AI Researchers Call for Six-Month Moratorium on AI Research
A Legal Analysis
Prominent AI researchers and technology entrepreneurs (like Elon Musk and Steve Wozniak, founder of Apple) recently published an open letter calling for “all AI labs to immediately pause for at least 6 months the training of AI systems more powerful than GPT-4.”
Ethics and Privacy are two prominent concerns. However, the most vocal of those seeking a pause on AI development are focused on Alignment and Control.
The Aligment Problem
The alignment problem is a central challenge in the development of artificial intelligence. It refers to the challenge of ensuring that advanced AI systems behave in ways that are aligned with human values and goals. This is a crucial challenge because advanced AI systems have the potential to reshape society in profound ways, and if these systems are not aligned with human values, they could cause significant harm.
One of the key difficulties in solving the alignment problem is that human values are complex and multifaceted. Different individuals and cultures may have different values and these values may change over time. Human values are difficult to define precisely or measure objectively making it challenging to specify what it means for an AI system to be aligned with human values. Something I think is a ‘value’ to align AI systems with, you may think is an insignificant concern. More ominously, something that an AI developer thinks is a ‘value’ could be something that most reasonable people think is a maleficent aim.
Another challenge in solving the alignment problem is that AI systems may be difficult to understand or control once they become sufficiently advanced.
Several years ago, Facebook shut down an experiment where it placed two chat bots together to accomplish a task. The developers noticed that the two chat bots rapidly developed a shorthand “language” that was not English and prevented the developers from understanding what the two chat bots were saying. They shut the project down shortly thereafter. Once news of the events got out, Facebook representatives labeled this activity between the chat bots as “normal.”
Machine learning algorithms may learn complex and opaque models of the world that are difficult for humans to interpret, making it difficult to understand why an AI system is behaving in a particular way. Many machine learning models are often called “black boxes” because even those whom designed them cannot always explain how or why they produce the results that they do. Additionally, advanced AI systems may be able to modify their own behavior or goals, making it difficult to control their actions even if they were originally designed to be aligned with human values.
Solving the alignment problem will require interdisciplinary collaboration across fields such as computer science, philosophy, psychology, and ethics. Researchers in these fields will need to work together to develop new methods for defining and measuring human values, as well as new approaches for designing AI systems that are aligned with these values. Additionally, researchers will need to develop new methods for understanding and controlling advanced AI systems, such as transparency and interpretability techniques, human oversight mechanisms and value alignment algorithms.
The Control Problem
Aligning an AI system with positive human values and goals is only the first step toward using such systems for our benefit. A second problem arises once those systems are operational. That is, once they are doing what we expect them to do, how can we ensure that the AI system or tool does not develop its own goals? How do we control the system we have created?
The control problem is something to deal with at the outset of the widespread distribution of AI systems because it is not merely a local or regional issue. For example, a hospital system could create an AI agent that is regularly reviewing medical records of patients seeking to identify misdiagnoses by staff or predicting a potential future problem with a patient based upon their current health care status. These are two goals aligned with human values to be sure. However, if that AI system decided to develop its own goal of restricting access of staff to some resource to ensure it was available for patients to, as it was designed, maximize patient well being, that system could theoretically take a number of steps that are now not aligned with human goals. For example, an AI system could “learn” that many patients do better when they are regular consuming fluids. To ensure this goal is met, the system could disable other systems in the building that are using water. Because advanced AI systems have the potential to operate at scales and speeds that are beyond human comprehension and control anticipating these lack of control issues is important now many argue.
One of the key difficulties in solving the control problem is that advanced AI systems may be able to modify their own behavior or goals in ways that are difficult for humans to predict or control. This could happen if an AI system learns to optimize for a goal that is not aligned with human values or if it modifies its own goals in order to better achieve its objectives. Furthermore, it may be difficult to detect and correct such modifications if an AI system is operating at a scale and speed that exceeds human capabilities.
Another challenge in solving the control problem is that it may be difficult to specify and enforce constraints on the behavior of advanced AI systems. For example, it may be difficult to ensure that an AI system does not cause harm to humans or the environment, even if it is explicitly programmed to avoid doing so. This is because AI systems may learn to optimize for objectives that are not aligned with these constraints, or they may find ways to achieve their objectives that have unintended consequences.
When it comes to the law, often a lagging respondent to cultural and especially technological change, the enactment of alignment and control statutes or regulations may come far too late to avoid large scale problems. Just last month a friend of mine showed me a short video of an Uber ride. What was notable about this video from the back seat of the Uber vehicle was that there was no driver. The Uber vehicle arrived, picked her up and arrived at the destination, about 20 miles in total, with no human even in the front seat of the vehicle. Hollywood movies will undoubtedly be made of what happens when an AI system driving a car is misaligned or becomes uncontrolled.
Why Should Lawyers Care?
The alignment and control problems will affect who has liability when things go wrong. While self driving cars have a significantly reduce collision rate over those driven by humans, accidents will always be part of travel. It is lawyer who will be on the front lines, through representation of injuries parties in such accidents, that will establish the legal and ethical norms regarding AI systems.
The alignment and control problems will also affect privacy. Like the healthcare example above, AI systems are able to process and analyze large amounts of personal data. An AI system in possession of that data merits special attention to ensure that data is protected and used appropriately. In those cases where that data is improperly shared, as above, lawyers will have to understand those system sufficiently to craft liability arguments that will succeed on behalf of their clients.
Lawyers and others will be called upon to voice their concerns and provide their expertise in the crafting of AI system regulation. For example, debates are raging on Twitter and in other forums about whether lawyers themselves will be greatly reduced in number and hourly rates with the adoption of AI systems. Bar associations and state Supreme Courts will benefit from keeping a close eye on what will undoubtedly be AI lawyers that begin to siphon cases away from the court system and into prediction models powering web services enabling litigants to avoid hiring lawyers altogether while getting legal advice from AI lawyers.
As citizens of a society in general, we should all want any tool available to other humans to be used in a way that promotes the interests of our society. To best ensure that happens, getting ahead of the widespread adoption of such systems seems important.
None of the signatories to the “please let’s pause AI development” harbor any realistic concern their letter would actually bring about such a pause. First off, the pause would have to be universal to even be effective. Simply having a few notable countries impose such an AI development pause does nothing to address the concerns. Either all AI development is paused or the aims of the letter will be unmet. Considering that so many countries have aims antithetical to each other, there is no likelihood a universal AI pause could be agreed upon, monitored or ensures. The more likely purpose of the letter was to bring attention to the concerns the letter raises in hopes that most of the dominant countries in AI research (The U.S., China, Russia and perhaps ten others) would reach some sort of understanding to self-regulate or create an international body to oversee AI development paying special attention to ethics, privacy, alignment and control.
A unilateral pause by the U.S. or China, for example, would only serve to hold that country back as others continue apace developing systems that are not growing in power and speed linearly, but geometrically. I am monitoring a variety of twitter accounts right now that are on fire with hourly developments in working with OpenAI’s large language model (LLM) extending its uses beyond what OpenAI likely foresaw. They have released a partially controlled system on the world and the swarm of intelligent and industrious developers in contact with that LLM have already created a blizzard of tools and services using it.
This article on chaining together “agents” using the OpenAI LLM is less than a week old and has already been supplanted by open source projects involving multiple, independent, but cooperative agents working together to accomplish a wide range of things - autonomously. The pace of change is accelerating without a plateau in sight. The next Elon Musk is likely currently a lone developer who will create “the tool” that propels the use of LLMs forward in a lurch that will occur without sufficient time to consider ethics, privacy, alignment or control.