内容来源：https://www.wired.com/story/documents-reveal-palantir-irs-contract-fraud-clean-energy-credits/

内容总结：

美国国税局引入大数据工具强化税务稽查，帕兰提尔公司获1.8亿美元合同

据公开记录显示，美国国税局（IRS）去年向硅谷大数据公司帕兰提尔（Palantir）支付180万美元，用于改进一款定制化税务稽查工具。该工具旨在帮助税务机构更精准地识别高价值审计案件、追缴欠税及潜在刑事调查对象。

国税局在签约文件中承认，其现有系统已严重落后。该机构依赖“超过100个业务系统和700种方法”筛选案件，这些系统历经数十年累积，已形成碎片化格局，导致效率低下、资源重复投入且案件筛选质量不佳。

帕兰提尔为此开发的“筛选与分析平台”（SNAP）试图整合国税局分散的数据库，通过分析非结构化数据（如合同、交易记录等），辅助稽查人员发现以往可能遗漏的税务异常信号。目前该平台仅处于试点阶段，但国税局已表现出深化合作的意向。公开合同记录显示，自2014年以来，国税局向帕兰提尔授予的合同总额已超过2亿美元。

试点项目重点关注三类税务领域：自然灾害受灾者的税收减免申请、居民清洁能源税收抵免（如安装太阳能板）以及高价值赠与税申报（如艺术品、股权等资产的赠与）。税务专家指出，若SNAP能有效解析赠与资产估值依据、交易双方关系等非结构化信息，或将提升对复杂涉税行为的监管能力。

长期以来，国税局主要依靠“判别函数得分”（DIF）机制选择审计对象，但该算法细节从未公开。近年来，该机构亦尝试与加密货币平台合作、分析社交媒体信息等方式拓展数据源，但自身技术升级屡受挫。有学者指出，国税局自1960年代以来“从未完全实现过系统现代化”，部分原因在于其公众形象不佳、易成为政治博弈的牺牲品，导致资源不足且缺乏持续的政治支持。

近年国税局人员流失严重：2025年2月至7月间，超过2.5万名员工离职或提前退休。同时，高层领导更替频繁，致使多年期重大项目难以持续推进。在此背景下，国税局转向与私营科技公司合作，试图以外部技术突破内部僵局，但其实际成效与数据应用边界仍待观察。

中文翻译：

根据《连线》杂志通过公开记录申请获取的文件显示，美国国税局去年向帕兰提尔公司支付了180万美元，用于改进一款定制工具。该工具旨在帮助税务机构识别"最高价值"案件，以进行税务审计、追缴欠税及潜在的刑事调查。

签约时美国国税局表示，其用于筛选纳税人可能误报税款或欠税案件的系统包含"超过100个业务系统和700种方法"，这些系统是"数十年间"逐步建立的。随着识别潜在税务差异的复杂性增加，该机构称其系统效率日益低下，亟需寻找解决方案。

国税局在一份概述合同范围的文件中写道："这种碎片化格局可能导致诸多不良后果，包括但不限于工作与成本重复、对覆盖范围缺口认知不足，以及案件筛选效果欠佳。"这份文件同样由《连线》杂志获取。

帕兰提尔为此问题开发的定制工具名为"筛选分析平台"（简称SNAP），旨在帮助国税局优化潜在欺诈案件的识别流程。文件显示，该软件目前仅作为试点项目的一部分投入使用。帕兰提尔公司与美国国税局均未回应置评请求。

政府合同记录表明，虽然帕兰提尔开发SNAP的具体时长尚不明确，但国税局自2014年起就已采购该公司技术。帕兰提尔累计从国税局获得超过2亿美元的合同与应付款项。文件显示，该机构目前有意深化与帕兰提尔的合作关系。

SNAP将如何与国税局现有技术系统整合尚不清楚。与帕兰提尔其他工具类似，该系统很可能架设于国税局高度分散的数据库之上，协助审计人员发现税务申报中可能被忽略的风险信号。合同显示国税局有意实现软件现代化，并寻求帕兰提尔协助。其中一份文件指出，帕兰提尔的SNAP试点项目旨在从"支持文件中的非结构化数据"中提取"关于合同、车辆与供应商的关键信息"。

国税局要求帕兰提尔针对现行税法部分条款设计三种"案件筛选方法"，涵盖灾区索赔（面向自然灾害受害者的税收减免形式）、住宅清洁能源税收抵免（抵消太阳能板或风力涡轮机等安装成本的抵免计划），以及709表格赠与税申报表（人们在赠与艺术品、股票或公司实体等贵重资产时需填写的文件）。

霍夫斯特拉大学专注于赠与税与遗产税研究的米切尔·甘斯教授指出，若SNAP正在分析支持文件中的非结构化数据，可能正在审查那些对赠与财产进行"充分披露"的表格。国税局规定此类披露必须包含财产价值评估依据的"详细说明"，以及赠与双方的关系信息。

甘斯举例说明，若某人将私营企业赠与他人，披露材料需包含企业评估的佐证信息，例如"资产负债表、净收益报表、经营业绩与股息记录"。

扬斯敦州立大学会计与金融学教授埃丽卡·诺伊曼补充道，Venmo等转账应用的公开记录，以及Etsy和Depop等网站上的公开店铺信息，同样可能包含国税局感兴趣的非结构化数据。

若帕兰提尔的SNAP工具在筛选审计案件时需参考Venmo或Depop的数据，前提是国税局已掌握这些信息。合同文件明确规定，该机构仅要求帕兰提尔使用"当前SNAP系统中的现有数据"。

数十年来，国税局确定审计案件的主要方式是为每位纳税人计算"判别信息函数"（简称DIF）分数。据国税局解释，"分数越高，审计可能性越大"。诺伊曼指出，该分数的计算方式如同黑箱，但研究人员普遍认为，国税局会比对当前纳税申报与过往最终触发审计的申报记录之间的相似性。

诺伊曼曾研究国税局为改进案件筛选流程尝试的其他方法，包括与Coinbase等公司合作分析加密货币交易信息，以及挖掘社交媒体公开帖子，寻找个人或企业可能少报收入的线索。

国税局也曾多次尝试更新自有技术，但诺伊曼表示其始终受技术难题困扰。她解释道："自1960年代以来，该机构基本上从未实现过完全成功的现代化改造。"

联邦政府其他部门同样面临现代化困境。例如政府工作人员申请退休时，仍需部分依赖存储在宾夕法尼亚州乡村石灰岩矿洞中的纸质文件办理手续。

但诺伊曼指出，国税局技术升级面临一个特殊阻碍：该机构在公众中极不受欢迎，使其成为政客争取选民的便利攻击目标。因此她认为，往往"缺乏为国税局改革倾尽全力的政治意愿"。

在特朗普执政期间，这种态势导致国税局资源缩减与人员裁减。2025年2月，该机构约有10.3万名员工；到2025年7月，已有超过2.5万人辞职或接受延期辞职及提前退休方案。

国税局领导层的任期通常较短，这意味着多年期大型项目难以持续推进或完成。诺伊曼表示："局长更替速度极快——近年来更是明显加速。但当资源不足且无人强力推动时，改革举措确实举步维艰。"

英文来源：

The IRS asked Palantir to craft three “case selection methods” related to parts of the existing tax code. The options included disaster zone claims, a form of tax relief for natural disaster victims, Residential Clean Energy Credits, a tax credit program that offset the cost of installing things like solar panels or wind turbines, and Form 709 Gift Tax Returns, which people may have to fill out when they give away valuable things like like artwork, stocks, or corporate entities.
Mitchell Gans, a professor at Hofstra University focused on gift and estate taxes, says that if SNAP is analyzing unstructured data from supporting documents, it may be examining forms providing “adequate disclosure” of property being gifted to another person. The IRS stipulates that these disclosures must include “a detailed description” of how the property's value was determined, and the relationship between the giver and recipient.
Gans says that if, for instance, a person gives someone else a private business, the disclosure would need supporting information about how it was appraised, such as “balance sheets and statements of net earnings, operating results, and dividends.”
Erica Neuman, an accounting and finance professor at Youngstown State University, adds that public logs from money transfer apps like Venmo, as well as public storefronts on websites like Etsy and Depop, could also contain unstructured data of interest to the IRS.
If Palantir’s SNAP tool were to factor in data from Venmo or Depop when selecting audit cases, the IRS would have to already possess it. The contract documents state that the agency only wants Palantir to use “existing data in SNAP today.”
For decades, the primary way the IRS decided which cases to audit was by calculating what is known as a Discriminant Information Function (DIF) score for each taxpayer. According to the IRS, “the higher the score, the greater the audit potential.” Neuman says that the way the scores are calculated is a black box. But researchers generally agree that the agency looks for similarities between current tax filings and past ones that eventually led to audits.
Neuman has studied other methods the IRS has experimented with to improve its case selection process, including contracting with companies like Coinbase to analyze information about crypto transactions, and mining public social media posts for clues that an individual or business may be underreporting their income.
The IRS has also tried updating its own technology several times, but Neuman says it has been plagued by technical difficulties. The agency “basically never had a full successful modernization since the 1960s,” she explains.
Other parts of the federal government have also struggled to modernize. When government workers apply for retirement, for instance, they have to go through a process that still partially involves paper documents stored in a limestone mine in rural Pennsylvania.
But Neuman says there’s one thing that has made it especially difficult for the IRS to upgrade its technology: The agency is incredibly unpopular with the public, making it a convenient punching bag for politicians looking to score points with voters. As a result, Neuman says, there's often “a lack of political will to die on the hill of the IRS.”
Under the Trump administration, that dynamic has translated into shrinking resources and staffing cuts at the IRS. In February 2025, about 103,000 people worked for the agency. By July 2025, more than 25,000 people resigned or accepted deferred resignation and early retirement offers.
Leaders at the IRS also tend not to stay in the job for very long, which means it can be difficult for big, multiyear projects to stay on track or get finished. “The commissioner turnover is incredibly quick—and that, of course, has accelerated in the most recent years,” Neuman says. “But it's just hard to get things done when there's not enough resources, and no one is really pushing that hard for it.”