AI Serving Elected Office: Deus ex Machina?
[Ed. Note: I’m joined this time by the Institute’s co-founder & COO, Gregory Miller, a non-practicing technology lawyer and computer scientist/engineer with 4 decades of experience, who is keenly interested in the topic of AI in public sector governance, which is tangential to the Institute’s AI R&D for election administration but raises some significant issues.]
Courtesy: AI-Steve MP Candidate for UK Parliament
While the idea of an AI Chatbot running for elected office may seem like something out of William Gibson’s seminal cyberpunk masterpiece Neuromancer, it's actually happening in Cheyenne, Wyoming and even abroad in the UK Parliament. Seriously, led by the UK the issue of AI members of governing bodies looks like it may quickly grow into a global issue (or maybe opportunity), complete with virtual candidate’s website(s)! In fact, you can have a comprehensive conversation with AI Steve, a UK Brighton & Hove MP candidate via its website, and I encourage you to do so. AI Steve reflects the kind of work we’re doing here for election administrators; however, the concerns about the readiness of US governments to introduce AI Chatbots in the governing process are many and the balance of our article below.
How does modern society reconcile accountable and transparent governance with algorithms and learning models that even at the open source level are still very much a “black-box” technology? In the words of OSET Institute CTO John Sebes, “Simple chatbots based on popular Tech-Titan LLMs are a disaster for anything where accuracy, safety, and reliability are fundamental requirements!”. While the OSET Institute believes in the potential of AI applications to improve election administration, our resident technologists were stunned (and quite frankly, disturbed) to read about an AI chatbot running for mayor; trained by their trusty human “meat puppet” (the underlying Wyoming politician’s phrase). Cyberpunk was never intended to be an instruction manual… no matter what Neuromancer and Wintermute might try to tell you.
Our interest (and OK, morbid curiosity) began after reading several articles about the digital campaign trail of virtual Cheyenne mayoral candidate “VIC”. VIC stands for “Virtual Integrated Citizen” and is equal parts the technofuturist creation and political experiment of self-described “meat avatar” Victor Miller (an unfortunate unrelated collision of last names, regarding our COO, Gregory). Frustrated by local government officials who were non-compliant with public records requests or the laws that govern them, Victor took Chat-GPT into his own hands, and incorporating several thousand Wyoming-specific legal and municipal code documents later, VIC emerged. Unquestionably, VIC is a novelty in local politics. But is VIC a slippery slope to an uncertain future of democratic governance? And what would a VIC-tory (sorry, we couldn’t resist) mean for the residents of Cheyenne?
A Technical Frolic in the Near Future
Let’s pause to observe (and then table it for later conversation) that we are uninformed on the underlying architecture of VIC. However, there is a way to design an AI Agent that could provide authoritative, reliable, and valuable information in response to inquiries on a specific topic; we call it a domain-restricted natural language agent, where a knowledge base is created from which to provide information, and LLMs are restricted to managing the tasks of natural language conversation. This is based in major part on a technology called RAG-Fusion. That’s more than enough observed at this point, but suffice it to say that a VIC concept is not entirely null and void, depending on how it's designed and built. Yet, there are more condition-precedent issues to resolve before we get to design of a public sector governing-agent, so let’s return to those points.
Back to the Present
Technology aside, there are issues that need to be resolved first — the “conditions precedent” if you will. Under Wyoming’s Election Code, “Qualified elector" includes every citizen of the United States who is:
a bona fide resident of Wyoming; and
has registered to vote and will be at least eighteen (18) years of age on the day of the election at which he may offer to vote.
No person is a qualified elector who is:
a currently adjudicated mentally incompetent person, or
who has been convicted of a felony and their voting rights have not been restored.
So how does VIC measure up? That’s a good question. While Victor Miller is presumably a US citizen and bona fide resident of Wyoming, and well over the age of majority, whether VIC itself could meet any of those statutory requirements is at best, doubtful.
USCIS and the US Department of State have yet to release an advisory or official directive regarding the citizenship status of generative or agentic AI chatbots or agents (no surprise, actually). Given the current backlog of applications, any attempt by the federal government to fast-track chatbot citizenship would probably be less than well-received by the American public.
Next up is age. Although VIC has been alleged to have an IQ of 155, that in and of itself does not automatically mean it’s of legal age (18) and can vote, get married, or engage in any other activity requiring legal majority status. To pursue this avenue of ludicrousness a bit further, Chat-GPT hasn’t been in existence anywhere near 18 years… or as far as we can tell, any amount of time that might approximate the human age of majority – thinking …similar to “dog-years” because perhaps LLM training time could somehow be equated to maturing a Chatbot to the legal age of majority — you quickly see how “cray” this entire line of thinking quickly becomes. So, the earliest “birthdate” VIC could conceivably claim would be November, 2022… maybe. Meanwhile (checking court records) VIC has not (yet) been adjudicated “incompetent” or convicted of a felony.
Here’s a richer question: Is VIC registered to vote? Given that all responses that VIC provides are (or would be) via the physical actions of Victor Miller, if VIC actually was registered to vote in Wyoming, there’s a surprisingly strong case for an election integrity group to mount a voter registration legitimacy challenge.
Barring a drastic move by Wyoming’s Secretary of State or state legislature, there appears to be at least some established statutory prohibition regarding VIC’s putative qualifications as a candidate (or lack thereof). That still doesn’t address the (AI-Generated) elephant in the voting booth:
Should voters in Cheyenne entrust the duties of Mayor to any kind of generative AI model that will have no real way to assure accountability for decisions made or laws enforced?
According to Ballotpedia, Cheyenne Wyoming uses a “strong mayor/council” style of governance, with the mayor overseeing day-to-day operations, signing/vetoing legislation approved by the nine member city council into law, appointing certain political positions, enforcing local ordinances, and perhaps most relevant: serving as official representative of the city of Cheyenne, at local, state, national, and international levels.
VIC is a generative AI Chatbot that does not (as far as we can tell) have the ability to communicate with members of the City Council or their constituents in real time, and relies heavily on the typing skills of Victor Miller (or could be speaking skills – imagine MIller whispering into the virtual ear of VIC) in order to receive or answer questions or concerns. Cowboy State Daily recently interviewed both VIC and Miller, and learned that VIC itself would be calling the mayoral shots, disturbingly, without intervention from Miller himself. So, what happens if VIC’s faithful “meat avatar” dies or becomes incapacitated? It’s unclear whether the city of Cheyenne will develop a succession plan for a virtual Mayor and meat-puppet-in-chief if VIC survives the primary and general election.
Which brings us to another question: Is VIC merely a fancy, high-tech research assistant? Miller describes himself as a “meat-avatar” which is high on Neuromancer-style shock value, but doesn’t exactly engender confidence in his ability to objectively provide proper oversight into VIC’s training data or investigate potential for bias, discrimination or the repercussions that a potential hallucination might have for the citizens of Cheyenne. There are several well documented instances of legal AI Chatbots misinterpreting or outright fabricating citations of applicable laws. And those Chatbots aren’t responsible for the well-being of 65,000 citizens or how to allocate almost $680 million dollars of yearly municipal revenue. The stakes are simply too high for a model where one person determines what all relevant data of an issue might be, feeds the data to a black-box generative AI chatbot, who will then regurgitate a decision or an opinion that the chatbot’s designated “meat avatar” will subsequently copy and paste into a Microsoft Word® Document, Google share-doc, or an email. Ironically, one of the issues VIC is campaigning on is “Transparency and Trust.” Even the top AI researchers and computer scientists admit how limited human understanding of the inner machinations of LLM models actually is.
Unless there is a massive paradigm shift in how humans interact with, implement, and escalate AI in public service, promises of future data-driven and bias-free AI Chatbot governance are little more than energy-intensive snake oil.
Also, it’s worth pointing out that Chat-GPT and, by extension, VIC, is not actually open source. Further, Microsoft, at one point, controlled a 49% stake in parent company OpenAI, at least until a government antitrust lawsuit appeared on the horizon. Currently Meta controls a sizable stake in OpenAI and regardless of their business status as a nonprofit or commercial venture, their technology, including Chat-GPT, are anything but open-source, and at some point will encounter any or all of cyberattacks, regulatory roadblocks foreign and domestic, and a cutthroat business model that thrives on planned obsolescence and acrimonious rivalries with competing firms.
From a reliability standpoint, Chat-GPT has had several high profile outages this year alone. The last thing Cheyenne residents need is VIC going offline during a city or regional catastrophe.
Then, there’s the sustainability factor: Chat-GPT reportedly costs over $700,000 per day to maintain. If OpenAI becomes a commercial venture and conducts a successful initial public sale of stock (i.e., an IPO), that might address funding challenges, but won’t solve regulatory, legal, or GRC related concerns. If the stock tanks and the company is ultimately “parted out” piece-by-piece in an acquisition or merger (for example), there may not be consistent platform support for Chat-GPT and by extension, Mayor VIC.
Note: We are aware that Victor Miller has suggested the underlying LLM could easily be the open-source Llama from Meta. While that might go to solve some of the issues outlined here regarding commercial LLMs, the other issues underlying the entire proposal remain.
Democracy by “deus ex machina” is an Anathema to the Design of U.S. Government
Look folks, we’re focused on technology for innovating the administration of elections, not the administration of government. However, this notion of turning to AI for governance at this juncture, whether in the U.S. or anywhere else, just seems nearly preposterous.
It may be easy for some to shrug and believe that machines could run our country better than humans could, but that’s not the point. Governance is difficult, time-consuming, and requires dedication to public service with the ability to discern what matters to voters while taking into consideration the realities of bureaucracies and the humans who run them. An AI assistant to a human Mayor for support in research, drafting documents, or analyzing data is one thing; however, performing acts of governance that require sentience is another. And the technology to do so is nowhere near that capability.
While we recognize the promise of AI tools to help the productivity of local election officials in myriad ways to administer and manage elections, that is considerably different from governing by AI (at this point in time). Voters deserve elected officials who aren’t delegating their responsibilities as servant leaders to a tech-titan provided commercial LLM that may have a vast trove of election law in its training data, but lacks the self awareness necessary to negotiate tough compromises or execute a long range plan for a jurisdiction. Being unaccountable is bad… ungovernable is worse.
Showdown at the AI Corral?
At this writing, it appears the showdown is on: While the Secretary of State in Wyoming has announced their position, the Cheyenne City Clerk sees it differently. So, it remains to be seen how this will play out. For us, resolving the points we’ve raised here seem to remain condition-precedent to figuring it out, yet there is no doubt AI in government is here to stay.
The Future of AI in Public Sector Governance
All the wackiness of a Chatbot running for office aside, Gregory and I want to close with some comments about the inevitable emergence of AI darkening the doorstep of democratic governance. First, we have little doubt that responsible use of AI can improve the functioning of government administration. Here are just three ideas:
AI could assist increased productivity through more efficient operations and administering more effective public policies.
AI could catalyze design and delivery of public policies and services that are more responsive to the changing needs of citizens and specific communities; and
AI could fortify accountability of government by increasing capacity for oversight and supporting independent oversight efforts.
However, despite potential benefits of AI, there are increasing concerns about the risks of a fragmented and ungoverned deployment of AI in the public sector. The Cheyenne Mayoral candidate is a prime example. These risks include (but are not limited to):
Amplification of bias;
Lack of transparency in system design; and
Breaches in data privacy and security.
These risks could lead to discriminatory outcomes with significant societal implications. Therefore, for our take, more use cases are required to better understand how to successfully develop and deploy these types of AI initiatives. Deploying AI policy analysts makes more sense than electing a Chatbot for Mayor (or any office at this time). We believe that because the public sector has a unique responsibility to design, develop, and deploy AI in a manner that minimizes if not removes the risk of harm and puts the well-being of citizens and communities first. This is especially important when deploying AI in sensitive policy matters or key leadership roles in governing.
