新罕布什尔州现已“由 Gemini”API 提供支持
4 分•作者: sans_souse•7 个月前
这是一个重要的问题，需要我们对技术发展保持前瞻性思考，并关注科技界内外法律与治理的走向。我个人认为，我们正明显地走向一个不可避免的“政府”，即大型科技公司与政府本身的融合。
新罕布什尔州就业保障局（NHES）推出了“AI裁决助手”来处理失业救济申请，旨在简化事实收集并加快福利发放流程。该系统通过对话界面与申请人互动，收集关于他们离职的详细信息。
最近的截图，见：https://imgur.com/a/34FtrCC，来自公开的实时门户网站（adjudication.assistant.nhuis.nh.gov），证实该系统“由Gemini提供支持”（由谷歌提供）。这代表了将商业、面向公众的大型语言模型（LLM）直接整合到核心州级政府行政职能中的首批案例之一（据我所知）。
虽然其既定目标是提高效率，但这种实施具有重要的技术和治理影响：
```
核心逻辑外包：
```
州政府不再仅仅将云提供商用于基础设施（IaaS）或软件（SaaS），而是将裁决逻辑的一部分外包给了谷歌的第三方AI模型（MaaS - 模型即服务）。将公民的案件总结给人工审核员的初始过程，正在由Gemini中的一个商业、闭源系统处理。
```
数据主权和安全：
```
该系统处理大量敏感的个人身份信息（PII）。现在，数据流扩展到州政府控制的服务器之外，延伸到谷歌的API端点。这引发了关于数据处理、包含PII的API调用保留策略，以及这些数据是否可用于训练未来模型等关键问题。攻击面从州政府运营的网站组件扩展到商业LLM的复杂安全环境。
```
问责制和“黑盒”：
```
NHES的常见问题解答确认，最终决定由人工裁决员做出。然而，这种人机结合的方式是基于AI生成的摘要。如果Gemini模型生成的有缺陷或有偏见的摘要导致福利被错误地拒绝，那么对于任何处于这种境地的人来说，问责链现在变得不明确，因为这仍处于早期阶段。
```
供应商锁定：
```
通过围绕特定的专有模型构建工作流程，州政府创造了显著的依赖性。未来迁移到不同的模型提供商或“内部解决方案”（旧方法）将需要大量的技术工作和再培训，从而赋予供应商可观的长期影响力。
在新罕布什尔州测试的这个新系统，是大型科技公司与州政府直接融合的一个案例研究。它突出了在使用强大、预先存在的模型所带来的即时效率提升，与依赖性、安全性和公共问责制相关的长期风险之间的权衡。
[来源]:
Route Fifty关于现代化推动的文章：https://www.route-fifty.com/artificial-intelligence/2025/03/new-hampshires-benefits-program-embraces-ai-amid-modernization-push/403448/
查看原文
This is an important issue when thinking ahead of the technological curve and where we are headed with Law and Governance <i>both in amd out of</i> the tech world. Personally I believe we are pretty clearly heading towards an inevitable 'Goovernment' which would be a merger of Big Tech and government itself.<p>New Hampshire Employment Security (NHES) has launched an "AI Adjudication Assistant" to process unemployment claims, aiming to streamline fact-gathering and speed up the benefits process. The system engages claimants in a conversational interface to collect details about their job separation.<p>Recent screenshots here; https://imgur.com/a/34FtrCC from the live public portal (adjudication.assistant.nhuis.nh.gov) confirm the system is "Powered by Gemini" (by Google). This represents one of the first direct integrations of a commercial, public-facing LLM into a core state-level government administrative function ( <i>afaik</i> )<p>While the stated goal is efficiency, this implementation has significant technical and governance implications:<p><pre><code> Outsourcing of Core Logic:
</code></pre>
The state is no longer just using a cloud provider for infrastructure (IaaS) or software (SaaS), but is now outsourcing a component of its adjudicative logic to Google's third-party AI model (MaaS - Model as a Service). The initial process of summarizing a citizen's case for a human reviewer is being handled by a commercial, closed-source system in Gemini.<p><pre><code> Data Sovereignty and Security:
</code></pre>
The system processes a high volume of sensitive Personally Identifiable Information (PII). The data flow now extends beyond state-controlled servers to Google's API endpoints. This raises critical questions about data handling, retention policies for API calls containing PII, and whether this data could be used for training future models. The attack surface expands from a state-run web site component to the complex security environment of a commercial LLM.<p><pre><code> Accountability and the "Black Box":</code></pre>
The NHES FAQ confirms that a human adjudicator makes the final decision. However, this human-in-the-loop is acting on a summary generated by the AI. If a flawed or biased summary from the Gemini model leads to a wrongful denial of benefits, the chain of accountability is now unclear for anyone in such a position, as this is all still in such early phases.<p><pre><code> Vendor Lock-In:
</code></pre>
By building the workflow around a specific proprietary model, the state creates a significant dependency. Migrating to a different model provider or an "in-house solution" (the old way) in the future would require substantial technical effort and retraining, granting the vendor considerable long-term leverage.<p>This new system being tested in New Hampshire serves as a case study in the direct merger of Big Tech and state governance. It highlights a trade-off between the immediate efficiency gains of using powerful, pre-existing models and the long-term risks associated with dependency, security, and public accountability.<p><i>[Sources]:</i>
<i>Route Fifty article on the modernization push: https://www.route-fifty.com/artificial-intelligence/2025/03/new-hampshires-benefits-program-embraces-ai-amid-modernization-push/403448/</i>