Do we need a legal framework for safe AI?
It is now clear that artificial intelligence (AI) will play a major role in the coming years. Accordingly, many groups - and nations - are striving to dominate AI or play a key role in shaping it. This also includes approaches for a uniform legal framework. This brings us to the subject of compliance, an important topic for systems engineering.
The following is about two questions: Do we need a legal framework at all, and if so, what should it look like? Elon Musk at least took a clear stance on this five years ago.
Is AI dangerous?
There should be a legal framework for things that are dangerous. So how dangerous is AI?
Unregulated AI could be the "greatest risk to our civilization"
Elon Musk, NGA 2017
As early as 2017 Elon Musk the need to create a legal framework for AI as follows: One problem in business is that competition is always breathing down your neck. Without a legal framework, taking safety into account is often a competitive disadvantage. Through appropriate regulation, all competitors must abide by the rules.
But is AI dangerous at all? That doesn't seem to be the case at the moment. But by the time the danger becomes apparent, it could already be too late. Because implementing a legal framework, especially a global one, takes years. This is beautifully expressed in this tweet:
In other words: We may not know whether we need a legal framework. But for the same reasons that we take out insurance, we should also set up an AI legal framework.
With a legal framework for world domination
A legal framework, should it achieve sufficient acceptance, also represents power for the controlling economic bloc. The EU's efforts to create clarity here are particularly visible to us. Under the name Artificial Intelligence Act (AIA) the EU is working on a corresponding draft. AIA is only one part of a AI package that guarantees the safety and fundamental rights of people and companies. should. Depending on the risk classification, a system must meet certain security requirements.
China is also trying to create a legal framework for AI. However, unlike the EU, the focus here is on data protection, not security. China is trying to position itself as a leading nation when it comes to AI. The legal framework is therefore primarily about legal clarity for commercial use.
The topic of data protection is addressed comparatively little in the EU in the context of AI. This is certainly due to the success of the GDPR which data protection already has a clean (albeit rather bureaucratic) legal framework.
The USA is also working on this issue, albeit nowhere near as visibly as in the EU or China. What is more interesting is that the USA is working closely with the EU to develop a common legal basis to create.
What should a legal framework look like?
It should come as no surprise that the ideas from both China and the EU are met with skepticism for various reasons. For example, human rights organizations criticize that the legal framework vague formulations and loopholes contains. Other groups are concerned about unreasonable requirements.
But much more worrying is the Criticism from Wired as to whether the EU even understands what AI actually is. Because if a legal framework is based on false assumptions, then it is worthless. According to the EU, AI is software developed using certain techniques that leads to goals set by humans. However, even in this broad version, there is no reference to clear human responsibility for these results. Several EU member states are currently trying to soften this definition even further by only considering self-learning AI systems. This opens the door to classifying a system as an AI or non-AI system, depending on what is more convenient at the time.
This also makes it clear where a legal framework should actually start: The safety of systems for people. This raises the question of how useful an independent legal framework is. After all, we need a domain for the application of every AI system. With regard to functional safety, we have ISO 26262 (automotive) or IEC 61511 (industrial processes), which are all based on IEC 61508 based. It certainly makes sense to set generic, risk-based guidelines for the various AI technologies, which then need to be fleshed out in the context of a domain. But this does not fit into the overall picture suggested by the hype surrounding AI.
Image by Garik Barseghyan from Pixabay
