Artificial Intelligence– Regulatory debate pt. IV

Europe fares better than I expected.

• Unless I missed something, the preliminary deal on how to regulate AI in the EU looks far more realistic and business-friendly than I had previously feared especially given how similar the original pronouncements of the EU and the CCP were.
• The EU has reached a deal on AI Regulation which will now need to be approved by the member states and the EU parliament to go into force.
• The proposed regulation falls into two main categories:
  • First, basic transparency: which applies to “general purpose” AI systems: by which the EU means large language foundation models that can be fine-tuned to be able to do a wide range of different tasks.
  • In practice, this means almost all generative AI, but interestingly, models that are free and open-source are exempt from this requirement.
  • The requirement is to have an acceptable-use policy, a record of how the model was trained, a record of the data used and a policy to ensure copyright is respected.
  • This is a major departure from how the EU was communicating a few months ago when it was looking at requiring models to have certain characteristics which are impossible to achieve.
  • These included things like requiring the model’s creator to be able to explain why the model does what it does or ensuring that it won’t make stuff up or include biases in its responses.
  • The nature of deep learning algorithms makes almost all of this practically impossible which is why China’s much harder line will slow its development of AI down.
  • This is another example of how the Chinese priority of national security and loyalty to the party has done more to hinder Chinese progress towards technological independence than the US Department of Commerce has managed to achieve.
  • This proposal from the EU looks to me to be fairly realistic in terms of what can and cannot be achieved and should also not be too onerous to comply with meaning that small companies are not meaningfully disadvantaged.
  • Second, “systemic risk” which will only apply to the largest models which have over 10,000 registered business users in the EU.
  • The threshold is proposed to be set at models that use more than 10 septillion operations per second and it is thought that only the likes of GPT-4 or Google’s Gemini will meet this threshold.
  • Here the requirements are a lot more onerous with vague requirements such as “mitigate systemic risks” and “adequate cybersecurity controls” which will be more difficult to comply with.
  • Interestingly, none of the European start-ups currently make the grade for this extra level of regulation and I am pretty sure that there will be some blowback with regard to protectionism.
• While I am not a big fan of regulation generally, this is much better than I had feared, and the real winner here is Meta Platforms.
• This is because its foundation models are all available in the open-source community meaning that as the rules appear to have been written, it will be exempt from this regulation in its entirety.
• Consequently, I am more optimistic about the development of generative AI in the EU than I was while the negotiations were going on.
• However, I still think that the best regulatory environment will be a low-touch system that is cheap and simple to comply with and targets restricting access of bad actors rather than the technology itself.
• I continue to think that the machines are as dumb as ever, but their size and complexity have greatly enhanced their linguistic skills even if they are simply calculating the probability of words occurring next to each other.
• This creates a convincing illusion of sentience which leads people to anthropomorphise these systems which in turn is what I think makes them much more capable of being used by bad actors.
• Hence humans remain in far more danger from other humans than they are from the machines, and it is this that I think any regulation would be best serviced to target.