The Powerlessness Cycle
To ensure AI is sufficiently safe, ethical, and legally compliant, we must immediately break the powerlessness cycle we are currently stuck in | Edition #313
Currently, there is no known plan, system, or enforceable framework to ensure that AI development and deployment occur in a sufficiently predictable, controlled, safe, compliant, and ethical way.
You read that right: there is no plan.
Most AI policy experts will not admit it publicly, but in practice, due to the lack of coordinated, proactive, and effective action, AI companies and the models they train are already shaping societal institutions, frameworks, and systems, the job market, our prospects for the future, and how we speak, write, read, think, and live our lives today.
Even if only indirectly for now, AI is already in the driver's seat, shaping our future, while we are stuck in a powerlessness cycle.
Below is a more detailed picture of what is happening, along with some thoughts on how to break this cycle:
How the Powerlessness Cycle Works
Over the past three and a half years, I have seen this cycle repeated multiple times:
1. AI capabilities are advancing fast: New AI models, systems, programs, interfaces, features, protocols, and computational layers are launched almost daily, creating new technological and cognitive layers that are aggressively embedded into societal institutions, systems, frameworks, and programs;
2. Fast and aggressive adoption is prioritized over everything else: Fast adoption is prioritized over literacy, ethics, and governance; people, institutions, systems, and frameworks do not have enough time to be adapted and reshaped in an orderly manner; there is a sense of chaos and meaninglessness; new risks emerge, people get harmed; it is difficult to react timely or plan for the future with existing tools that have already become outdated; a widespread sense of chaos and insecurity spreads;
3. Policies and laws are late, reactive, and erratic: After chaos spreads or people get harmed, policymakers, lawmakers, and decision-makers are reminded that they need to act. Due to pressure from the electorate, every few weeks a new bill, policy proposal, or general guideline is published in an attempt to address the emerging risks and changes. Most measures come too late and are designed in ways that will make them obsolete soon;
4. Policies and laws are ineffective at dealing with widespread AI-driven societal disruption: When a policy or law is fully implemented, it is often already too late, and the challenges it was supposed to solve have already evolved or taken a new form that cannot be addressed through the same mechanism. The cycle then starts again; go back to item 1 above.
-
Let us look at an example of the powerlessness cycle playing out in practice:
Case Study: Europe and the EU AI Act
Take the AI Act, for example.
So much time, effort, and taxpayer money have been poured into it over the past few years to ensure that Europe would be the first continent to have a consistent and robust legal framework for AI.
I have been teaching the EU AI Act over the past three years. If you ask me whether it will be effective in regulating and governing AI in Europe, my answer is: probably not.
Two years after its enactment, it has already undergone a major review process that included amendments and the postponement of a large portion of the rules.
The EU standards for AI, which would accelerate compliance, are not yet ready, and there is uncertainty about when they will be.
Given the lack of broader future-proofing mechanisms, much of the current text of the EU AI Act will likely become obsolete in the next few years, or even before the new, postponed enforcement dates arrive.
Enforcement authorities will face major challenges in implementing the rules, especially given the AI race and the pressure on the EU to be more innovative and competitive in relation to the U.S. and China.
Lastly, companies are finding all sorts of ways to circumvent the EU AI Act and avoid having to comply with specific rules, rendering it ineffective, especially in the areas it was supposed to regulate more strictly.
How We Got Here
Democratic lawmaking and policymaking take time, and over the past three and a half years, authorities have not been able to respond to the pace of AI development in a timely, orderly, and comprehensive manner.
Let me be clear: law and policy have traditionally been behind technological progress, especially over the past two decades.
It is unrealistic to expect lawmaking and policymaking to proceed at the same pace and in the same timeframe as technological development, and that is not a problem if the gap remains manageable.
However, law and policy have never had to address the pace of technological acceleration and societal disruption we are experiencing now, especially when the goal is literally full societal disruption.
In less than four years of the generative AI wave, we have seen cognitive automation lead to the near-complete disruption of learning, teaching, writing, research, and coding, as well as in how people talk, express themselves, think, create art, and develop intimacy. It has also partially or fully reshaped various professional fields in a very short period, leaving many unsure about their own ability to make a living.
AI has not yet helped any hostile actor fully take over a country’s power grid (though it does not seem far off).
It has also not autonomously decided to infect all humans with a deadly pathogen (though it recently gained the ability to create new pathogens).
Recently, we have also experienced scary incidents, such as AI models escaping training environments and hacking into other companies’ systems (OpenAI, Anthropic), suggesting that, to some extent, sci-fi-like scenarios like these are becoming increasingly likely.
Even though there seems to be a serious new security incident almost every week by now, our ability to control AI has not improved.
To make things worse, in recent years, AI development has become closely tied to national security and military supremacy.
The stakes in the geopolitical arena are high, and global powers will do whatever they can to achieve what they consider AI supremacy, as the U.S., for example, has made clear in its AI Action Plan.
There are currently not enough political, economic, or technological incentives for countries to take decisive, proactive, and comprehensive AI governance actions (as MIRI, groups like Pause AI and Stop AI, and individual politicians have called for).
As grim as it might sound, it is unclear whether we will ever find the right conditions and the geopolitical appetite to govern AI predictably and comprehensively, and to ensure that future developments and incidents do not harm people and society.
How We Escape From It
One could read what I wrote above and ask: “So what? We have always found ways to address new challenges, risks, and incidents; that is what we will do over and over again. That is how AI governance works.”
That is an optimistic thing to say, and it is nice and healthy to be optimistic.
But we must look at the facts:
Today, every single frontier AI developer’s goal is to create smarter-than-human superintelligent AIs that will autonomously train other smarter-than-human AIs (recursive self-improvement) and fully overtake society from the inside out.
It sounds like a sci-fi movie or book, but that is literally the plan, and hundreds of billions have been invested in this project.
Zuckerberg and others have recently been calling for “superintelligence for everyone.”
Even though it sounds like a nice and noble goal, it's much more complicated than it seems, as I wrote yesterday (click the post below to read it in full):
Superintelligence is a form of power, and for the first time, we will have people, organizations, and countries with access to tools that enable them to function in an order-of-magnitude smarter and more powerful way than others.
Superintelligence will not be homogeneously distributed, just as power is not.
Failing to break the powerlessness cycle in AI will, in practice, lead to never-before-seen situations of power asymmetry, violence, and oppression at multiple levels, involving people, organizations, and countries.
The later we break the cycle, the more difficult it will be to govern AI and the real-world challenges it will cause.
We have to acknowledge that AI and the race for superintelligence pose never-before-seen existential challenges that must be addressed with innovative, never-before-seen policy, legislative, and governance mechanisms.
I have been thinking a lot about the topic and how to create the right legal, political, and administrative environment to govern AI effectively and proactively (instead of merely reacting to emerging AI risks).
These are some ideas that have come to my mind on different occasions while teaching or writing about AI governance, which together form a plan to escape the powerlessness cycle:
Countries should have a continuously monitored AI governance plan. Many countries today have a strategic plan for AI, focused on their broader economic and technological strategy, such as America’s AI Action Plan and Europe’s AI Continent Action Plan. These plans focus on AI development, deployment strategies, and how to remain globally competitive. Countries should also have a coordinated plan focused on safety and governance that has specific targets, guidelines, and triggers, and that describes how data should be collected, monitored, and assessed. The plan should have calls to action designed for specific authorities and institutions to ensure that if specific disruptions, deviations, or risks occur, there is a plan of action. The plan should be highly detailed and specific, focused on the scope of each administrative unit, with ongoing updates, reviews, and assessments to ensure the targets still match the reality on the ground. The plan ensures that there is a path and a destination and that countries are not merely reacting to emerging AI risks.
We must create binding AI laws at a much faster pace. Traditional lawmaking involves a long, multistakeholder process that has been designed that way because the resulting law would make sense for many years, maybe a decade or more, so legal certainty and broad democratic validation were prioritized. With AI, we should assume that the technological reality will be radically different in one to two years, but we still need legal certainty, predictability, safety, and control mechanisms. The time to discuss a new, binding rule or mechanism should be much, much shorter than under traditional lawmaking procedures, no more than a few weeks, and countries should establish fast-track lawmaking mechanisms to regulate AI.
Every single policy or law on AI should have a scheduled review date, which should occur no more than one year after its enactment, and often only a few months after the law is published. AI advances much faster than any other field before it, and administrative procedures and review schedules are not made for this pace of change. Lawmakers and policymakers should create laws and policies knowing that they might not make sense one year later (but they are still necessary to govern AI effectively).
Every single law should contain future-proofing mechanisms. Preferably, every single material provision should be future-proof. By that, I mean that if the technological reality changes faster than expected, there should be fast-track mechanisms to update any single provision, making it suitable for the new reality on the ground. The lack or scarcity of future-proofing mechanisms creates legal uncertainty as it becomes unclear if a provision is still applicable if the reality it describes no longer exists, or if the rest of the provisions still apply if one or more of the provisions have been made obsolete.
We need to popularize AI governance triggers. Every single institution, organization, and government body at every administrative level must understand how AI impacts it, create monitoring mechanisms, and establish AI governance triggers to provide an alert when something seems to be going off track. Currently, this level of literacy, awareness, and monitoring does not exist in most places, especially not in a fully decentralized and specialized way. In specific government bodies and institutions, there are safety triggers that are connected to existing cybersecurity mechanisms. We must expand that, track AI’s impact more closely, and monitor for specific and contextual disruptions so we can act accordingly.
There is much more to say and do about this, and the discussion should continue. I would be glad to read your additional input and comments below.
I am confident that merely reacting to what happens in AI week after week and hoping for the best is not rational and will not work. This is not AI governance.
We must break the powerlessness cycle.