The question is not so much to know if the United States or China will win the race for AI, but to prevent the machines from winning it. It would be the worst scenarios.
How do you explain the rapid boom of the Chinese actor Deepseek, who managed to emerge in such a short time and at a cost much lower than that ofOPENAI And other models?
It is simply a LEGO brick placed at the top of a fully built tower on the previous work of scientists. Many people say that this brick has not cost dearly, but it is based on a technique called “mixture of experts” [NDLR : qui permet de faire des calculs plus rapides]developed partly by Mistral, and Mistral himself relies on Meta Llama [un modèle linguistique rendu publiquement accessible]and so on. Behind Deepseek, there is a whole open source ecosystem including R1, their latest language model, is only the last brick. It is not even the last, because, since then, Perplexity has taken this base and has made its own version. We cannot therefore really say that Deepseek did not cost dearly, because this AI is based on this tower, produced by other people.
To what extent can Deepseek become an influence tool for China?
If you do not want to use the R1 model, from Deepseek, there are now many other options, equivalent, even better, and at a lower cost. So, even if for a week, it was “the last Lego brick”, the latest innovation, this is no longer the case now. There is therefore no reason to panic about this.
It is also necessary to distinguish between the Lego brick, which is the R1 model, and the application that is Deepseek. The basic model can answer certain questions, but this service is censored [NDLR : le massacre de Tiananmen de 1989 n’apparaît jamais dans les réponses, par exemple, et Taïwan est décrit comme une partie “inaliénable” de la Chine]. Excessive dependence on this application is therefore quite dangerous, which is why the Taiwan government and its public sector prohibit its use. But if you want to download this Lego brick and study it, there is no problem.
Does China use AI to destabilize Taiwan and launch cyber attacks?
We have millions of cyber attacks per day in Taiwan: for more than a decade, we have been the country most affected by this type of attack. Since 2022, when the president of the House of Representatives of the United States, Nancy Pelosi, has visited Taiwan, we also observe hybrid actions, with cyber attacks and disinformation. For example, a cyber attack that blocks access to the site or the services of certain ministries in order to create a void conducive to fake newsfollowed by an informational attack which aims to supply advertising panels with hateful messages [NDLR : ce qu’il s’est produit en 2022, durant la visite de Nancy Pelosi]. Access to the results of an election can also be paralyzed, to replace all kinds of false rumors. But there are other hybrid tactics, such as the “accidental” cut [en 2023] Submarine cables between Taiwan and the Matsu Islands [qui font partie de l’archipel taïwanais].
How would you define the Chinese disinformation strategy and the threat it represents for Taiwanese democracy?
Authoritarian regimes always have the same speech: “Democracy does not work.” They say that “democracy only leads to chaos and polarization”, or that “only authoritarian regimes can control a pandemic”. At the time, our answer was to set up even more democratic tools, and to show that we had better results in the fight against COVVID.
How do you defend yourself?
We have also organized withdrawal solutions, such as low orbit satellites to ensure internet connection even if the cables are cut. Against cyber attacks, we have created threat detection and monitoring teams. We also use AI to spot them as soon as possible.
On the disinformation side, we do “prebunking“In order for citizens to be informed and prepared. Before an election, we thus warn the population that it will be confronted with fake news claiming that there is violence in certain polling stations, or that the results are rigged. These fake news end up being broadcast, but as we had talked about it upstream, they don’t really take.
We have set up a database that citizens can consult when a rumor becomes viral to compare it to others fake news who have already been debt [NDLR : réfutées]. We can then very quickly generate an answer using AI, but we also count on journalists, in particular, to perform the final verification.
Finally, the National Cybersecurity Institute is responsible for defending critical infrastructure. In this area, we are also collaborating with private sector players, including Taiwanese companies such as Trend Micro, but also international companies such as major cloud service providers: Microsoft, Google, Amazon Cloud, and many others. We organize exercises with these groups, with attack simulations.
Does the election of Donald Trump, which we do not know if he would defend Taiwan, make a Chinese attack on the island more likely?
Taiwan considers all its democratic allies as peers in the defense of democracy. If one of our allies is gangrenous by divisions, this is not good news for Taiwan. However, if we compare the last American election to that of four years ago, or even that of 2016, the result is clearer, there are fewer internal conflicts on the results. And it is rather a good thing for Taiwan.
Trump has already announced that he wanted the Taiwanese to pay more for their defense, and he threatens to tax semiconductors, isn’t that worrying?
Our Minister of Economy and our President have both indicated that there was reinforced communication between the Taiwanese government and the United States, because we are already spending a lot, and more and more, for the defense. Our partnership on semiconductors has always been a win-win agreement, because their design is done in the United States and manufacturing in Taiwan. This is a beneficial agreement for both parties, so we have to rely on the excellent relationships that we already have.
Faced with the United States and China, has Europe delayed irrattilutable?
If you build your AI from an open source, it is better to get late because many things have already been built and you can use existing components as you wish.
Is it possible to create a healthy AI?
The time has come for those who are very attached to ethics to impregnate the AI. To correct the movement that has imposed itself since 2016, when platforms like YouTube or Instagram have changed their algorithms so that users spend as much time as possible on the screens. They have created automatic learning systems intended to encourage people to continuously watch short videos. It has become a mental health problem worldwide, and a parasite for societies, which unfortunately adhered to this logic.
Today, people collectively awaken to the idea that AI can be regulated. We see everywhere in the world of laws, regulations, standards to supervise the use of social networks by the youngest or in classrooms. We must impregnate these algorithms with a different ethics, taking into account the moral traditions of antiquity or with the objective that people understand each other better.
At the same time, with these technologies, new threats appear: criminals create techniques of phishing [ou hameçonnage : technique consistant à se faire passer pour un organisme pour obtenir des données bancaires, NDLR] or other increasingly effective scams.
This is why the means to defend yourself against your scams must also be in open source. We must embark on a safety race: the more we work on this distribution of the AI in a safe way, by integrating it into each culture and each company, the less it will be likely that large laboratories or countries monopolize this technology .
Can AI be at the service of democracy?
There are actually two different races at current AI. The first is vertical: we try to increase capacities, to obtain more data and to develop larger, more efficient language models. But there is also another, horizontal race, that of the distribution and distribution of the AI.
The easiest way to put AI at the service of democracy is to ensure that people have access to specialized artificial intelligence, with small models of language, which they can master. I work for example with the intelligence project CIP.ORG collective. We are launching an international dialogue on AI so that people around the world can ask questions about how they want AI to behave. The objective is that AI can adapt to the needs of different communities according to their culture and standards.
There cannot be an AI universally aligned with democracy, because it has a very different meaning in each democratic country. This is why, instead of a universal AI that would colonize the rest of the world, we need pluralism, so that people who agree on certain very fundamental values, can translate them into different applications according to their culture, their legislation, local standards.
.