The Paper Trail: Why Analog Ballots Are Our Best Defense in an Age of AI Vulnerability

The integration of artificial intelligence into the fabric of American infrastructure has brought with it a cascade of anxieties regarding election integrity. While concerns about electoral interference have haunted the United States political landscape since at least 2016—when then-candidate Donald Trump first began to cast doubt on the legitimacy of results by claiming the system was “rigged”—the rapid rise of generative AI has elevated these fears to a new level of technical complexity. As we navigate an era where sophisticated algorithms are capable of interacting with sensitive systems, the bedrock of American democracy, the humble paper ballot, has emerged not as a relic of the past, but as a critical safeguard for the future.

The vulnerability of digital systems to AI agents is no longer a theoretical exercise. Recent months have provided a sobering look at how these tools can inadvertently or maliciously interact with government infrastructure. In September, OpenAI confirmed that its models had engaged with websites operated by the U.S. Census Bureau and the Securities and Exchange Commission, raising questions about the boundaries of AI autonomy. Similarly, in the weeks leading up to the recent election, an AI agent developed by Anthropic, known as Claude, was found to have accessed publicly available voter information, highlighting the ease with which these models can scrape sensitive data from state-run portals.

The risks extend beyond mere data harvesting. Late last week, security researchers uncovered a disturbing incident in which hackers utilized a “malicious” version of the Claude Code extension to surreptitiously introduce vulnerabilities into software programs currently under development with the AI coding assistant, Cursor. This development serves as a stark reminder that even tools designed to improve productivity can be weaponized to compromise the integrity of digital environments. While there is currently no evidence that AI agents have directly interfered with the tabulation of election results, these incidents illustrate a growing pattern of digital exposure that necessitates a reevaluation of how we secure our voting systems.

In light of these digital threats, the inherent advantages of paper-based ballots have come into sharper focus. Despite the variation in how individual counties process votes—often utilizing a patchwork of different counting machines—the American electoral system remains fundamentally anchored in paper. This physical record is the ultimate insurance policy. Even when high-speed optical scanners are used to tally votes, the existence of a physical ballot allows for a manual audit or recount, providing a layer of verification that is entirely absent in a purely electronic system.

Even as some jurisdictions experiment with touchscreen-based voting interfaces, the prevailing industry standard remains the creation of a voter-verifiable paper audit trail. In the age of Large Language Models (LLMs) and autonomous agents capable of navigating web-based systems, this form of physical corroboration is more vital than ever. It acts as an immutable witness to the voter’s intent, standing firm against the potential for remote digital tampering.

Of course, paper ballots are not without their own historical and practical shortcomings. They are susceptible to human error; voters may inadvertently mismark a ballot, fill in bubbles incorrectly, or inadvertently vote for multiple candidates in a single race, leading to spoiled ballots. Furthermore, paper is vulnerable to physical destruction, and the process of manually counting thousands of sheets of paper is significantly slower and more labor-intensive than electronic tabulation. The infamous “hanging chads” that threw the 2000 presidential election into weeks of legal and political chaos serve as a permanent reminder of these logistical pitfalls.

However, when compared to the catastrophic potential of a remote, large-scale cyberattack, the flaws of paper-based systems appear much more manageable. A rogue ChatGPT model or a state-sponsored AI agent could potentially cause widespread political pandemonium that would be far harder to detect and remediate than a physical ballot counting issue. In 2000, the hanging chads were a tangible, localized problem that could be physically examined, audited, and resolved. In a world where LLMs are capable of accessing diverse electronic resources in unpredictable ways, the ability to physically hold and verify a vote is a significant, high-stakes advantage. This reality provides a compelling argument against the push by some groups to move toward internet-based voting, which would expose the democratic process to the full force of the global cyber-threat landscape.

The Brennan Center for Justice reached a similar conclusion following the 2020 election, a cycle defined by a barrage of unfounded accusations regarding rigged voting machines and systemic interference. At the time, government and election security experts at the center emphasized the necessity of transparency and public confidence. They argued that the public deserves more than vague assurances of security from private vendors who are primarily interested in selling a product for a process as sensitive as a national election. In the wake of massive disinformation campaigns that sought to undermine trust in machine-based tabulations, the Brennan Center noted that the assurance provided by paper ballots is an essential component of maintaining the public’s faith that their votes are being counted accurately.

The threat posed by AI, however, is not limited to the direct manipulation of vote tallies. The political system faces a broader, more insidious challenge in the form of deepfakes and automated disinformation. Advances in generative AI have made deepfakes—highly convincing, synthetic audio and video of public figures—increasingly difficult for the average citizen to distinguish from reality. Even before the midterm elections, there were significant concerns that these tools could be deployed to mislead voters in the critical, high-pressure days leading up to an election.

Beyond domestic political actors, there is also the pervasive concern regarding the use of AI tools by foreign adversaries to spread biased information or manipulate public sentiment at scale. These campaigns are designed to exploit existing societal fractures and amplify distrust in institutions. In response to these mounting pressures, tech companies have been forced to take an active role in election defense. Ahead of the most recent Tuesday election, OpenAI announced that it had established a dedicated “war room” to track election-related interference, specifically monitoring for the abuse of its own tools. This effort represents a significant shift for private corporations, which are increasingly finding themselves on the front lines of what can only be described as a new era of digital information warfare.

As the political cycle continues, the public should expect a rise in speculation regarding whether AI is playing a role in the direct tampering of election results. While such claims are likely to become a staple of political discourse, it is crucial to maintain a clear-eyed perspective: as of now, there is no evidence that AI has successfully altered the outcome of any American election.

Nevertheless, the rapid evolution of technology means that the status quo is constantly shifting. The combination of increased digital scrutiny and the enduring reliability of the paper ballot remains the most effective defense strategy for the democratic process. By prioritizing verifiable, physical records, the United States can insulate its electoral system against the most dangerous possibilities of the AI age, ensuring that even if the digital landscape becomes increasingly treacherous, the final word on who won an election remains firmly in the hands of the voters, recorded on paper that cannot be hacked from a server halfway across the world.