🎙️ Q&A with David Evan Harris

David Evan Harris was named one of Business Insider’s top 100 people in AI in 2023. Harris recently began serving as a Member of the Artificial Intelligence (AI) and Election Security Advisory Committee, convened by Arizona’s Secretary of State. As the Chancellor's Public Scholar at UC Berkeley and Continuing Lecturer at the Haas School of Business, he teaches courses on AI Ethics, Social Movements & Social Media, and Civic Technology. Harris has held positions in the U.S. government and founded the Global Lives Project in 2004, an initiative building a video library of daily life worldwide. We sat down with Harris to discuss his views on equitable AI, the militarization of AI, and fostering ethical tech ecosystems.

Sudeshna Mukherjee (SM): Your career spans academia, the private sector, and the nonprofit world. How do you leverage these experiences in your work on Artificial Intelligence (AI)??

David Evan Harris (DEH): Now that I'm doing more work in academia and the nonprofit sector, people value my firsthand experience of how decisions are made inside big tech companies. People in government have called upon me to share my views because they lack access to people with inside experience in tech companies who aren't currently employed there. In the past year, I’ve testified almost a dozen times before legislative bodies—in San Francisco, Sacramento, and most recently, the US Senate.

Right now, my big focus is trying to get AI legislation passed. We had two big wins recently with AB 2839 and AB 2655—two bills banning election-related deepfakes—in California just getting signed by the governor. This was the result of the work of the team at the California Initiative for Technology and Democracy (CITED).

SM: What does equitable AI mean to you? As someone who worked on AI fairness and inclusion at Meta, and having been named one of Business Insider's top 100 people in AI, what do you think are the most pressing challenges to ensuring equitable AI systems, particularly for marginalized communities?

DEH: Today's AI products have the potential to cause harm. For instance, when I worked at Meta (formerly Facebook), the company was sued by the U.S. Department of Housing and Urban Development for facilitating housing discrimination through its AI-based ad targeting systems. This is just one example of how automated decision-making systems can be biased and discriminatory.

Language is another critical area. While generative AI systems can work in many different languages simultaneously, I suspect that companies aren't testing their AI systems in every language they're deploying them in. Meta's LLaMA 2, for example, functions in many languages - but they mentioned they did not do safety testing in any languages other than English. This can have severe consequences. A stark example of this is what happened with Meta in Myanmar. For a time, when genocide was being promoted on the platform, the company had no content moderators who spoke the critical languages involved in that conflict. I firmly believe that you shouldn't be able to operate a social media platform in a country if your company doesn't understand its history and linguistic nuances. This principle extends to AI systems as well.

We also face issues of representational harm, where AI systems generate biased or offensive representations of different groups. For instance, AI image generators have been documented to sexualize or over-sexualize images of women compared to men. 

Lastly, we need to consider the distribution of access to generative AI tools and whether this access is equitable. Many of these tools require subscriptions that can be quite expensive, sometimes $10 or $20 a month, which is not accessible to large parts of the world. While these tools can provide real value in certain professional circumstances, we need to think carefully about balancing equitable access with the potential impacts of these systems.

SM: You founded the Global Lives Project, aimed at building a video library of daily life around the world. How might initiatives like this inform our approach to developing more culturally inclusive and globally representative AI products and services?

DEH: The Global Lives Project is primarily a film and video art effort, not intended to be about AI or AI technology. I've been concerned that AI companies might, without consent from me or the filmmakers I've worked with, decide to use our videos for training data simply because they're available online. While I can see potential benefits in using this video content for AI, I don't believe it is ethical to use it for training AI systems since that wasn't the original intent of the work.

This raises important questions about consent, copyright, and the ethical use of publicly available data for AI training. It's a dilemma that many creators face, and we need to address it as an industry and society. The Biden Administration's Executive Order on Safe, Secure, and Trustworthy Artificial Intelligence did have a series of provisions related to copyright and AI. They asked the Library of Congress, the Copyright Office, and the Patent and Trademark Office to conduct research and provide recommendations about what to do about AI and copyright. However, it didn't change any legal status or resolve the ambiguity surrounding these issues.

This is still an emerging space with lots of work to be done. On an individual basis, there really should be the option to opt out of having your information used to train AI models. Earlier this year, Meta declared it would use users’ Facebook and Instagram photos to train its text-to-image model. And while they offer a form of opt-out functionality, they have made the process deeply convoluted and uncertain. Beyond this functionality, no U.S. law currently prevents Meta from taking away this option in the future. 

This is also an opportunity to consider the benefits of highly curated datasets. Hypothetically, highly curated datasets could solve content consent issues by only including content from those who have approved its use for training. Further, increased curation could help prevent harmful content from ‘poisoning’ the training dataset and impacting model outputs. A prime example of this was when independent researchers discovered suspected CSAM in the training data for Stable Diffusion. Curating training data has become a central focus for many AI companies in recent months, especially as legal and regulatory pressures begin to mount. 

SM: What are your thoughts on the militarization of AI and its potential global impacts?

DEH: The militarization of AI is a complex and concerning issue. We're seeing AI-like technologies in autonomous weapon systems and drones. My colleague Stuart Russell at Berkeley has been working to push for treaties against the use of lethal autonomous weapons.

Countries like Russia, China, North Korea, and Iran have been caught attempting to militarize tools from Microsoft-backed OpenAI. We can expect more of this disinformation militarization, as well as the development of lethal autonomous weapons. One author I like, Renee DiResta, uses the term "Digital mercenaries" to describe groups for hire that will do this kind of work. These are groups that engage in disinformation campaigns and influence operations, often on behalf of state actors. The EU AI Act, which I consider the most important AI legislation globally, has broad exemptions for military and policing use of AI. Unfortunately, in the current geopolitical climate, the world's leading countries aren't inclined to voluntarily slow down weapons development.

There are also concerns about chemical, biological, radiological, and nuclear risks, where AI could potentially be used to develop new weapons in these categories. Unfortunately, I think we need to play the long game. The world isn't ready to constrain military uses of AI right now. Geopolitically, the world is at a difficult place with the conflicts in Ukraine and Gaza. The world's leading countries are not looking to slow down any kind of weapons development. We're lucky so far that we haven't seen nuclear weapons used in either of these two conflicts. But it often seems like we're at the brink of that. 

The most promising effort for meaningful agreements might be the series of AI safety institute meetings, which began at Bletchley Park in the UK last fall. The UK government initiated these meetings in response to growing concerns about the potential risks of advanced AI systems. The primary goal is to foster international cooperation on AI safety and establish common principles and practices for responsible AI development. Importantly, these meetings included participation from countries like China, Saudi Arabia, and the United Arab Emirates – nations with significant AI infrastructure investments that need to be involved for any restrictions to be meaningful. The organizers aim to create a global framework for AI governance and safety standards, recognizing that the challenges posed by AI transcend national boundaries and require collaborative solutions.

SM: In your articles for The Guardian and the Tech Policy Press, you express significant concern about the risks posed by unsecured AI systems, particularly in the context of elections and society more broadly. How can policymakers and tech companies work together to mitigate these risks while still fostering innovation? What specific regulatory measures are going to be the most crucial to address the misuse of AI by sophisticated threat actors, especially as we continue to go through major elections worldwide?

DEH: It's crucial to be extremely careful when releasing unsecured AI systems because you cannot recall them, once they are out. A prime example is the AI image generation tool Stable Diffusion. A version of the generator was recently exposed as being inadvertently trained on thousands of images of child sexual abuse material (CSAM). Even after this was discovered and reported, Stable Diffusion 1.5 was left available online on a website called Hugging Face, which is a well-known code repository for unsecured or "open weights" AI systems. And even after its recent removal from Hugging Face, the model can still be downloaded from an alternate repository, known for hosting many models that are fine-tuned to create pornographic content.

Regarding the upcoming elections, I'm particularly concerned about non-consensual imagery or deepfakes of political candidates and election officials. We need much clearer warnings when something is a deepfake and not the authentic voice or image of a politician. A recent example is an audio deepfake of Kamala Harris narrating a video. Elon Musk shared it on X (formerly Twitter) without a clear warning that it was a deepfake or parody. 

We've seen disinformation campaigns and information operations before, but now it's much less expensive to produce them. Some companies even advertise they can make tens of thousands of simultaneous political phone calls using an AI-generated voice that has real conversations with voters. This could also be done through text-based chats or a combination of text, audio messages, and deepfake images.

These developments underscore the urgent need for robust regulatory measures to address the misuse of AI in electoral contexts, particularly by sophisticated threat actors. We need to focus on transparency, quick response mechanisms, and cross-platform cooperation to mitigate these risks while still fostering innovation in AI technology.

Follow David's work on LinkedIn.

Edited by Estelle Ciesla and Vance Ricks