CIFAR DLRL Summer School: Ethics in AI

I recently attended the CIFAR DLRL Summer School, which brings together participants to engage with foundational research developments in the fields of deep learning and reinforcement learning. Traditionally, the school is held every summer at MILA, but due to the ongoing COVID-19 pandemic, the organizers decided to hold a virtual summer school.

The week featured talks, panel discussions, breakout sessions, and 1:1 chats with speakers. I thought all the talks and sessions were top-notch and were differentiated from conference or seminar talks since they covered more ground and explored multiple frontiers of research. A common thread woven through all the sessions was the willingness to engage in tempered speculation about future and evolving research directions. Naturally, it’d be difficult to succinctly summarize all the sessions. Instead, I’ll focus on one particularly thought-provoking session, the “Ethics in AI” panel discussion, Yoshua Bengio and Doina Precup, moderated by Sasha Luccioni.

Context for Discussion

The fields of machine learning and artificial intelligence, more broadly, have made incredible progress in recent years! They’ve left from being a pure research exercise in academic and industrial labs, and have percolated into many areas of society. A sampling of workshops from a conference I attended last year (NeurIPS 2019) is a case in point:

Tackling Climate Change with ML

Since developments in the field now have real-world consequences, ethics shall become a central part of the conversation.

Note: Any factual errors or omissions are solely my own responsibility, and do not reflect the views of any speakers or participants. I’ll try to minimize editing and avoid interpretation bias.

Q/A

Why are we thinking about ethics if we’re not philosophers?

Real-world applications require multi-faceted expertise and healthy communication between ML researchers, healthcare workers, and social scientists (for example). As ML researchers it is our responsibility to anchor the discussion by exposing the technical limits of the systems and algorithms that we design. We should defer to other people who know much better than us how society functions. Science and engineering have always gone hand in hand. For AI this is even more important since these algorithms ingest data and produce results that are not always visible. If we end up succeeding, we may have non-biological intelligence, and everyone must be a part of this discussion.

How do we regulate AI such that it won’t hurt its’ development since policymakers don’t understand AI?

Hurting is a false question, since if we don’t design the right framework, then ultimately the implementation will backfire, and the broader public will not trust the technology. We need to do the right things, even if in the short term it’ll put in hurdles and costs. There’s an important education piece that needs to happen. We need people to understand what the tech can (and cannot) do. Anyone can understand what the limits and the power of the tech are. It’s a conversation and education which is completely necessary. Everyone who votes needs to understand enough about the collective decisions that we take. Think about the core science knowledge which we teach young students. AI should be a core part of the curriculum. We need to keep the conversation going. Our voice is currently completely framed by images of sci-fi that people receive in movies.

What are the difficulties which you see?

The datasets that we’re using are very important. We need to try and ensure that the datasets which we use are appropriate. We need to understand if people understand how these methods work. As researchers, we love the ML core, but the user interface is extremely important. The AI component needs to come at the right time, and be presented in the right way. Everyone should get to participate. We CS scientists need to learn how to think about social impact since we often don’t have the right training for this. Several areas within ML are related to ethics, such as bias, AI safety, and explainability. Many computations happen, which are non-explainable. We perhaps need some equivalent of GDPR to be present. We need to explain how the system was trained (e.g. what was the objective function?). How do we ensure that the match between our intentions and what the agent does isn’t catastrophic? This leads to interesting questions about generalization capability as well.

What does that mean for us? What should we prepare for?

Ideally, we would have mandatory courses on the social impact of tech for all CS students. There’s a bit of that in engineering, but typically none in CS. We have a responsibility to create training courses for our grad students (instead of waiting for this framework to evolve). Ideally, it shouldn’t be an isolated course, and definitely shouldn’t be attended by students to just “check a box”. The material should be woven inside the ML class.

How do you keep the big picture in mind when you’re working on fundamental research?

The main exercise is to think about what this algorithm be used for. What could this affect? Not all research that we do is easy to think about in those terms. You can imagine some uses for your algorithm that you hope could be deployed and then elaborate on that. We have to take the time to ask those questions ourselves. Conscious attention is required for us to see things that we might else miss.

Might something like the Geneva Convention be necessary for AI?

I have qualms about international laws or treaties, that mention AI. Among the programs that I’m running, which ones count as AI? In the case of weaponry, legal and CS experts have come up with definitions of lethal autonomous weapon systems. The system can make complex decisions that involve immoral components that would take the human a long time. For current systems, no machine can understand social or moral aspects. Humans could decide to not push the trigger. The machine will just do it. It could become real if we don’t regulate these weapons at the international level. Tax AI? That doesn’t make sense. We need to be clear on what types of things to target.

Can we define morals and ethics as part of the objective function?

This is a major consideration in AI alignment. How do we make sure that the moral goals which we have (i.e. made explicit in ambiguous laws, regulations, constitutions, etc), are well understood in computers? Especially when they’re trying to optimize these objectives? One approach is to look at how humans behave, and what we consider as virtuous behavior. The other technique is to understand that some aspects of behavior are not articulated explicitly. For such situations, this aspect of a “human in the loop” is going to become crucial. We’re all biased! Introducing a human in the loop also makes the entire system biased. For example, most people aren’t fond of academic committees, but they are necessary to have robustness in the system.

We aren’t always aware of how AI is being used. How do keep an eye on how things are being used in applications?

It’s critical to maintain public transparency. Everything you do should be open source and documented. You could also have this available to auditors which are chosen by our government. Companies should be forced to document what a system does, how it does it, etc. when it is deployed to millions of people. This wouldn’t be too dissimilar from how airplanes have auditors. Having a bit of experimentation is good though since we don’t know what is the right way to do this.

Since we have fields that already have regulation (i.e. healthcare), how do we integrate AI?

Data handling/consent, all of these things need to be done. We need to tell regulatory bodies, what kinds of stats would be compelling to look at. We need more work done in this space.

Since we’re doing things that are impacting the general public, how do we reach out to them?

We talk about it. Suppose you’re shopping, should you have a pop up that details how you’re data is being used? What about having a data trust? The idea here is to have a 3rd party that controls the data of many people, where they are not going to make money from the data itself. They only represent your interests, so that if there are many people, this system has huge leverage. There’s something wrong when we accept detailed privacy policies for online platforms without reading them since contract law means that the two people understand what they’re signing up for. This applies to healthcare data as well.

As a grad student, choosing topics, making sure that our research is not going to be misused, is this going to be important to bring good in the world?

Due to the competitive behavior that we have, we tend to overvalue theory and algorithms, as opposed to applications. We need to apply our skills to help the world. People underestimate how many people are doing things in AI for Good. There’s tremendous fulfillment in pursuing these applications. It’s necessary how the field views how it views work. We need a change in culture within our community.

How do we make the social and ethical questions more relevant within our discussions?

We need to change our value systems, share objectives. How do we change a culture? I don’t know, but it can change! You need to understand how decisions are made. It is important to cultivate a cohesive community.

How do we prevent ML from exploiting human weaknesses?

This is exactly why we need regulations that force transparency. To make sure that ML tools aren’t abused in a way that goes against our ethics. If you bring this in the open and explain it to the public, many people won’t like it, but it’s a fairly robust way of evaluating what we consider good in society. This has to be integrated into a local context, with the local laws. This is why we need diversity in all parts of the AI pipeline. If there’s only one type of person, you’ll only see one type of vision. People from different walks of life will be interested in different types of visions.

As ethics applied to each field, why is it so concerning for AI?

AI research is on a path to create very powerful tools for humans. Consider the knife vs nuclear bomb for example, in their potential to cause destruction and loss of life. As ML becomes more powerful, it can manipulate us in different ways. If AI was weak, we wouldn’t bother. As far as regulation goes, there are artifacts (e.g. GPT-3 can be regulated), and there are the algorithms that can produce more artifacts. When you create powerful tools, a natural consequence of this is that it gives more power to certain entities. There’s a vicious concentration of power, tools -> power -> better tools. This is why we need world level governance for international regulation on how AI is used. I believe that AI can create tremendous good.

CIFAR DLRL Summer School: Ethics in AI

Context for Discussion

Q/A

Further Reading