Can Democracy Survive Artificial General Intelligence?Seth Lazar, Alex Pascal / Feb 13, 2024
2023 was the year of AI, as new products brought recent advances in the field to universal attention, and the world’s most powerful tech companies declared their ambitions to achieve ‘Artificial General Intelligence’ (AGI). 2024 will be the year of democratic elections, with a record-breaking 40-plus countries (including the US, India, UK, Ukraine, Taiwan, and South Africa) representing more than 40% of the world’s population going to the polls. Many are already, justifiably, cautioning about the direct impacts of the former on the latter, as our information and communication environment becomes ever more polluted with AI-generated deepfakes, disinformation, and potentially cyber attacks. How much difference AI will make in 2024 remains an open question. Beyond these immediate threats, however: the advent of AGI could finish democracy once and for all.
To many, this fear will seem remote, even hypothetical. That’s understandable. There are, at present, many more pressing concerns raised by existing AI systems—including harms ranging from bias and discrimination to labor exploitation and mass surveillance, as well as economic disruption from job displacement to turning the creative economy on its head. Moreover, one of the biggest problems with AI today is how it falls short of expectations, not how astonishingly capable it is.
So, if we are a ways from AGI, why worry about it now? Because the people building the most advanced AI systems are explicitly and aggressively working to bring it about, and think they’ll get there in 2-5 years. Even some of the most publicly skeptical AI researchers don’t rule out AGI within this decade. If we, the affected public, do not shape this techno-quest now, we may miss the chance to do so at all. We face a fundamental question: is the very pursuit of AGI the kind of aim democracies should allow?
What even is AGI? Defining it sometimes feels like pinning jello to a wall. But as progress accelerates, something like a consensus is emerging. Synthesizing a vast literature, AGI would be a non-biological computational system that can perform any cognitive function currently performed by humans at the level of the median human or better (acknowledging the crude quantification and assumptions this implies). Google DeepMind's recent paper, “Levels of AGI: Operationalizing Progress on the Path to AGI,” mentions “linguistic intelligence, mathematical and logical reasoning, spatial reasoning, interpersonal and intra-personal social intelligences, the ability to learn new skills and creativity” as significant milestones. Other AI researchers would also add: instrumental rationality; causal reasoning; tool use; and at least some ability to distinguish truth from falsehood. OpenAI calls it, simply, “AI systems that are generally smarter than humans.”
Existing AI systems are undoubtedly far from AGI, on every criterion just mentioned, besides perhaps linguistic competence. And yet, GPT-4, OpenAI’s most advanced model, is significantly more general and capable than earlier systems. The feasibility horizon of AI research is rapidly expanding outwards. And while technologists do not yet have AGI's ingredients, many think they know where to look—in terms of both building on GPT-4's successes, and backfilling its limitations. And the leading AI labs and Big Tech companies—responsible for so much progress over the last ten years—have the creation of AGI as their explicit mission. Whether you think that’s just hype or else that achieving AGI is inevitable, we should at least ask, now, whether pursuing that goal is itself consistent with democratic values.
A democratic greenlight, not just guardrails
At a first pass, the answer seems to be no. AGI could do more than any preceding innovation to shape and disrupt our economies, politics, culture, and communities. In democracies, the people are sovereign. All should stand as equals and govern together (at least through our representatives). There is nothing inevitable about AGI’s arrival. It is a choice. One that will affect all of us profoundly. The question is, who’s making it?
Right now, the answer is a few people at even fewer tech companies. Allowing private companies to unilaterally pursue the development of technologies as potentially transformative as AGI is self-evidently undemocratic. Notwithstanding the well-intentioned experiments in corporate governance to make some leading AI labs more public-interested than regular businesses, a Board of Directors simply cannot adequately represent the societies and people whose trajectories and lives AGI will radically transform. Thus far, the public debate about how to rein in the societal impacts of AI has focused only on identifying guardrails to shape its development, mainly to mitigate AI’s harms and risks. This is necessary, but not nearly sufficient. We also need to ask the more fundamental question of whether we actually want to build AGI in the first place. The advent of future technologies is not a fait accompli—we have shown in the past, for example with human cloning, that we can slow or stop development if we choose to do so. Whatever you think our AI future should be, it should be one that we have consciously chosen together. We, the People, have to call the question, lest we sleepwalk into an AGI future we don’t in fact want.
A democratic path to AGI?
Suppose then that democratic publics explicitly and affirmatively decide that they want AGI. Could we develop it in accordance with democratic values? Some of the leading AI labs clearly recognize this question’s urgency, and are making respectable efforts in this direction. Incorporating democratic inputs into AI development has already led to some noteworthy improvements (and we welcome experiments aiming to use AI more generally to enhance political participation). But democratic inputs are not the same as democratic control. Accepting inputs presupposes controlling the agenda and dictating where inputs are welcome. There will be many branching paths on the road to AGI, and at many of those junctures—for example, how to source and filter the data on which AI systems are trained—the public interest will predictably conflict with the pursuit of returns on investment. When tens of billions of dollars have been invested in a company, those billions will ultimately set the agenda, irrespective of corporate structure. OpenAI’s recent governance crisis is case in point.
In addition to the kind of deep democratic oversight of AI development that some have proposed, the only reliably democratic path to AGI would likely involve complementing any private sector R&D with a robust and capable counterpart driven purely by the public interest—an AI ‘public option,’ as some have called it. Investments like the National AI Research Resource could light a path to such an approach, but would require an order of magnitude greater commitment and ambition to succeed.
What happens if we ‘succeed’?
But does the path to AGI lead somewhere that democracies should go? In 2023, many loudly argued ‘no,’ not because of the implications for democracy, but because they think AGI poses an existential threat to humanity at large. A range of experts have presented scenarios, ranging from speculative to compelling, in which AGI is humanity’s final, civilization-ending invention. The leading AI labs also feel this critique keenly, and have built research teams aiming to ‘align’ AGI (and ASI, its superintelligent successor) to make it ‘safe, beneficial, and controllable.’ But even if we can align AGI to meet these criteria (assuming we can decide, democratically, what they mean and how to achieve them), this would still not be enough for AGI to be safe for democracy.
Here’s why. If AGI is better than most humans at all cognitive tasks, it is very likely to be better than humans at the numerous tasks of governing—that is, designing, implementing, and enforcing the rules by which a community or institution operates. This will create a compelling incentive to invest AGI with governing power at all levels of society—from clubs, schools, and workplaces, to the administrative agencies that regulate and help steward the economy, labor, the environment, transport and healthcare, and even provide for public safety, criminal justice, and election administration. If in fact AGI is much better at executing the tasks that we give it than humans (as its would-be creators intend), there will be a strong, perhaps irresistible temptation to have it identify and select which tasks to pursue, then to have it set our priorities, not just make and enforce our rules in particular domains. As new threats and problems arise faster than we can process them, we may very well entrust AGI with a blanket authority to prioritize, decide and act on our behalf. We would de facto be kissing good-bye to democracy in any real sense of its value and practice. Think of this threat as an absent-minded walk down a political primrose path, not the more widely-discussed ‘rogue AI’ scenarios. We already see this kind of easy deference to existing, deeply flawed computational systems. It would only be exacerbated with AGI.
From decades of work on automation, we know that in every domain, from manufacturing to algorithmic trading, automating a task and then relying on humans for oversight at critical moments is a doomed project. The goal of making future AGI systems ‘controllable’ cannot be achieved through technology design alone. For anything to be controllable, we have to presuppose something or someone doing the controlling. It is not enough to design systems that could in principle be controlled, but where we can reliably predict, based on past experience, that humans will fail to use the controls that we have designed for them. Nor is having some AGIs control others an adequate answer. For AGI to be safe for democracy, democratic institutions run by people must be able and expected to exercise meaningful control. This may well require rethinking the aging institutions of constitutional democracy itself—something that only we, the People, can legitimately do.
Setting AI entirely aside, this year will prove for many democracies their sternest test yet, and may see more voters than ever before choose candidates who have explicitly promised an anti-democratic agenda. These developments show that we cannot take the value of democracy for granted—we can’t treat it as such a sacrosanct and shared ideal that nobody could ever credibly make an argument against it. Some might embrace the idea of replacing our messy, disputatious political systems with “efficient,” “impartial,” “optimizing” technocratic AGI rule. We do not.
But let’s have that debate, and not underestimate the gravity of the choice we’re now making passively, by default. Otherwise in 2024 we might save democracy from the would-be autocrats, only to pave the way for AGI to deliver it an even more decisive blow.