数月前，一位纽约金融家表示他在 2025 年暑期实习生中遇见了第一批「真正的 AI natives」；但当资深金融人员追问这些实习生的想法后，发现其见解令人警惕地肤浅。因此该公司减少了返聘名额，并将招聘重心从 STEM 转向人文学科，原因是「我们要的是批判性思维，不只是 AI」。这个个案虽不具代表性，却映照出一个趋势：在金融业掀起 AI 热潮的同时，AI 并未立即实现科技乐观论者所预期的盈利天堂，也未立即引发末日预言中的崩溃；Nvidia 的调查显示 89% 的金融高管认为 AI 提升收入，73% 认为其对未来成功关键，但仅是部分指标而非全域答案。

风险面已成为监管核心。文章提到，金融稳定委员会警告，主要来自大规模私募资金支持 AI 资料中心的「瞠目结局」风险在本周被放大；IMF 亦警示网路攻击可引发金融体系危机。FSB 进一步指出，电脑交易「拥挤同调」、对少数云端供应商的集中依赖、以及模型幻觉，都是系统性风险。监管面已启动高层会谈，US Treasury Secretary Scott Bessent 与 Federal Reserve Chair Jay Powell 针对 Anthropic 的 Mythos 进行讨论，Federal Reserve Vice-chair Michelle Bowman 则呼吁在银行风险模型评估中提高灵活性。英国 FCA 也创新推出 supercharged sandbox，延续与 Nvidia 合作的试点经验，提供有限条件下免费算力与资料，供金融科技在受管控的安全环境中实测。

第三项关键是「落实率」与「效益测量」。Cambridge 的 Judge Business School 调查（628 家金融与 AI 机构）显示，81% 私营金融机构正在使用 AI，其中约 50% 采用 agentic AI，主要模型来源为外部供应商且以 OpenAI 为首。Yet, 76% 的大型金融集团、55% 的所有受访机构，仍难以衡量 AI 价值；只有 40% 报告获得利润提升，43% 则认为没有改变。仅约四分之一预期裁员，58% 反而预期增加聘用或再培训。Hyland 指出仅 45% 的企业认为 AI 成效符合预期。总体而言，AI 的部署存在时间差与认知差距，尤其监管部门的导入与知识常约为私营机构的一半。金融系统最终取胜者将是那些培养既懂 AI 又有批判性思考的人才者，且 2026 年实习生成为观察重点。

A few months ago, a New York financier said his 2025 summer interns were the first true AI natives he had seen. Their digital fluency looked impressive at first, but their ideas were found shallow when senior staff probed deeper, so the firm cut return offers and shifted hiring from STEM toward humanities, stating: “We want critical thinking, not just AI.” This anecdote is small, but it illustrates a broader pattern in finance: while AI hype is high, it has not yet delivered the profit utopia promised by tech evangelists, nor has it triggered the doom some warned about. A Nvidia survey cited in the text found 89% of finance executives say AI raises revenue and 73% call it crucial for future success, showing mixed but still limited confidence, not certainty.

Regulatory risk is now central. The Financial Stability Board is warning that heavy private credit supporting AI data centres can likely produce defaults, and the IMF notes that cyberattacks could trigger a systemic crisis. The FSB also cites herd behavior in automated trading, concentration on a few cloud providers outside strong oversight, and model hallucinations as systemic vulnerabilities. Policy engagement is rising: US Treasury Secretary Scott Bessent and Federal Reserve Chair Jay Powell met senior financiers over Anthropic’s Mythos model; Federal Reserve Vice-chair Michelle Bowman called for more flexibility in assessing banks’ risk models. In the UK, the FCA introduced a supercharged sandbox after a prior pilot with Nvidia, offering selected fintechs free computing power and data to test AI in a controlled, accountable environment.

A larger gap is between AI rhetoric and measurable outcomes. In a survey of 628 finance and AI firms, 81% of private-sector finance groups use AI and about half use agentic AI, with most relying on external models and OpenAI the top provider. Yet 76% of large financial groups and 55% of all respondents say they struggle to measure AI value. Only 40% report profit increases, while 43% see no change; only about one quarter expect job losses, but 58% expect more hiring or reskilling. Other surveys echo this: Hyland reports just 45% of businesses say AI delivers expected results. The evidence suggests deployment is still uneven and lagging in understanding, with regulators often less advanced than the institutions they supervise. The likely winners are firms that build AI-native talent with critical thinking, and the 2026 intern cohort becomes the key group to watch.