AI 预测系统对伊朗在 2026 年底前出现政权崩溃或更替的概率估计为 20%，高于许多专家的判断；与此同时，近两年的机器学习和大语言模型进展，使冲突预测重新受到关注。当前趋势是把犯罪、公共卫生、罢工、天气、经济、民主倒退、社交媒体以及卫星和无人机图像等数据，综合用于早期预警。

最成熟的案例之一是 ACLED 的 CAST：它由全球约 150 名研究人员持续跟踪骚乱、镇压、帮派战、军事袭击等事件，并在有较好数据时，可对未来最长 6 个月的有组织政治暴力做出有效预测。IOM 的新模型则在 4 月 1 日对索马里未来 3 个月的流离失所人数给出 304,362 人的预测，到 4 月 20 日时，误差据称可能只有几个百分点。

尽管如此，质疑仍然存在，尤其是在预测新冲突爆发方面。用户中也有人认为，诸如极端高温、油气或钻石等“风险放大器”确实能提高模型效果，但虚假信息、数据缺口和先发制人的镇压会削弱预测价值；因此，最新趋势是把模型与人类来源、加密群聊和更广数据网络结合使用。

An AI forecasting system put the chance of regime collapse or replacement in Iran by the end of 2026 at 20%, higher than many experts would estimate; at the same time, progress in machine learning and large language models over the past two years has renewed interest in conflict prediction. The main trend is to combine crime, public health, labor strikes, weather, the economy, democratic backsliding, social media, and satellite or drone imagery into early-warning systems.

One of the most developed examples is ACLED’s CAST: it is powered by about 150 researchers worldwide who track riots, crackdowns, gang warfare, military attacks, and other violence, and when good data exist, it can predict organized political violence up to 6 months ahead. The IOM’s new model, meanwhile, forecast on April 1 that drought, floods, and fighting would displace 304,362 people in Somalia over the next 3 months, and by April 20 the error was said to be only a few percent.

Even so, skepticism remains, especially about predicting the onset of new conflicts. Some users say risk amplifiers such as heatwaves and extractable resources like oil or diamonds can improve models, but disinformation, data gaps, and the possibility of preemptive repression weaken predictive value; the latest response is to combine models with human sources, encrypted group chats, and broader data networks.

Source: AI models are being used to predict conflict

Subtitle: Good data are hard to come by

Dateline: 5月 14, 2026 11:25 上午