真爱趁现在
A GT investigation examines the funding, government ties and questionable business practices of seven US entities subject to China’s countermeasures_我的网站
一 | （来源：麻省理工科技评论）
以后找工作时，你的简历很可能还没到招聘经理手里，就已经先被 AI 看过一遍了。 A piece of cotton fabric, an AI supply-chain map, a "human rights" report and an electronics certification document appear to have little in common.
Behind them, however, are the US entities that have helped build and operate Washington's machinery of restrictions against China - tracing materials linked to Xinjiang region, mapping supply chains for enforcement, producing reports used to advance "forced labor" allegations, converting those allegations into corporate compliance pressure, and lobbying to exclude Chinese laboratories from the US market.
On August 5, China responded.
China's Ministry of Commerce announced the countermeasures against Applied DNA Sciences, Inc., Stratum Reservoir, LLC., Altana Technologies, Inc., Responsible Business Alliance, Verite Group, Inc. and Human Rights in China, the Xinhua News Agency reported.
These countermeasures, which took effect on Wednesday, were adopted in response to the US sanctions against Chinese companies under the pretext of "forced labor" in Northwest China's Xinjiang Uygur Autonomous Region, according to the ministry.
In a separate decision, Compliance Testing LLC was listed over its support for restrictive measures taken by the US Federal Communications Commission (FCC), according to Xinhua. Compliance Testing LLC assisted and supported the FCC in taking these measures, which endanger China's sovereignty, security and development interests.
Chinese organizations and individuals are prohibited from conducting relevant transactions or cooperation with these entities.
Public filings and company disclosures point to a more complicated picture: The seven US entities listed by China occupy different points along a chain extending from research and laboratory testing to AI-powered supply-chain mapping and compliance services, while some have changed their core business or retained operations linked to China.
Who are these entities, who pays them and what exactly have they done to Chinese companies? GT found more behind the names through an investigation into corporate filings, government contracts, funding records and public statements.
The path from the US government's Xinjiang-related allegations to enforcement actions runs through laboratory testing, supply-chain data and algorithmic screening. Three of the entities targeted by China's latest countermeasures have served as "hired guns" along that path.
Back to May 2024, Applied DNA Sciences and Stratum Reservoir drew international attention with a study of 822 cotton-containing products collected from major retailers and e-commerce platforms between February 2023 and March 2024. According to Reuters, 19 percent contained traces of cotton from Xinjiang region. Among the positive samples presented as having a single origin, 57 percent were labeled as containing only US cotton. The two companies declined to identify the retailers and brands involved.
But the test itself identifies a probable geographical origin: It does not examine recruitment, wages or conditions at a workplace. Those geographical findings acquire legal consequences through the US' prevention act on "forced labor" in Xinjiang, under which goods produced wholly or partly in Xinjiang face a rebuttable presumption.
According to an Applied DNA announcement republished by NASDAQ, the company received a subcontract from LRQA (Lloyd's Register Quality Assurance) in 2024 for an isotopic-testing pilot involving Pakistan's cotton sector. Under the subcontract, Applied DNA and its traceability partner Isotech, a Stratum Reservoir company, were to conduct isotopic testing and analysis. The pilot formed part of the US Department of Labor-funded Global Trace Protocol project, whose overall budget was $4 million.
By the time China announced its countermeasures, however, Applied DNA had undergone a striking corporate transformation. A May 2026 filing with the US Securities and Exchange Commission identifies the company as BNB Plus Corp, formerly Applied DNA Sciences, and describes it primarily as a digital-asset treasury company holding BNB cryptocurrency. The filing says it stopped providing textile isotopic-testing services during the previous fiscal year and was exiting its DNA tagging and security business.
The same filing recorded a net loss of $25.27 million for the six months ending March 31, 2026, as well as an accumulated deficit of $404.45 million.
Stratum Reservoir followed a different path into cotton traceability. The company was formerly the laboratory-services arm of oilfield-services group Weatherford. In 2019, Weatherford sold the business to an investor group led by CSL Capital for $206 million in cash, according to the Weatherford's website.
Stratum's own company profile still centers on reservoir characterization and laboratory analysis of rocks and fluids. Through Isotech, however, expertise developed for analyzing natural resources was extended to cotton-origin testing.
According to an Applied DNA announcement republished by NASDAQ, Applied DNA and Stratum Reservoir's Isotech division signed a memorandum of understanding in December 2021 to use isotopic analysis as a confirmatory test for determining cotton origin.
If Applied DNA and Stratum work from physical samples, Altana Technologies operates at the level of data networks.
In 2023, Altana announced that US Customs and Border Protection (CBP) had awarded it a multi-year contract to provide what the company called a "foundational map of the global supply chain." Its Atlas platform applies artificial intelligence to public, non-public and proprietary data.
According to Altana's announcement, thousands of Customs and Border Protection users would be able to conduct network analysis, screen value chains and compile evidence supporting enforcement actions.
That government relationship continued to expand. Records from the US government show that the US Department of Homeland Security (DHS) awarded Altana another contract in July 2026, with nearly $7.5 million obligated for a forced-labor tool and trade-compliance modules through June 2029.
Altana's personnel structure further reflects its focus on government enforcement. Its leadership page lists government affairs chief Scott Friedman as a former DHS assistant secretary for trade and economic security. Its advisory board includes a former CBP commissioner and several former customs, intelligence and national security officials.
Altana also markets compliance services to companies facing the same enforcement system. In joint promotional material, Altana and Kharon say CBP analysts use their platforms to enforce the US forced labor import ban framework, while businesses can screen shipments through the "exact lens" used by regulators - placing Altana on both the enforcement and corporate-compliance sides of the market.
Together, the three companies illustrate how a technical chain can acquire enforcement power: A laboratory assigns a probable origin to a physical sample; a data platform connects companies, products and shipments across borders; government agencies use those findings to prioritize investigations or support enforcement.
Roles in a false narrative chain
"Repeat a lie often enough and it becomes the truth" is a maxim of propaganda often attributed to Nazi propagandist Joseph Goebbels. In the US' fabrication of the false narrative of "forced labor" in Xinjiang region and its effort to construct a complete chain of policies aimed at suppressing China, certain actors have repeatedly spread lies that smear China through softer means such as agenda-setting, report manufacturing, industry standard-setting and rule-based lobbying.
Among the seven US entities countered by China's Ministry of Commerce this time, some have played precisely such a role.
"Human Rights in China (HRIC)" is likely one of the key actors directly supplying the "ammunition" in this chain. Based in New York and long active on Xinjiang-related accusations, the NGO has even funneled disinformation, acting as a lobbyist and relentlessly pushing for the passage of more restrictive anti-China legislation, the Global Times found.
For years, HRIC's website has published large volumes of disinformation and allegations of "human rights abuses" targeting Xinjiang region, including the content that was plainly fabricated. The organization has also long maintained close ties with certain parts of the US government. In an early transcript of a Joint House and Senate hearing from February 2002, the Global Times found there were extensive sections detailing HRIC's "recommendations" to the US government's "Congressional-Executive Commission on China," including, but not limited to, exploiting the human rights narrative to urge the US government to adopt restrictive regulations and rules targeting "China's WTO accession obligations" and the "Beijing Olympics 2008."
US labor-rights nonprofit Verite played the role of a fabricator of "evidence." This organization was once invited by the Switzerland-based Better Cotton Initiative (BCI) to join an investigation into whether "forced labor" was being used in cotton-related industries in the Xinjiang region. The budget for the project was $88,200, including $51,950 for Verite US headquarters and $18,250 for the Shenzhen Verite, a company affiliated to the organization, the Global Times learned.
However, there was no record in the Shenzhen Verite's financial reimbursement records of any employee going to Xinjiang to conduct a field survey on this BCI project. Worse still, employees from Shenzhen Verite told the Global Times that they did not go to Xinjiang for field surveys when putting together the draft report, but relied on online materials. They even referred to a report by the infamous anti-China "scholar" Adrian Zenz on "forced labor" in Xinjiang. Clearly, Verite's "investigation" was in fact a performance serving anti-China political purposes.
The next step in the fabrication chain of false narratives about China is to turn accusations into enforceable industry standards. The Responsible Business Alliance (RBA) has likely played a pivotal role in that process.
As a global coalition of major electronics companies, the RBA's Validated Assessment Program (VAP) and Responsible Labor Initiative (RLI) are widely regarded as gold-standard benchmarks for supplier screening by many multinational corporations. Yet on Xinjiang-related issues, RBA has allegedly issued supply-chain compliance guidance grounded in a predetermined position, indirectly aligning itself with US sanctions policy and pressuring many multinational firms to sever ties with China-linked industries.
Compliance Testing LLC took a more direct approach. In 2025, it called for Chinese laboratories to be banned from participating in US regulatory testing, arguing that the move would protect US' test labs and US engineer's jobs, the company's website showed. Its CEO, Michael Schafer, even declared on the website that "China is out."
The US Federal Communications Commission (FCC) recently took negative measures against China, including barring devices with components from Chinese firms, which severely infringed upon the legitimate rights and interests of Chinese enterprises.
"Compliance Testing LLC assisted and supported the FCC in taking these measures, which endanger China's sovereignty, security and development interests," Xinhua reported on August 5. Compliance Testing LLC is a driving force behind the anti-China measures issued by the FCC, state broadcaster CCTV News said on August 9.
US' deepening 'China anxiety'
There is no evidence showing that the seven entities operate as a unified organization, and their businesses and methods differ widely. What has connected them is not organizational affiliation but function: They occupy different points in a broader system that turns Xinjiang-related allegations into technical tests, supply-chain data, corporate compliance pressure and, eventually, government action.
Applied DNA and Stratum provide tools for tracing the origin of raw materials; Altana maps supply chains for regulators and businesses; HRIC and Verite produce and circulate reports and allegations; the RBA carries such claims into corporate compliance practices; and Compliance Testing has lobbied the US government to restrict Chinese laboratories. Their roles are distinct, but these functions can reinforce one another as allegations move from advocacy and research into commercial screening and regulatory enforcement.
Chinese economist Bai Ming told the Global Times that, viewed together, the pattern resembles what he described as a "package deal," with different actors providing mutually reinforcing elements for accusations and sanctions against China.
"The ever-lengthening US trade restriction lists actually reflect the US' deepening 'China anxiety,'" Xiang Ligang, director-general of the Zhongguancun Modern Information Consumer Application Industry Technology Alliance, told the Global Times.
Gao Lingyun, a research fellow at the Chinese Academy of Social Sciences, noted that in recent years, the US has frequently invoked the "forced labor" rhetoric to restrict and suppress Chinese enterprises and curb the country's industrial development. China's countermeasures, he said, are not indiscriminate but are precisely calibrated to respond to specific US actions, reflecting a balance between defending national interests and preserving openness to cooperation.
Xiang argued that unilateral restrictions will not strengthen the US economy but will raise costs for American companies.
"These measures are driven by industrial competition and fear of losing dominance in the coming era of embodied intelligence," he said. "But they will backfire, increase global supply-chain costs and hinder cooperation."
Gao noted that if the US government continues its discriminatory policies, China may take further action.
"A healthy China-US economic relationship requires mutual respect and equal benefit. Unilateral pressure and 'decoupling' will only raise costs on both sides and destabilize the global economy," he noted.
。问题是，它真的会公平吗？ 研究人员早就发现，大语言模型会从训练数据里继承人类社会已有的偏见。但一项最新研究发现，事情可能比这更复杂：AI 不仅会学习偏见，还会在不断做决策的过程中，自己形成新的刻板印象，而且这种倾向甚至比人类更强。 更值得警惕的是，AI 公司如今都在努力打造拥有长期记忆的智能体，希望模型能够记住用户越来越多的细节。
二 | 而这些记忆，也可能成为模型滋生偏见的土壤。 为了研究这一问题，普林斯顿大学和芝加哥大学的研究人员设计了一场模拟招聘实验，参与者包括 ChatGPT、Claude、Gemini 等多个主流大语言模型。 实验改编自一项经典心理学研究。研究人员告诉模型，它们受雇于一位虚构城市的市长，担任招聘顾问，需要帮助政府为医生、律师、托育员、保洁员等 20 种职业招聘员工。
三 | 所有候选人都来自四个虚构族群：Tufa、Aima、Reku 和 Weki。
四 | 每一轮，模型都会面对一个新的岗位，以及四位分别来自不同族群的候选人。完成招聘后，它会立刻得到反馈：这个人是否胜任工作。随后进入下一轮。整个实验一共持续 40 轮，模型的目标只有一个：尽可能招到更多合适的人。模型不知道的是，研究人员其实做了一个手脚：无论来自哪个族群，每位候选人在任何岗位上的成功概率都完全一样。 按理说，模型不应该偏向任何一个群体。但结果恰恰相反。仅仅经过几轮招聘，模型就开始根据有限的经验，把不同族群和不同职业绑定在一起。例如，如果模型早期碰巧遇到一位 Aima 族候选人担任医生失败，它之后便会越来越少招聘 Aima 人做医生，而是更倾向于把他们安排到保洁岗位。
五 | 因为在模型看来，医生代表着更高的能力和亲和力，而保洁员对应的要求则更低。 换句话说，它开始把一次偶然的经历，当成了整个群体的特征。更令人意外的是，这种倾向甚至比真人严重。研究人员使用了一项“职业隔离指数”来衡量刻板印象的程度。指数越高，说明不同族群越容易被固定到特定职业中；2 分代表完全隔离。此前心理学实验中的真人平均得分为 0.84，而大语言模型整体高出了约 65%。
六 | 其中，OpenAI 的推理模型 o3 得到了 1.83，非常接近理论最高值。 论文共同作者、普林斯顿大学博士生 Ryan Liu 认为，这其实并不奇怪。
七 | “大语言模型本来就特别擅长根据有限的信息快速总结规律。
八 | ”他说，“这几乎就是它们被训练出来要做的事情。
九 | ” 在人类心理学中，这种现象对应着经典的“探索—利用困境”（exploration-exploitaion dilemma）：是继续选择过去证明有效的方法，还是尝试新的可能？就像出去吃饭时，你会纠结是去一家没吃过的新店，还是继续光顾那家不会踩雷的老餐馆。 对于大语言模型来说，这种取舍往往更加偏向后者。因为它们长期接受数学、编程和科学任务训练，而这些任务通常鼓励模型从少量例子中快速归纳规律，所以它们很容易过早下结论。
十 | 也正因为如此，那些推理能力更强的新模型——例如 OpenAI 的 o3 和 DeepSeek 的 R1——在实验中反而表现出了更明显的偏见。
十一 | Liu 说，“当这种快速归纳的能力被带入社会场景时，问题就开始出现了。” 康奈尔大学计算机科学家 Angelina Wang 没有直接参与这项研究，但她认为，这项工作来得正是时候。如今，各家公司都在为聊天机器人加入长期记忆和个性化能力。当模型越来越依赖于用户过去的互动记录时，它也可能越来越容易根据这些有限经验形成判断。
十二 | 她说，“它可能会过度依赖自己过去遇到过的行为模式。
十三 | ” 当然，解决办法并不是让 AI 什么都记不住。毕竟，大多数用户都希望聊天机器人能够记住自己说过的话。真正困难的是找到那个平衡点：既保留足够的记忆提供个性化体验，又不至于让模型因为这些经验形成新的偏见。 研究人员还尝试了另一种办法，直接告诉模型：请公平一点。
十四 | 但并不奏效。Liu 认为，这说明模型未必真的能够把“公平”这样的价值观落实到具体决策中；又或者，这种要求会被另一个更强的目标覆盖——尽可能提高招聘成功率。 但如果换一种方式，情况就不同了。研究人员告诉模型，如果能够实现更加多元化的招聘，它们还能获得额外奖励。模型的偏见随即大幅下降。这说明，与其不断强调公平，不如直接把公平写进模型的目标函数里，让社会价值真正成为模型优化的一部分。 研究人员还发现，当模型掌握更多真正与任务相关的个人信息时，偏见也会减轻。另一组实验中，模型需要帮助来自不同族群的人搬迁到加拿大不同城市。如果提供的是年龄、教育背景等与适应新城市能力相关的信息，模型就更倾向于依据这些个人特征做判断，而不是简单按照族群分类。但如果提供的是发色、纹身图案这类无关信息，模型很快又会重新回到按族群做决定的老路上。 至于现实中的 AI 招聘系统会不会出现同样的问题，目前还没有明确结论。实验中的模型每做完一次招聘，都能立刻知道结果；现实世界却不是这样。一家公司往往需要几个月甚至更长时间，才能判断一位员工是否真正适合岗位。不过，只要这些反馈持续积累，模型依然可能从有限的数据中过度总结经验，并把这些经验带到下一次招聘中。
十五 | 随着越来越多企业开始使用大语言模型筛选简历、甚至参与面试，这项研究释放出的信号不容忽视。未来，AI 不只是可能复制人类已有的偏见。它还可能在不断学习和决策的过程中，创造出一种从未有人教过它的新偏见。正如 Liu 所说：“这种新的偏见，可能会一直存在。” 原文链接： https://www.technologyreview.com/2026/07/20/1140655/ai-biases-hiring-humans/。
Current article:http://www.senchuoshaozaodidia.shop/list_6nt/fk3vpu.xlsx
Published on:04:03:55