AI & Democracy:

In 2016, the U.S. Presidential election served as a sobering wake-up call for Silicon Valley about the role that social media can play in impacting electoral integrity. Since then, efforts have been made across technology companies to build systems, partnerships, policies, and tools to mitigate the spread of disinformation and deceptive behaviors during elections. However, two alarming trends have emerged. Firstly, the current context of economic rigor and political debate around election integrity risks undermining some of the genuine progress in this space over previous years. In particular, teams, procedures, research, or tools necessary to tackle demonstrable attempts at undermining electoral processes are under pressure.

Secondly, we are in the early stages of a seismic shift in the threat landscape due to the advent of generative AI technologies. These two intertwined issues come at a time of rising global instability and a year teeming with critical elections around the world.

Finally, global trends show 2024 is the tipping point for democracy. Last year, V-Dem assessed that 60% of the world was under autocratic rule; this year, that changed to 72%. There are 65 elections in 54 countries in 2024, including the world’s three largest democracies (India, Indonesia, United States) and critical areas such as Taiwan and the EU. The question we must ask is: How can we prepare for what lies ahead?

Currently, a pivotal and urgent battle is unfolding that pits proponents of open and accessible AI models, against those who caution against wide dissemination of such technologies. The landscape is complex, within the Open Source proponent camp, there is no agreed upon nomenclature of what open really means, what an open source stack looks like, and whether anyone is really achieving it. And within the critics camp, some talk of the risk of misuse while others cite existential doom.

What's at stake is far-reaching: the question is whether the burgeoning powers of AI will be concentrated within a few corporate giants or whether these capabilities will be democratized. The aim is not just academic; it directly influences whether open-source AI can provide researchers and advocates with tools that enable transparency and greater understanding of these technologies, responsible innovation and solutions to problems not necessarily on the radar of large Silicon Valley firms. This urgency is amplified by employee movements within Silicon Valley, raising red flags about the potential security threats of open-source AI models. Here, the policy debate is critical but lacks solid grounding. While the technology industry has established definitions and understanding of what open source means in the realm of software, this lexicon and shared understanding are yet to be developed for open-source AI.

For the past decade, policymakers from Europe to the United States have grappled with the challenge of understanding and regulating the ever-complicated systems platforms use for content moderation. Existing frameworks, like the Digital Services Act (DSA) in the European Union, are vastly premised on the current state of moderation technologies. These are the primary bulwarks against the dissemination of harmful content, such as hate speech, disinformation, and exploitative material. Yet, an unanticipated transformation is brewing within the tech hubs of Silicon Valley. The advent of Generative AI is poised to introduce a fundamentally new paradigm for content moderation, one that brings both challenges and opportunities for the Trust and Safety field. This shift carries sweeping implications for transparency and accountability in moderation mechanisms, as well as for our understanding of the very meaning of content moderation. The need to comprehend and adapt to this change is immediate, and with the pace of innovation, almost nothing on this new content moderation paradigm is available to help guide the policy conversation.

The consensus is that Generative AI models ought to be rigorously vetted for safety before their public release is growing wide and deep. The challenge arises in defining the nature and scope of this vetting.In recent years, the concept of “safety by design” has gained traction amongst regulators and industry leaders, and the cybersecurity-inspired approach of "red teaming" has gained currency as a go-to method for stress-testing these AI models to fathom their various safety implications. However, the term "red teaming" has been applied to a plethora of diverse methods, contributing to semantic dilution and leading to what might be termed 'safety theater.' This is a precarious situation wherein a buzzword is presented to regulators and the general public as a cure-all for understanding the complex sociotechnical risks associated with AI. This is particularly critical now, as policy-makers from the European Union to the United States are inching closer to mandating such assessments. A granular, methodologically sound understanding of what it truly takes to examine these models for their societal and democratic implications is an imperative.