platform algorithms have been implicated in spreading inflammatory content during ethnic conflicts.73 Although the causal relationship between exposure to algorithmically curated content and changes in political attitudes or behavior is not fully established, algorithmic curation appears to exacerbate existing polarization in societies with weak institutions, limited media literacy, and deep ethnic or religious cleavages. Even if platforms do not create these divisions, they can intensify them by making emotionally charged content more visible than content encouraging compromise or mutual understanding. Algorithmic curation is not the only way AI shapes public discourse. As citizens, journalists, civil servants, and analysts increasingly draft, summarize, and reason with the same chatbots—most of them trained abroad—the range of positions people express may narrow toward whatever those models treat as default. Controlled experiments show that exposure to opinionated language models can shift users’ own expressed views,74 that groups writing with a shared model produce less diverse text,75 and that the same flattening appears in creative output.76 These are short-term effects measured in experiments; whether they aggregate into societywide homogenization is not established. But the possibility matters most where pluralism, rather than consensus, is what holds a fragile political settlement together. In such settings, homogenization is a risk distinct from disinformation, and addressing it points toward supporting a plurality of models so that no single foreign-trained system becomes the default voice in a national public sphere. These challenges are especially salient in lower-​ income countries because platforms have made limited investments in content moderation. For instance, Meta provides content moderation for only a fraction of the world’s 7,000 ­languages.77 When AI systems trained predominantly on high-resource languages make decisions about content in more regional or local languages like Amharic, Burmese, or Swahili, the error rates are predictably higher and result in both underdetection of hate speech and overflagging of legitimate discourse. Polarization limits citizens’ willingness to process information that contradicts their group identity.78 Consequently, in contexts with significant societal divisions, algorithmically amplified polarization can make cooperation across groups even more difficult; weaken institutional accountability; and erode trust in government, media, and expertise. However, recent evidence suggests that low-cost interventions can reduce susceptibility to misinformation. In South Africa, biweekly fact-checks delivered via WhatsApp over six months improved citizens’ ability to identify new misinformation, particularly when they received financial incentives to review those fact-checks.79 In Mexico, simply providing voters with accurate information about local government performance during the COVID-19 pandemic backfired: Providing voters with unfavorable information about relative COVID-19 cases and deaths increased the vote share of incumbent candidates.80 A short antipolarization message delivered alongside the same information reversed the backlash, restoring the link between performance and electoral accountability. If governments can identify which citizens are susceptible to which messages, they can target disinformation and curtail it instead of disseminating it. Conversely, governments can use concerns about disinformation to justify expanded surveillance. This interaction means that the individual dynamics described earlier may compound in contexts in which institutional checks on both surveillance and information manipulation are weak. But even when governments are well intentioned, cybersecurity risks may pose challenges to their efforts, which can perpetuate mistrust in AI. AI’s Political Impact: Reshaping Power Within and Across Countries 219