Ethan Mollick: AI labs and policymakers should build national-scale public-good AI projects now

emollick · x · 2026-11-06

Ethan Mollick argues that too little effort goes into near-term, achievable national-scale AI projects for public good — from AI labs, policymakers, or NGOs. He cites obvious candidates: universal tutoring, healthcare second opinions, financial literacy/scam avoidance, and accessible government services. Drawing on 20 years building tutoring and financial-literacy programs, he says these hard problems were always tech-limited and now real gains are achievable without AI beyond current models — though careful design and studies are needed, and some efforts will fail.

Attached evidence: an MIT/Stanford paper showing GPT-5.2 and Gemini 3 Flash financial advice moves most respondents closer to life-cycle-optimal decisions (though advice differences could accumulate into 4-5% retirement wealth gaps); a Lancet-published study of Google's AMIE in primary care (98 patients, zero safety stops, PCPs found it helpful in 33 of 44 reviewed cases); and an RCT showing AI-as-tutor boosts learning that persists a week later, while AI-as-automation gains fade.

---

Summary in Chinese

Ethan Mollick 感叹，无论 AI 实验室还是政策制定者与 NGO，都太少有人在琢磨用当下就能实现的 AI 项目做国家级的公共利益工程。

他列出几个数据已经支持的明显方向：

全民辅导：AI 助教支持课堂学习
医疗第二意见：为患者提供低成本问诊参考
金融素养与防诈骗：普及理财知识、识别骗局
政务服务可及性：让公共服务的获取变简单

他基于自己 20 年做辅导与金融素养项目的经验指出，这些难题过去一直受限于技术，如今 AI 有望带来实质突破——而且不需要超出当前模型能力的 AI 就能落地。当然需要严谨的设计与研究、也可能失败，但若做对了，收益巨大。

配图佐证了「现有研究已铺路」的判断：一篇 MIT/斯坦福论文发现 GPT-5.2 与 Gemini 3 Flash 的理财建议能让多数受访者更贴近生命周期理论的最优决策；Google AMIE 在《柳叶刀》发表的前瞻性研究显示 98 名患者零安全中断、医生认为 44 例中 33 例有助诊前准备；另有 RCT 表明用 AI 做「辅导」的学生成绩提升且一周后仍保留，「自动化代写」则增益消退。

Original post →

Ethan Mollick: AI labs and policymakers should build national-scale public-good AI projects now

More from AGI Musings