What we did, in a sentence: we built an index that attempts to score countries by their current vulnerability to AI-amplified democratic backsliding. It’s an early pilot with several points that we flag, so push-back is highly encouraged.
In what ways does artificial intelligence (AI) affect democratic systems? We’d wager that many would agree that there's great potential for both positive and negative effects; our investigation covers those that drive countries towards authoritarianism. In this decidedly ‘negative’ realm, we identify five preliminary pathways for AI-amplified backsliding:
We’ve named these Drivers of Political Change [1]. Per our definition, they're factors that erode a country's democratic practices and have historically preceded shifts towards authoritarianism. Sometimes they coincide with races to the bottom, sometimes reactions to crises, and other times foreign intervention. Often, they arrive as a result of many systematic and sudden factors. Nonetheless, our reference point is the authoritarian regime — these factors, when yielding negative outcomes, push states towards authoritarianism; if the state is already largely authoritarian, it helps consolidate state power in favor of that authoritarian system.
How on earth does an index really help, anyway? First and foremost, they’re awareness-building tools for citizens and organizations. If you want support for proposed countermeasures, the awareness must exist to make the policy salient in the first place. Second, we aim to fill a gap that we believe AI risk evaluations miss. Audits of this nature often focus on organizational structure and model risk. And since AI development is currently led by multiple private actors and deployed around the world, we think that evaluations fail in addressing the impact of AI development, deployment, and usage at a country level; this is worth exploring because AI interacts with existing political systems and social structures.
Now, it’s not like AI’s effect on democratic practices and stability is some unexplored phenomenon — Freedom House’s Freedom on the Net [2] and CAIDP’s
Preexisting political conditions like economic strife, a degraded information environment, and fractured elites are ripe to see their democratic knock-on effects amplified by AI. A state's exposure is therefore largely a function of how many of these AI-amplifiable vulnerabilities are already present. As such, existing indices of this type are designed for a pre-AI threat environment, and they only capture materialized threats, which makes them less suitable for measuring exposure to risks that have not yet materialized. We also see limited utility in measuring AI democratic risks after they materialize. Because the transition may be harder to reverse than in previous eras, it's more useful to assess risk through small early triggers.
We followed the OECD's composite indicator guidelines to help us structure our methodology. We collected proxy indicators related to our five drivers, including 29 AI-specific indicators and 88 non-AI-specific (democracy health indicators). The selection for the 27 countries relied on geographical and political diversity. Linear interpolation was only used for imputing data that is missing at random. We applied a 30% missingness threshold and min-max normalization to keep indicators comparable.
Non-AI democracy indicators were included so that, for a given level of AI capability, authoritarian states are penalized more heavily than democracies. This design decision follows a conversation with our colleague Jeba Sania about their Future of Life Institute case-study project [4], which found that the AI capability threshold at which democratic risks emerge varies by regime type: because authoritarian regimes are more prone to power concentration, even low-level AI poses risks there that it wouldn't in a democracy.
Partial weighting is used between and within drivers. "Partial" here means an indicator is weighted in a country's score only if it exists for that country. The decision not to penalize countries for missing indicators was made due to data limitations, but we think it relates to one of the methodological limitations in our index design (other limitations are discussed in detail on our website). We applied sensitivity tests for our indicators, and the ranking was almost always robust across different imputation, normalization, and weighting choices except when using geometric instead of linear aggregation.
In this section, we outline a few of our pilot’s findings. These are what we find most interesting, but we’re certainly open to sharing more if it arises as part of any resulting discussion.
The top ten are all liberal democracies
Our scores ranged from 86.2 (Denmark) to 28.6 (China), a spread of 57.6. The top ten, fairly predictably, are all classified as liberal democracies with resilient institutions that plausibly provide a structural buffer against the amplification mechanisms our framework identifies. That being said, this buffer is uneven across drivers. The polarization question
Of all the hypothesized drivers, polarization demonstrates the most interesting trend. For the top two thirds of countries (listed from highest to lowest by CORDA score, as partially pictured above), the driver is always the lowest value, whereas for the bottom third it almost always isn’t. Now, this reveals a couple of things. First, among liberal and electoral democracies, polarization tends to be the greatest structural weakness with respect to the other drivers in the basket. Second, and perhaps more crucially, the driver may be statistically unique from the other four; background testing revealed that the driver has far lower internal consistency and may be doing more cross-cutting, cross-driver work than first theorized.
Economic inequality and state capacity erosion are the biggest differentiators
D5, despite its shortcomings, reveals an interesting emergent separation between the top two thirds and bottom third. This isn't because its scores swing, they don't; D5 is the most homogeneous driver, with the lowest spread at 32.3 against roughly 60 for the rest. What separates the groups is where D5 sits relative to a country's other drivers, not its own value.
So what drives the bottom third's low overall scores? Economic inequality and state capacity erosion. Among authoritarian regimes, all drivers score low. Hybrid regimes, by contrast, tend to have relatively healthy scores for information environment, elite defection, and polarization while scoring comparatively weak on economic inequality and state capacity erosion.
The importance of design choices
Some readers might have already been thinking “why choose arithmetic aggregation for scoring rather than geometric?” The simplest answer is, well, simplicity. Our index rewards overall strength by taking the simple mean to calculate the CORDA score. However, rewarding driver balance is a perfectly arguable approach, and one that we continue to grapple with.
Take New Zealand. Under arithmetic aggregation, it sits at 10 of 27 per our ranking. Under geometric aggregation, it falls to 19 given a relatively weaker polarization score. By making the choice to employ arithmetic over geometric aggregation, we encode the assumption that strong drivers can compensate for the others. While not a pure limitation (some of which we address in a later section), it demonstrates an example of how extractable insights can swing widely following subtle changes in theoretical and methodological approach.
Initially, for our state capacity erosion driver, we defined the risk as AI outpacing state regulation capacity and regulatory fatigue by tech firms. This definition does not include the risk of AI integration in the executive branch. Cullen O'Keefe,[5]Alan Z. Rozenshtein,[6]Christoph Winter, [7] and the Forethought Foundation have argued in related works that AI integration within the executive branch could lead to the removal of bureaucratic checks and balances that democracy depends on. If an AI system is functionally beholden to the head of the state, gradual integration could lead to extreme concentrations of power. These scenarios are hard to operationalize given that they rest on strong assumptions about how a system is deployed. Still, we think it’s worth spending time thinking about their specification and potential dynamics that make them possible to produce a more comprehensive index.
As it stands, our index treats AI as a tool. Future work should consider whether that assumption holds for more agentic systems, ones better modeled as actors with their own objectives, which would change what the index needs to measure. Further, there is a need to account for the positive impacts of AI in the index construction; things like AI improving epistemic and collective sense-making via deliberative and fact-checking tools are worth exploring. Another positive impact would be the improvement of state capacity, but operationalizing this should consider executive AI risks.
Whilst we are planning to scale up our project, we acknowledge the limitations of our work, and so should others (and we welcome any individuals or teams to replicate our work). Here are a few we’ll be working on going forward:
Theoretical framework
Our five drivers are not comprehensive or definitive. We think it’s worth spending time thinking about how new threat models can be constructed, like AI magnifying existing authoritarian capacities (e.g., surveillance, loyalty screening and securitization) by unlocking feasibility bottlenecks.
*Systemic risks *
Our framework considers strategic risks where political agents take advantage of AI impacts and capabilities. However, we think it’s plausible for AI to create risks to democracy without malicious actors. For example, the impact of gradual disempowerment and erosion of critical thinking can change the democratic culture without someone intending it to do so.
Methodological future research direction
In the methodological realm, we hope to expand our country coverage, rearrange our indicator grouping to match statistical findings, and verify our data missingness audits.
We’d like to thank the following individuals for their consultation:
Jeba Sania, Michael L. Bąk, and Omer Alawed for thinking with us about our theoretical framework. We are also happy to talk to others who are interested in thinking with us about how AI will impact democracy.
This post's content is based on collaborative work by Coleman Snell, Eilaf Mohamed, Jason Hung, and Casimir Wypyski.
Admittedly, "vectors of democratic backsliding" is more accurate. For the sake of presentation we've just noted this to avoid any confusion; the term "drivers of political change" is maintained for internal consistency. It's very much an 'alpha version' artifact.
This report is as yet unpublished; support for our methodology emerges from our conversation with Jeba rather than an official, finalized research output.