There are teams from researchers in academia and with large AI labs are currently working on the problem of AI ethics, or the moral concerns raised by AI systems. These efforts tend to focus on data privacy issues and on what is known as AI bias – AI systems that, using training data with often built-in biases, produce racist or sexist results, such as women deny credit card limits they would grant a man with identical qualifications.
There are also teams from researchers in academia and with some (although fewer) AI labs working on the problem of AI alignment. This is the risk that, as our AI systems become more powerful, our surveillance methods and training approaches will become increasingly useless to get them to do what we actually want. Ultimately, we will have handed over the future of humanity to systems with goals and priorities that we do not understand and that we can no longer influence.
Today, that often means AI ethicists and those who are AI alignment are working on similar issues. Improving the understanding of the inner workings of current AI systems is an approach to solving AI tuning and is crucial to understanding when and where models are misleading or discriminatory.
And in some ways, AI tuning is just the problem of AI bias that’s (terrifyingly) big: we’re assigning more societal decision-making power to systems that we don’t fully understand and can’t always control, and that lawmakers don’t. t knows almost well enough to regulate effectively.
As impressive as modern artificial intelligence may seem, right now those AI systems are in a sense “stupid.” They usually have a very limited range and limited computing power. To the extent that they can cause damage, they usually do so by: replicating the damage in the datasets used to train them or by intentional abuse by bad actors.
But AI won’t stay dumb forever, as many people work diligently to make it as smart as possible.
Part of what limits current AI systems in the dangers they pose is that they don’t have a good model of the world. Still, teams are working to train models who to do have a good understanding of the world. The other reason current systems are limited is that they aren’t integrated with the power levers in our world — but other teams are working hard to build AI-powered drones, bombs, factories, and precision manufacturing tools.
That dynamic – where we continue to make AI systems smarter and smarter, without really understanding their goals or having a good way to monitor or control them – sets us up for disaster.
And not in the distant future, but in a few decades. Therefore, it is critical that AI ethics research focuses on mastering the implications of modern AI, and AI alignment research focuses on preparing for powerful future systems.
Not just two sides of the same coin
So can these two groups of experts responsible for making AI really get along?
These are two camps, and they are two camps that sometimes hate each other.
From the perspective of people working on AI ethics, experts focused on alignment are ignoring the real-world problems we’re already experiencing today, in favor of obsessing over future problems that may never come. Often the reconciliation camp doesn’t even know what issues the ethics people are working on.
From the perspective of many AI alignment people, however, many “ethics” working at top AI labs is basically just glorified public relations, designed primarily so tech companies can say they care about ethics and avoid embarrassing PR snafus — but do nothing to change the overall trajectory of AI development. In surveys by AI ethics experts, most say they: don’t expect development practices at top companies change to prioritize moral and societal concerns.
(Just to be clear, lots of AI alignment folks also straight away this complaint to others in the alignment camp. Many people are working to make AI systems more powerful and dangerous, with various justifications for how this helps to learn how to make them more secure. From a more pessimistic perspective, almost all AI ethics, AI security, and AI tuning is really just working towards building more powerful AIs — but with better PR.)
For their part, many AI ethics researchers say they’d like to do more, but are thwarted by corporate cultures that don’t take them very seriously and don’t see their work as a major technical priority, like former Google AI ethics researcher Meredith Whittaker noticed in a tweet:
I have an AI ethics joke, but it has to be approved by PR, the legal department, and our partners in the Department of Defense before I can tell.
— Meredith Whittaker (@mer__edith) July 26, 2020
A healthier AI ecosystem
The battle against AI ethics/AI alignment does not have to exist. After all, climate researchers who study the current effects of warming do not tend to bitterly condemn climate researchers who study long-term effects, and researchers who work on projecting the worst-case scenarios don’t be tempted to argue that anyone working on heat waves these days is wasting time.
You could easily imagine a world where the AI field was similar – and much healthier.
Why is not that the world we are in?
My instinct is that the AI battle is related to the very limited public understanding of what is happening with artificial intelligence. When public attention and resources seem scarce, people find misguided projects threatening – after all, those other projects get involvement at the expense of their own.
Many people – even many AI researchers – are not concerned about the safety effects of their work very serious.
At the various large-scale labs (where large-scale = several thousand GPUs) there are different opinions about how important security is. Some people care a lot about safety, others care very little about it. If security issues turn out to be real, uh oh!
— Jack Clark (@jackclarkSF) August 6, 2022
Sometimes leaders dismiss long-term security concerns out of a genuine belief that AI will be very good for the world, so the moral thing to do is accelerate development.
Sometimes it is based on the belief that AI won’t be transforming at all, at least not in our lives, so all this fuss is unnecessary.
Sometimes, though, it’s out of cynicism — experts know how powerful AI is likely to be, and they does not want supervision or responsibility because they think they are superior to any institution that would hold them accountable.
The public is only vaguely aware that experts have serious concerns about the security of advanced AI systems, and most people have no idea which projects are priorities for the long-term success of AI alignment, which concerns are related are to AI bias, and what exactly AI ethicists do every day anyway. Internally, AI ethical people are often isolated and isolated in the organizations where they work, and have to fight to get their colleagues to take their work seriously.
It’s these major gaps with AI as a field that, in my opinion, cause most of the divides between short-term and long-term AI security researchers. In a healthy field, there is plenty of room for people to work on various problems.
But in a field that struggles to define itself and is afraid of achieving nothing at all? Not so much.
A version of this story was initially published in the Future Perfect newsletter. Sign up here to subscribe!