AI’s political stakes are rising fast enough that the old framing—Anthropic versus OpenAI, model A versus model B, lab rivalry as the main story—has started to feel like a category error. The more capable these systems become, the less the debate is about who built the better chatbot and the more it becomes about who gets to set the rules for how power is exercised through software.
That shift is already visible in the way governments, regulators, and even courts talk about AI. The language has moved from “innovation” and “safety” in the abstract to questions that sound uncomfortably like governance: Who is accountable when an AI system influences a decision? What counts as acceptable risk when the output can shape public opinion or operational security? How do you audit a system whose behavior changes with updates, tools, and context? And perhaps most importantly, how do you coordinate across borders when the systems themselves don’t respect them?
In other words, the timeline has changed. When AI capabilities begin to affect political processes—public debate, election infrastructure, intelligence workflows, crisis response, and the everyday mechanics of policy implementation—the response can’t be left to individual companies or isolated national efforts. It becomes a collective action problem.
Not because everyone suddenly agrees on what “good governance” looks like, but because the consequences are shared. If one provider deploys a system that can be used to manipulate information at scale, the harm doesn’t stay inside that provider’s ecosystem. If one country sets a permissive standard for high-risk uses, other countries will feel pressure to compete—or to compensate—regardless of their own preferences. If one lab releases capabilities faster than oversight can adapt, the political system that absorbs the shock will be dealing with it long after the lab has moved on to the next iteration.
This is why the “model wars” narrative is increasingly inadequate. It treats AI as if it were a consumer product with incremental feature differences. But AI is becoming infrastructure for persuasion, analysis, automation, and decision support. Infrastructure doesn’t just change industries; it changes leverage. And leverage is political.
A new kind of influence: from content to capability
For years, the public conversation about AI’s political impact focused on content generation: synthetic text, deepfakes, and the ability to flood channels with persuasive material. Those concerns remain real, but they’re no longer the whole picture. The more consequential shift is that modern AI systems are not only producing content—they are compressing time and lowering the cost of complex tasks.
That matters politically because politics is, at its core, a contest over speed, framing, and coordination. Campaigns, advocacy groups, and state actors all rely on rapid research, message testing, and operational planning. When AI reduces the friction of those activities, it changes the competitive landscape. It also changes the asymmetry between actors with different resources. A small group with access to powerful tools can behave more like a large organization. A well-funded actor can move faster than institutions designed for slower cycles of deliberation.
The result is not simply “more misinformation.” It’s a broader acceleration of influence operations. AI can help draft arguments, tailor them to specific audiences, translate them across languages, and generate variations optimized for engagement. It can also help adversaries probe defenses—testing what filters catch, what narratives slip through, and which platforms respond slowly. Even when the content is not overtly deceptive, the ability to produce plausible, targeted messaging at scale can distort the informational environment in ways that are hard to measure and harder to reverse.
And then there’s the second-order effect: AI doesn’t just amplify messages; it can shape decisions. Decision-makers increasingly use AI for summarization, analysis, and drafting. That means AI outputs can influence what gets considered “evidence,” what is framed as “risk,” and what options appear feasible. In governance, perception is policy. If AI changes perception, it changes outcomes.
Governance is catching up, but not fast enough
Most regulatory frameworks were built for technologies that have clearer boundaries: a product ships, it behaves within known constraints, and compliance can be tested against a stable specification. AI systems are different. They evolve. They can be integrated into workflows that vary by organization. They can be augmented with tools—search, code execution, retrieval systems—that expand their practical reach beyond what a single model card might suggest.
This creates a governance mismatch. Oversight mechanisms often assume that risk can be assessed at deployment time. But with AI, risk can increase after deployment as models are updated, prompts are refined, and new capabilities are added through integrations. A system that seems manageable today can become more powerful tomorrow without the same level of public scrutiny.
That’s one reason the conversation is moving toward ongoing monitoring rather than one-time certification. It’s also why accountability is becoming a central theme. If AI influences political processes, then someone must be able to answer for how it was used, what safeguards were in place, and what went wrong when safeguards failed.
But accountability is not a simple matter of assigning blame to a single actor. In many real deployments, responsibility is distributed: the model provider supplies the underlying system; the platform operator provides distribution; the integrator designs the workflow; the end user decides how to apply it. Political consequences emerge from the combined system, not from any single component.
So the question becomes: how do you design accountability structures that match the technical reality?
Collective action: standards, coordination, and shared expectations
The phrase “collective action” can sound vague, but it points to concrete needs.
First, shared standards. Without common benchmarks and definitions, every regulator and every company ends up reinventing the wheel. One jurisdiction may treat certain uses as permissible while another treats them as prohibited. One company may claim a safety approach that another company cannot verify. Shared standards don’t eliminate disagreement, but they reduce the space for strategic ambiguity.
Second, coordinated policy. AI capabilities cross borders instantly. Information operations do too. If one country tightens rules on certain high-risk uses, actors can shift to jurisdictions with weaker enforcement. Coordinated policy helps prevent a race to the bottom and reduces the incentive to exploit regulatory gaps.
Third, ongoing public oversight. Transparency isn’t just about publishing technical details. It’s about enabling independent evaluation of claims, monitoring real-world impacts, and ensuring that the public can understand how AI is being used in political contexts. Oversight also includes the ability to investigate incidents—when AI systems are implicated in election interference, critical infrastructure disruptions, or large-scale manipulation campaigns.
Fourth, international coordination on incident response. When AI-enabled attacks occur, the response needs to be timely and interoperable. That means shared reporting formats, shared threat intelligence practices, and agreed protocols for how to communicate risks without amplifying panic or giving adversaries useful information.
None of this requires governments to agree on every technical detail. It requires agreement on process: how risk is assessed, how incidents are reported, how updates are handled, and how accountability is enforced.
Why “company competition” can’t solve the governance problem
It’s tempting to believe that market competition will naturally improve safety. After all, if one provider takes a stronger stance, others might follow. But governance problems rarely resolve themselves through competition alone, because the incentives are misaligned.
Companies want adoption. Adoption depends on performance, speed, and usability. Safety measures can slow deployment or reduce flexibility. Even when companies care about safety, they face pressure to meet customer demand and maintain competitiveness. Meanwhile, the political harms of AI are often externalities: the people affected are not necessarily the ones paying for the system.
This is where collective action becomes essential. Governance requires mechanisms that are not purely market-driven. It requires rules that constrain behavior even when doing so is inconvenient for individual actors. It also requires enforcement capacity—something markets don’t provide.
There’s also a deeper issue: political legitimacy. When AI influences public life, the public needs confidence that the system is governed fairly. That confidence cannot be manufactured solely through corporate assurances. It has to be grounded in institutions that the public recognizes as legitimate: courts, regulators, auditors, and civil society organizations with the ability to scrutinize claims.
Civil society’s role: transparency, monitoring, and pressure
If governments and companies are the obvious actors, civil society is often treated as an afterthought. But in practice, civil society can be the early warning system. It can monitor real-world effects, document patterns of misuse, and push for transparency where official reporting is incomplete.
Civil society also plays a crucial role in translating technical risk into public understanding. Political consequences are not only about whether harm occurs; they’re about whether the public can recognize what’s happening and demand corrective action. When AI systems are used to manipulate narratives, the ability to detect and explain manipulation becomes part of democratic resilience.
That doesn’t mean every activist group will be technically competent or unbiased. It means the ecosystem needs multiple independent observers. The alternative is a closed loop where only insiders see the evidence and only insiders control the narrative.
A unique take: governance is becoming a “capability arms race” problem
One way to think about the governance challenge is to treat it like an arms race—but not between countries in the traditional sense. It’s between capabilities and oversight.
As AI capabilities improve, the window for effective governance shrinks. Oversight systems—laws, regulatory frameworks, auditing practices, institutional expertise—tend to evolve more slowly than model development. That creates a recurring pattern: new capabilities arrive, society adapts imperfectly, harms surface, and then governance catches up. The cycle repeats.
The political danger is that the harms can become normalized before governance stabilizes. Once AI-enabled influence operations become routine, it becomes harder to roll back. Once AI-assisted decision-making becomes embedded in government workflows, it becomes harder to unwind. Once public trust erodes, rebuilding it takes longer than the technology’s improvement cycle.
So the goal shouldn’t be to “stop innovation.” It should be to shorten the governance lag. That means investing in institutional capacity: technical expertise inside regulators, standardized evaluation methods, and mechanisms for continuous oversight rather than one-time approvals