The EU have stated their intention to implement some pretty stringent laws in their AI Act. Even Macron has warned that EU AI legislation will hamper European tech companies compared to rivals in the US, UK and China. Mistral is a very successful French company who may be hit by the AI Acts attack in foundational models.
It will, of course take forever, there is a two year process, with three different drafts from here from three different institutions – Parliament, Commission and Council. It is all a bit of a technocratic mess, with lots of potential bun fights.
Here are the problems:
A ban on emotion recognition may well hamper useful work in autism, mental health. In truth there wiggle room here, as it is absent in two of the drafts.
A ban on facial recognition will face fierce resistance from police and intelligence agencies, on cases as wide as deepfake recognition, age recognition in child protection, accessibility features, safety applications, terrorism, trafficking and child porn. Expect a lot of back-tracking and exceptions here.
The ban on social scoring sounds good but you are being assessed all the time for credit, loans, insurance and mortgages. In truth, social scoring Chinese style does not exist, so this is an unnecessary law tackling a non-existent problem.
The big one is the move on foundational models and their training data. This will be fiercely fought, hence Macron’s early move. Attacking foundation models is like attacking the wind not the boat and its destination. Ohers, such as the US and UK have urged a more application-driven, sector driven approach, that focus on uses not the underlying models. France, Germany and Italy has fought this, so I’d expect backtracking.
The Act has several other shortcomings in terms of definition and implementation. It sees AI as a fixed entity. AI is intrinsically dynamic, as it learns and morphs as it operates. This makes it a moving target in terms of the object of legislation. On top of this AI is an integrated set of services, networks and organisations with various data sources. If I am a small AI company I am likely to be using various foundational models and delivery services from OpenAI, Microsoft, Facebook and Amazon, all as cloud services. One general-purpose system may be used for several separate services. It is difficult to see how responsibility can be assigned.
While the rest of the world gets on and uses this technology for astounding benefits in learning and healthcare, the EU seems determined to become a mere regulator. Not just a regulator but one that loves complex, red-tape driven solutions. Billions, if not trillions, in terms of productivity growth and innovation may be a stake here.
My guess is that the act will facing challenges in terms of reaching consensus among member states, given the diverse interests and viewpoints on AI regulation. There will also be difficulties in implementing the guidelines and regulations proposed in the AI Act, due to technological complexities or resistance from AI stakeholders. The act will therefore need revisions or updates to address emerging AI technologies and scenarios that were not initially considered.
The final problem with EU law is its fixity. Once mixed and set, it is like concrete – hard and unchangeable. Unlike common law, such as exists in England, US, Canada and Australia, where things are less codified and easier to adapt to new circumstances, EU Roman Law is top down and requires new changes in the law through legislative acts. If mistakes are made, you can’t fix them.
At only 5.8% of the world’s population, there is the illusion that it speaks for the whole world. It does not.
PS
When the Pope called for a global treaty on AI, we have surely reached Peak BS on AI and ethics. The Catholic Church put Galileo:house arrest for rest of life, they burnt Bruno at stake for scientific heresy. Their banned books list Galileo, Descartes, Voltaire, Sartre & Kant only in 1966! We have surely reached Peak BS on ethics & AI?
No comments:
Post a Comment