After DeepSeek, France Should Have Called Off The AI Action Summit

Before DeepSeek, the France AI Action Summit to be held in Paris from February 10 – 11, already looked like it would make little difference, because they did not seem to revise any assumption about AI safety in their agenda [Public interest AI, Future of work, Innovation and culture, Trust in AI, Global AI Governance]. You can download the program here.

Since over two years that AI has been in the news, there were several initial safety, alignment and regulatory assumptions that were made that were subsequently obvious would be the wrong direction.

Now, after DeepSeek, France should have called off the summit and organized something smaller, at a later time, just for their homeland. There were lots of prior conclusions that it was better to build bigger AI somewhere, so that it is not built elsewhere. However, with DeepSeek, it should be clear that the race is no longer for an AI advancement edge, but for AI safety superiority.

Whatever country can build the safest AI wins. If a country has a powerful but unsafe AI, with the capability to answer any question or be easily jail broken to do so, it would enable vices within the country against the country, deeper than it can, for some distant adversary.

There are lots of answers that AI can give that could result in negative actions that will be effected locally first, in whatever country, making unsafe AI a risk to the country. Also, access to unsafe AI elsewhere is also a risk. It is simply not about blocking access because those can be easily bypassed, once actors know that a less safe AI is out there.

The national security project for France or any country is not just to build safe AIs, but to have safe AI monitor unsafe AI from any source on the networks within the country, or be able to explore safety models that can be used as regulatory tools, so that AI models can adopt them for broader safety. AI safety tools, as viable products, would be the most profitable as well, as AI percolates social and productivity purposes.

Some countries in Europe are rushing to ban access to DeepSeek when they neither have a matching AI nor a safe AI to counter. What they could do is to aggressively fund labs in most universities in their countries to work on general AI safety and alignment, to have technical defense in a vital area, not to have nothing and assume AI is some sort of social media or website. Also, AI from a seemingly unallied country does not automatically mean bad and from an allied country good. AI models from a friendly nation can replace jobs or be used to make AI voices or fake images or videos, or generate malicious codes. It is not necessarily about where AI comes from, it is about what it can irreversibly and consequentially wrought, in general, especially for those who have nothing in safety or alignment.

The EU has an AI office, which seems for regulation more than for technical advances. DeepSeek has already made that office irrevocably obsolete. The UK has an AI safety institute that was evaluating AI models, but the AI companies would not give access to the weights, just in private, for proprietary reasons and to avoid other nations asking. The UK AISI were careful not to irk the AI labs. DeepSeek made their weights open, not just privately, for everyone. The UK AISI also does not seem to have anything promising in the lead for AI safety for over a year since inception, for the UK homeland—at minimum against fake AI videos, images, or voices—questioning their strategy, direction and innovative might. Several countries launched AI safety institutes after the UK, they possibly do not seem to recognize how important safe AI is to their nations, in spite of current misuses, the future and the possible emergence or a kind of unexpected model like DeepSeek. France, the host, however, does not even have one.

DeepSeek is not a wake-up call to ban semiconductors, or access to the model, or to accelerate AI development, it is a wake up call to pursue intense research in every possible direction on general AI safety. Some star AI professors already said no one knows how to control powerful AI systems. This means that every possible channel, from theoretical neuroscience to math, new deep learning architectures and whatever areas, should be explored in priority research, towards technical development, by serious countries for safety.

Governments would continue to make laws against misuses of AI, but AI is an intelligent tool, where evasion is far easier. Penalization directly for the AI tool responsible, tracking the tool, its outputs and preventing it from causing harm would be better technical channels for safety than laws against people—as AI ploughs on. Those laws may apprehend some, but many more may elude them. Backing up laws by technical tools would give hope to AI safety, alignment and regulation.

Even for workforce security, should AI replace people, it is also a major risk. Pursuing novel economic paths is another important research that is not simply universal basic income or the future of work.

The AI Action Summit would proceed in France and likely to have some major announcements. Just like the last AI safety summit in the UK did little as well as the one in Seoul, the surprises ahead could quickly make this summit unremarkable.

There is a recent feature in Le Monde, Artificial intelligence: The first measures of the European AI Act regulation take effect, stating that, “The first measures of the European AI Act will come into effect on Sunday, February 2 –coincidentally, just a few days before the Summit for Action on Artificial Intelligence (AI) on February 10 and 11 in Paris. Although the symbolism is strong, for the time being, this first part only concerns certain prohibited uses. In concrete terms, this Sunday will see a ban on certain uses of AI deemed unacceptable by the AI Act. These include social rating software, whether private or public, such as that used by China, individual predictive policing AI aimed at profiling people by estimating their propensity to commit offenses, and also emotion recognition at work or school, to analyze the behavior of an employee or pupil. Also banned are the exploitation of people’s vulnerabilities, manipulation or subliminal techniques.”