Dario Amodei grabbed headlines on Saturday when he released a paper calling for a slow down in the pace of AI development and this morning the UK Parliament Joint Committee on Human Rights released its “Human Rights and the Regulation of AI”. I’ve spent the morning reading this report and tracking back through Amodei’s last few essays on AI, as well as pondering the intersection of these calls with actual current AI law – in particular the EU AI Act.
In Amodei’s most recent piece “We Must Pace the Frontier” he writes about the Anthropic vision to build AI that sees humanity benefitting from AI that doesn’t sit in the hands of authoritarian powers, whilst increasingly recognising that “building it too fast is reckless”. As he announced on Saturday, the essay urges the AI industry worldwide to adopt a three-part plan that would see them all “prioritize caution over speed and prudence over profit” but even more crucially, that would require developers to pace the rate of advancement in AI capabilities specifically “so that risk prevention has time to keep up”. The notion that AI would build itself in a process of ‘recursive self improvement’ has always been the stuff of sci-fi nightmare, but Amodei recognises that this point was reached “roughly this summer”, (ie already) and that “greater capabilities” coupled with AI “malignment” creates the perfect storm for AI to be used to cause massive economic disruption, and even weapons against humanity, hence the urgency of his call.
Amodei acknowledges that “pacing” was discussed in 2023, but that it (apparently) made no sense then because AI was “not powerful enough to act as agents in the [real] world …, and were not capable of significant deception, manipulation, cheating, or cyberattacks.” To say that just three years later the boss of the (financially) largest AI business in the world thinks AI now has these capabilities is worrying enough, but that his call was answered with unanimous consent by the Tech bros means regulation cannot wait. Sam Altman of OpenAi posted “I agree with Dario”; Elon Musk agreed , “Dario is right”. Whether there is a perception that there is little that wins votes in AI regulation or there has been a giant international political poker game being undertaken by Governments wanting their national developers to progress as far as possible before guard rails were instituted, politicians must step up now and demand a place in the pacing strategy because, as and Clement Delangue of Hugging Face notes “It’s now clear that alignment is critical and won’t be solved behind the closed doors of a handful of frontier labs!”
So where is the “Law of AI”? You are more likely to discover the position of it in Asimov than in Hansards, but the Council of Europe Framework Convention on Artificial Intelligence and Human Rights, Democracy and the Rule of Law (to which the UK and the US are signatories, but which is not yet in force) and the JCHR Report make a start. The Framework Convention would largely facilitate the ‘pacing’ Amodei outlines, requiring signatories to it to adopt or maintain measures to: • respect human dignity and individual autonomy; • ensure adequate transparency and oversight; • ensure accountability and responsibility; • ensure that activities respect equality and non-discrimination; • ensure privacy and personal data rights are protected; • promote reliability; and • enable safe innovation. The co-ordination from and within industry that Amodei foresees is an essential factor in the success of pacing and to preserving human rights in an AI world.
In the UK guardrails for AI development, adoption and use, for example in medicine regulation, data protection, competition, product safety and Algorithmic Transparency Reporting Standard (ATRS) is sector-specific. Whilst ATRS requires transparency in the public sector, there is little transparency in the private sector and the JCHR Report cites many criticisms of this. The common theme is that risks to human rights, business and security exist at every stage of the complex AI lifecycle. UK laws that could apply to AI systems do so primarily at the point of deployment, making deployers the primary holders of responsibility for harms arising from AI systems but providing regulators with no actual power to evaluate those AI systems prior to release. Additionally, the fragmented legal landscape and uncertain legal liability for AI-generated harms (para 134) all mean that legal regulation of AI in the UK is currently inadequate
The EU AI Act almost certainly doesn’t get regulation or oversight right either. Over the summer I’ve sat in various briefings and training sessions to better understand the obligations placed on organisations under Art 4.1 of the EU AI Act which requires all organisations to support the development of AI literacy of their staff and other persons dealing with the operation and use of AI systems. This legislation doesn’t only apply to EU based AI deployers, but to training in all AI that placed on the EU market or used in the EU or that has an impact on people located in the EU and which extends to all output from use of that AI. This is a massive assessment for businesses to undertake and it is not a on-size fits all exercise, but one that requires assessment of the use, application, position, knowledge etc of the user, the program and the impact. I can’t help feeling this section of the AIA is wide of the mark; auditing the human experience of interacting with AI and not the AI itself, more legislation that is obsolete already at the point it is implemented.
The JCHR Report asserts that the “problem of pace, where AI develops more quickly than regulation designed to address its risks” can be overcome with a “principles-based approach” where codes of practice can be applied more quickly than primary legislation – essentially what Amodie recognises in his 3-step plan.
Saturday’s call for pacing by Amodei has spooked the financial markets today. With President Xi Jinping also calling for open-source technology sharing and renewing hopes of a global AI governance pact over the weekend, it seems that industry and governments are in fierce agreement (albeit that Donald Trump has accused “negative forces that are bringing it up that shouldn’t be bringing it up”). If it feels like we’ve been having these discussions for years, we have. The question is whether there is sufficient urgency, momentum and integrity to meet these calls with actions.
Time will tell if we act before AI does.
3-Step Pacing: In brief Anthropic proposes:
-
Embedded Evaluators who verify the pacing commitments who have employee-like access to third-party evaluators verify adherence to safety practices, incidents, pipelines and processes.
-
Coordinated Standards for Safety and Progress Limits where frontier AI companies (within democratic countries) coordinate to establish common safety standards as well as limits on the rate of AI progress.
-
Global Coordination between democratic and authoritarian governments to verify compliance with all steps.
For more on the “pacing” proposals read Amodei’s piece here.
This update is for general purpose and guidance only and does not constitute legal advice. Specific legal advice should be taken before acting on any of the topics covered. No part of this update may be used, reproduced, stored or transmitted in any form, or by any means without the prior permission of Brecher LLP.
