AI 招聘工具与算法歧视 雅思精读

英文原文

来源：Will AI give you the job? Automated hiring tools spark discrimination and secrecy lawsuits；The Guardian，Arielle Pardes，2026 年 8 月 19 日。

[1] For the last four years, Erin Kistler has applied for thousands of jobs at companies like Paypal, Microsoft and Netflix, only to find her résumé disappear into a black hole. A product manager with nearly 20 years of experience, Kistler believes she was qualified for every role, yet she never received a single interview.

[2] Now, Kistler is suing Eightfold AI, the Silicon Valley maker of hiring software used by hundreds of companies, including those where she applied, in a class-action lawsuit. The case, filed in January in California court, is one of the first to argue that automated screening functions as an undisclosed consumer report or applicant dossier, ranking job applicants on their likelihood of success without giving them the chance to see or challenge the results, according to Kistler's legal team.

[3] The lawsuit is one of several new legal battles over companies' use of AI in employment decisions. Workers are also suing Meta over an internal AI system that allegedly targeted them for layoffs because they took parental or medical leave. And a recent lawsuit against IBM alleges that AI tools discriminated against older workers.

[4] Companies in the US are increasingly turning to AI to make faster and more efficient workplace decisions, often under the guise of objectivity. But experts who study this type of software say it can also introduce or exacerbate bias that could follow candidates across companies. Lawmakers are taking notice. The outcomes of the legal battles could shape future employment practices, transparency around them and the power candidates have in the process.

[5] "There's actually no law requiring a notice or disclosure of the use of these AI hiring systems," said Ifeoma Ajunwa, a professor at Emory University School of Law and founding director of the AI and future of work program. "So companies are not necessarily telling workers when they're being evaluated with AI."

AI use in hiring is growing

[6] Last year, 90% of employers used some form of automation in the hiring process, according to a report from the World Economic Forum. These tools range from basic filtering – like excluding candidates who don't have a four-year degree or another prerequisite – to using AI to conduct assessments of a job applicant's skills or even to perform the initial phone interview.

[7] In the case of Eightfold, the company bills itself as "the world's largest, self-refreshing source of talent data". It constantly updates an internal database with information from the résumés, LinkedIn profiles, and social media profiles of over a billion workers who have applied for jobs through its platform. Using this data, it uses AI to score applicants on a scale of 0 to 5, predicting how well they would perform in a given job. Ultimately, the score could prioritize some candidates over others for interviews.

[8] "A large part of the problem is that job applicants don't know … what the reports say," said Rachel Dempsey, an attorney representing Kistler.

[9] Dempsey argues that job applicants deserve the same transparency into what goes into the hiring process that Americans already get in their credit reports. With that information, they can either dispute inaccuracies or work to improve on factors affecting their candidacy. But without any insight, it's impossible to know if these algorithms discriminate against job candidates in explicit ways, Dempsey said.

[10] "The concept of a black box is very scary," she says.

[11] In a statement to the Guardian, Eightfold AI denied the allegations. "We believe the claims asserted are without merit and intend to defend ourselves vigorously," an Eightfold AI spokesperson wrote in an email.

[12] In response to the lawsuit filed against IBM, the company denied using AI to automatically screen out candidates and in a statement said it "does not condone or tolerate discrimination of any kind." Meta did not respond to a request for comment on its lawsuit.

The fear of algorithmic blacklisting

[13] Though employment software providers claim to free up hiring managers to spend more time with qualified candidates and remove human personal biases, experts disagree. They say AI can exacerbate biases like favoring workers of certain demographics.

[14] "As we've done more of the research on AI hiring systems, we actually see that they tend to replicate a lot of the same biases that human managers have," Ajunwa said.

[15] Nearly all AI systems are trained to look for patterns, which can reinforce stereotypes or legacy biases. Amazon, for example, once used an AI tool that downranked women's résumés because the company's top performers were men. Amazon discontinued its use after discovering the issue. In other cases, AI hiring systems can introduce their own biases. In research for her book, The Quantified Worker, Ajunwa found that AI scored job applicants with southern accents poorly, as it struggled to understand them during voice interviews.

[16] While human hiring managers can be unfair – choosing the applicant who was in the same fraternity, for example – "the magnitude of bias is even stronger [with AI]," says Xuechunzi Bai, assistant professor at the University of Chicago who has researched hiring bias.

[17] Bai co-authored a recent study where AI models were asked to make hiring decisions about fictional applicants assigned to made-up demographic groups called Tufa, Aima, Reku and Weki. After a few rounds of hiring decisions, the AI models made inferences about candidates that had nothing to do with their qualifications, in effect stereotyping them. If one Tufa was a good doctor, then Tufa candidates were more likely to be chosen as doctors, and Weki candidates were more likely to be hired as janitors.

[18] The study found greater bias in AI than in human hiring decisions she examined last year. It also found that newer and more advanced models produced hiring decisions with more bias.

[19] "We were quite shocked to see this result," Bai said. AI models are very good at solving problems where there is one objective solution, as in math or coding. "But … they're getting worse in terms of exploring the other alternatives," Bai said. "In the context of hiring, exploration is quite important, but these models are not trained to do it."

[20] The consequences of AI's biases for an applicant can extend beyond one job opening. When a worker is rejected by a person, they can try again at a different company with a different hiring manager. But when a worker is rejected by an algorithmic system used across companies, the system "remembers the decision that was already made and makes it again for the sake of efficiency", said Ajunwa. "In reality, you've been algorithmically blackballed."

[21] Because so many companies use the hiring software built on the same foundational AI models, the concern is that a negative flag could potentially make it much harder for a candidate to get a job at all, said Katie Creel, co-author of a recent study examining the risks of "algorithmic monoculture" in hiring systems. "People are going to be shut out of jobs more than they would have otherwise been," she said.

[22] Eightfold AI did not comment on the matter.

Can AI be responsibly integrated into hiring?

[23] While AI tools may save time or help test candidates' skills, hiring managers shouldn't allow AI to make decisions, said Iman Abuzeid, CEO of Incredible Health, an AI hiring platform used by more than 1,500 healthcare employers. Incredible Health offers an AI agent to perform phone interviews, but the AI doesn't autoreject, score, rank or decide who moves forward, according to Abuzeid.

[24] "It's about collecting information from the candidates so a human is better armed to make the decision," Abuzeid said.

[25] Abuzeid also said about 10% of Incredible Health's AI interviews are audited for bias by a human, a process that's becoming a legal requirement by a growing number of cities and states.

[26] In New York City, a law that took effect in 2023 requires employers using automated systems in hiring to conduct annual bias audits and notify candidates in advance about the tech's usage. The law, however, only applies to software that "substantially assists" or replaces decision-making, creating a loophole for cases where humans are part of the process. Recent laws from Illinois and Colorado also prohibit employers' use of AI tools that result in unlawful discrimination.

[27] Increasingly, city and state regulations aim to make notifying job applicants that AI is being used a baseline requirement. But ideally, candidates would have transparency into what their algorithmic dossiers say about them, said Jenny Yang, a partner at Outten & Golden, the law firm representing the plaintiffs in the lawsuit against Eightfold AI.

[28] "This initial transparency is helpful in better identifying where there may be problems," Yang said, referring to notifications that AI was used. "There's a lot of growing concern among workers that they may be denied opportunity for reasons that no one understands."

文章脉络

报道从一位求职者的经历切入：四年投递数千个岗位、一次面试都没有。由此展开三条线索——多起诉Meta、IBM、Eightfold 的官司，算法偏见的实验证据，以及各地刚开始建立的审计与告知规则。

文中出现过两类容易被混为一谈的百分比与刻度，对照如下：

阅读提醒：报道没有断言这些系统「确实」歧视了具体个人——除已认定的事实外，多数说法来自原告律师、研究者与公司回应三方，注意区分。

地道表达

共 12 条，均可在英文原文中字面检索到；例句为原创，附中文释义。

长难句拆解

多重插入语包裹的指控内容

原句：[2] The case, filed in January in California court, is one of the first to argue that automated screening functions as an undisclosed consumer report or applicant dossier, ranking job applicants on their likelihood of success without giving them the chance to see or challenge the results, according to Kistler's legal team.

主干：主句 The case ... is one of the first to argue that ...；argue 后接宾语从句 automated screening functions as ...。

层次与易错点：filed in January in California court 是插入的过去分词，修饰 case；宾语从句后接 ranking ... 与 without giving ... 两个说明动作；句末 according to ... 标明这一指控的出处。易错点：functions as 里的 functions 是动词（起……的作用），不是名词「功能」；读到句末的 according to 要意识到整段指控属于原告律师的说法，而非记者认定。

逻辑关系：定义 + 限定来源——报道先给出指控内容，再声明它来自谁。

参考译文：这起 1 月在加州法院提起的案件，是最早主张「自动化筛选实际上等同于一份未披露的消费者报告或申请人档案」的诉讼之一：它按成功可能性给求职者排序，却不给他们查看或质疑结果的机会——这一说法来自基斯特勒的律师团队。

写作迁移：The report, published last month, argues that online platforms function as unregulated marketplaces, ranking sellers by popularity without letting them appeal. 上月发布的这份报告认为，网络平台实际上等同于不受监管的市场，按人气给卖家排序却不给他们申诉的机会。

关系从句叠加与跨句指代

原句：[4] But experts who study this type of software say it can also introduce or exacerbate bias that could follow candidates across companies.

主干：主句 experts ... say ...；宾语从句 it can also introduce or exacerbate bias；bias 后接关系从句 that could follow candidates across companies。

层次与易错点：who study this type of software 修饰 experts；宾语从句里的 it 指前面提到的 AI；follow candidates across companies 意为「跟着候选人跨越不同公司」。易错点：把 that 从句理解成修饰 software 会读成「软件会跟随候选人」；across companies 是跨公司，不是「在公司内部」。

逻辑关系：转折（前面说效率与客观）+ 递进后果。

参考译文：但研究这类软件的专家表示，它也可能引入或加剧偏见，而这种偏见会跟着求职者跨越不同公司。

写作迁移：Researchers who track social platforms say they can amplify content that spreads far beyond its original audience. 追踪社交平台的研究者表示，平台会放大那些远远超出原始受众范围传播的内容。

引语嵌入与同位语

原句：[16] While human hiring managers can be unfair – choosing the applicant who was in the same fraternity, for example – "the magnitude of bias is even stronger [with AI]," says Xuechunzi Bai, assistant professor at the University of Chicago who has researched hiring bias.

主干：主句是一个引语结构 "the magnitude of bias is even stronger," says Xuechunzi Bai ...；前面 While 从句作让步背景。

层次与易错点：破折号之间是举例说明「不公平」；引语内的对比对象由前文提供，即人类招聘经理的偏见；says Xuechunzi Bai 后接同位语说明身份。易错点：英文引语把说话人放在后面（says X），中文习惯先说人；方括号 [with AI] 是记者补入的说明，不是受访者原话。

逻辑关系：让步 + 转折强调——人类也会偏见，但 AI 的幅度更大。

参考译文：人类招聘经理也可能不公平——比如选了和自己同一个兄弟会的人——但「（在 AI 那里）偏见的程度甚至更强」，研究招聘偏见的芝加哥大学助理教授白雪春子说。

写作迁移：While paper records can be lost – a fire destroyed the archive in 1998, for instance – "digital loss is far more permanent," says the archivist. 纸质记录也可能丢失——例如 1998 年一场火烧毁了档案——但「数字形式的丢失要持久得多」，这位档案管理员说。

分词结果状语揭示的制度漏洞

原句：[26] The law, however, only applies to software that "substantially assists" or replaces decision-making, creating a loophole for cases where humans are part of the process.

主干：主句 The law ... only applies to software；software 后接关系从句说明限制条件。

层次与易错点：creating a loophole ... 是现在分词作结果状语，逻辑主语是前面整句所述的法律适用范围；where humans are part of the process 修饰 cases。易错点：把 creating 当成修饰 software 会读成「软件制造漏洞」；substantially assists 加了引号，说明这是法条原文用语，含义比 assist 更强。

逻辑关系：结果——正因为法条只覆盖「实质协助或取代」决策的软件，人在流程中时便出现漏洞。

参考译文：然而，这项法律只适用于「实质性地协助」或取代决策的软件，这就为「人在流程中参与」的情形留下了一个漏洞。

写作迁移：The rule covers only full-time staff, leaving a gap for part-time workers who do the same job. 该规定只覆盖全职员工，这给从事同样工作的兼职者留下了缺口。

术语与行业黑话

• bias audit｜本文含义：对招聘系统是否产生歧视性结果的定期审查｜使用背景：纽约等地的法规要求每年进行，是最早落地的算法问责形式｜近词辨析：与 financial audit 不同，bias audit 关注群体间结果差异，不涉及账目｜可否迁移到雅思：可（科技监管类话题）。
• black box｜本文含义：输入与输出可见、内部逻辑不透明的系统｜使用背景：机器学习与治理讨论的常用比喻，强调不可解释性｜近词辨析：black box 指不透明，algorithm 指算法本身，二者不等同｜可否迁移到雅思：可（需解释）。
• algorithmic blackballing｜本文含义：被算法系统标记后反复淘汰｜使用背景：由 blackball（投票否决某人入会）引申，报道用来形容跨公司的连锁排除｜近词辨析：与 rejected（单次被拒）不同，强调系统性且持续｜可否迁移到雅思：不可，属报道用语。
• algorithmic monoculture｜本文含义：大量机构依赖同一批基础模型，导致风险同质化｜使用背景：借自农业的 monoculture（单一作物），指系统缺乏多样性｜近词辨析：与 standardization 不同，monoculture 强调共同失效的风险｜可否迁移到雅思：可（科技与社会风险）。
• applicant dossier｜本文含义：关于申请人的成套资料档案｜使用背景：原多用于情报与人事，报道借指算法生成的候选人画像｜近词辨析：dossier 强调汇总与分类，不等于 résumé（简历本身）｜可否迁移到雅思：不可。
• class action｜本文含义：由多名处境相同的原告共同提起的诉讼｜使用背景：美国法律程序，便于个体难以单独承担的案件合并审理｜近词辨析：与 test case（试探性个案）不同，class action 追求群体赔偿｜可否迁移到雅思：可（法律类话题）。

背景补充

• 事实：去年 90% 的雇主在招聘中使用某种形式的自动化（[6]），从筛选学历到 AI 电话初面都有。这个比例是理解「为什么一场官司会影响很多人」的前提。
• 事实：纽约市 2023 年生效的法律要求年度偏见审计与提前告知；伊利诺伊与科罗拉多也有类似规定（[26]）。各地规则适用范围不同，正是报道所说「漏洞」的来源。
• 事实：亚马逊曾因训练数据反映男性高绩效者而降低女性简历排序，发现问题后停用（[15]）。这是「偏见来自历史数据」的具体例证。
• 观点（研究者）：Ajunwa 与 Bai 都认为 AI 会复制甚至放大人类偏见，且模型越新偏见越强（[14][18]）。这是研究结论，不是已被监管机构确认的事实。

雅思迁移

写作

题目：Some people believe artificial intelligence should not be used to make hiring decisions. To what extent do you agree or disagree?

提纲：

• 立场：部分同意——AI 可以辅助筛选，不宜单独决策。
• 主体一：AI 能节省时间、统一初筛标准，对大量招聘的机构有效率价值。
• 主体二：但训练数据带有历史偏见，且链条不透明，候选人难以申诉。
• 结论：应当把 AI 定位为信息收集工具，决策与问责留在人身上，并配套审计与告知。

示范段（84 词）：

Using AI to shortlist candidates can save time, especially for employers who receive thousands of applications. The problem arises when algorithms decide rather than assist. Because these systems learn from past hiring, they can reproduce old patterns of exclusion, and applicants may never know why they were rejected. A reasonable compromise is to let AI gather and organise information while humans make the final judgement, supported by regular audits and clear notice to candidates about how their data is used.

阅读

1. 基斯特勒的诉讼核心主张是什么？ 定位：[2] 依据："automated screening functions as an undisclosed consumer report or applicant dossier" 答案：她主张自动化筛选实际上是一份未披露的申请人档案，在未告知、也不给查看或质疑机会的情况下给人评分排序。

2. 为什么报道说现行法律存在空白？ 定位：[5] 依据："There's actually no law requiring a notice or disclosure" 答案：法律并未要求企业告知求职者正在被 AI 评估，因此企业不一定会主动说明。

3. 亚马逊的案例说明了什么？ 定位：[15] 依据："Amazon discontinued its use after discovering the issue" 答案：说明 AI 会复制历史数据中的偏见——因为高绩效者多为男性，系统就压低女性简历；发现后亚马逊停用了该工具。

4. 什么叫「算法黑名单」？ 定位：[20] 依据："remembers the decision that was already made and makes it again for the sake of efficiency" 答案：一旦被跨公司使用的算法系统标记，系统会延续此前的负面判断，使求职者在多家公司反复被淘汰。

5. 报道给出哪些可行的改进方向？ 定位：[23]–[26] 依据："hiring managers shouldn't allow AI to make decisions" 答案：AI 只收集信息、由人做决定；定期人工复核与年度偏见审计；提前告知候选人并允许其查看算法档案。

口语延伸

Part 2

Describe a technology that worries you. You should say:

• what it is
• how it is used
• who it affects and explain why it worries you.

2 分钟提纲（57 词）：I would talk about AI tools used to screen job applications. Companies use them to rank candidates quickly, often without saying so. They affect anyone applying for work, especially people whose backgrounds look unusual to the system. It worries me because a rejection can follow you across employers and you never learn the reason.

Part 3

1. Should companies be required to tell candidates that AI is used?（出处：[5]） 思路：从知情权与可申诉性两点答。 Yes, because people cannot challenge a decision they do not know about. A short notice costs very little and makes the whole process easier to review.

2. Do you think AI can ever be completely fair in hiring?（出处：[15]–[18]） 思路：承认数据来源决定结果，指出「完全公平」难以验证。 Probably not completely, since these systems learn from past decisions that were already unequal. Fairness also depends on what we mean by it — equal treatment and equal outcomes are different goals.

3. How might regulation change how employers use these tools?（出处：[26]） 思路：讲审计成本、告知义务与「人在流程中」的漏洞。 Regulation mainly adds paperwork and audits, which larger firms can absorb more easily. It also pushes employers to keep a human in the loop, although that can become a box-ticking exercise if no one checks the result.