尽管地缘政治的波动性日益增加，大型数据集和人工智慧模型的进步正使世界变得越来越可预测。一个主要的例子是短期气象学，其中现代 AI 驱动的七天天气预报达到了与 1980 年三天预报相同的预测品质。为了预测以人类为中心的事件，Anthony Vinci 确定了四种主要方法：个人超级预测员、集体组织智慧、预测博弈市场，以及合成这些输入以产生更深层见解的受训 AI 模型。

应用这种 AI 整合方法，Vico Technologies 利用生成式 AI 模型来分析复杂的数据集并预测未来事件。在一个显著的案例中，Vico 的模型预测美国对伊朗发动袭击的机率在 3 月 31 日前为 89%，显著优于 Polymarket 在 2 月 28 日实际袭击发生前估计的 63.5% 机率。Vico 报告的 Brier 评分为 0.15 以下，反映出超越几乎所有人类超级预测员的预测准确度和信心水准。

尽管如此，AI 预测仍面临重大脆弱性，包括来自刻意「大型语言模型修饰（LLM grooming）」的数据污染，以及可能引发模型自噬失调（Mad）的合成数据。此外，预测市场仍然容易受到操纵，正如记者 Emanuel Fabian 被威胁修改报导以影响 Polymarket 派彩的事件所显示的。像 Carissa Véliz 这样的哲学家警告说，数据是人类的产物，且预测往往反映了下游的权力关系，这需要模型不断适应以保持其效用。

Despite growing geopolitical volatility, advancements in large datasets and AI models are rendering the world increasingly predictable. A prime example is short-term meteorology, where modern AI-driven seven-day weather forecasts achieve the same predictive quality as three-day forecasts did in 1980. To forecast human-centric events, Anthony Vinci identifies four primary methods: individual superforecasters, collective organizational intelligence, prediction betting markets, and trained AI models that synthesize these inputs to generate deeper insights.

Applying this AI-integrated approach, Vico Technologies utilizes generative AI models to analyze complex datasets and forecast future events. In a notable case, Vico's model predicted a US attack on Iran with an 89% probability by March 31, significantly outperforming Polymarket's 63.5% probability estimate prior to the actual attack on February 28. Vico reports a Brier score below 0.15, reflecting a predictive accuracy and confidence level that surpasses almost all human superforecasters.

Nonetheless, AI forecasting faces significant vulnerabilities, including data pollution from intentional "LLM grooming" and synthetic data that can trigger model autophagy disorder (Mad). Additionally, prediction markets remain susceptible to manipulation, as illustrated when journalist Emanuel Fabian was threatened to alter a report to affect Polymarket payouts. Philosophers like Carissa Véliz caution that data is a human artifact and predictions often reflect downstream power dynamics, requiring models to continuously adapt to stay relevant.