Employment
See also: Employment ChatGPT Plugins
The use of artificial intelligence in employment covers software and statistical systems that companies deploy to recruit, screen, interview, hire, monitor, evaluate, promote, and dismiss workers. The category includes resume parsing engines, applicant tracking systems, video interview scoring tools, conversational chatbots that handle high volume hiring, productivity surveillance platforms, and generative AI assistants used by recruiters and managers. Employment is one of the most heavily regulated AI domains because hiring and firing decisions are protected by anti-discrimination law in nearly every developed economy, and because the same systems that promise to reduce bias have produced several of the most prominent examples of algorithmic discrimination in the past decade.
Overview
A majority of large employers now use some form of machine learning in the hiring pipeline, usually before any human recruiter ever sees an application. Resume parsers convert PDF and Word documents into structured fields. Match engines rank candidates against a job description. Knockout questionnaires filter out applicants who fail screening criteria. Chatbots handle scheduling and basic Q&A. Video platforms record interview answers and produce scores on traits the vendor claims to measure. Once a worker is hired, productivity software tracks keystrokes, mouse activity, application usage, and meeting behavior, and some platforms produce a single composite score that managers can compare across teams.
The market is dominated by a small number of human capital management vendors. Workday, Oracle, SAP SuccessFactors, and ADP supply enterprise recruiting and HR modules that ship with embedded AI features. Standalone applicant tracking vendors such as Greenhouse, iCIMS, SmartRecruiters, Lever, and Jobvite layer AI scoring on top of the core workflow. Specialized players such as HireVue, Paradox, Eightfold AI, Beamery, and Phenom focus on video interviewing, chatbots, or talent intelligence. Resume parsing is largely commoditized through Textkernel (which owns Sovren),[31] Daxtra, and HireAbility.
Regulators caught up to the market later than the technology. The first dedicated US law on AI hiring took effect in Illinois on January 1, 2020.[4] New York City began enforcing a bias audit requirement on July 5, 2023.[5] The US Equal Employment Opportunity Commission settled its first AI hiring case against iTutorGroup in August 2023,[7] and the federal class action Mobley v. Workday survived a motion to dismiss in July 2024,[10] then received conditional collective certification in May 2025.[11] The European Union AI Act classifies most employment uses as high risk, with the main set of obligations for those uses originally scheduled to apply from August 2, 2026 and since deferred to December 2, 2027.[16][46]
The industry sells AI in hiring as a remedy for bias and a path to better hires. The peer reviewed evidence has been mixed. Several audits and academic studies have found that mainstream systems amplify rather than reduce disparate impact, especially when training data reflects historical hiring patterns or when generative models are asked to rank resumes that differ only by candidate name.[18][19]
History
Automated resume sorting predates the modern wave of AI. Early applicant tracking systems in the 1990s ran keyword searches against scanned resumes. Sovren launched in 1996 and became one of the most widely licensed resume parsing engines, embedded inside thousands of HR products that customers never saw branded. Taleo, founded in 1999 and acquired by Oracle in 2012, established the modern ATS workflow of structured pipelines, requisitions, and stage based candidate movement.
The shift from keyword matching to statistical learning began in the mid 2010s. Companies started running supervised models on past hiring outcomes to predict which applicants were most likely to be hired or to perform well. Amazon built one of the earliest examples between 2014 and 2017. Reuters reported in October 2018 that the project had been scrapped because the model had taught itself to penalize resumes that contained the word "women's" and that downgraded graduates of two all women's colleges.[1] The training data, mostly resumes received over the prior decade, was dominated by men, and the model learned to reproduce that pattern.[1]
Video interview AI emerged from a parallel track. HireVue, founded in 2004 as a video interviewing service, added algorithmic scoring around 2014 and began marketing claims that the platform could assess personality traits, cognitive ability, and "emotional intelligence" from facial expressions, voice patterns, and word choice.[21] By 2019 the Electronic Privacy Information Center had filed a complaint with the US Federal Trade Commission alleging that HireVue's scoring was opaque, unproven, and likely biased.[2] In January 2021 HireVue released the results of an external audit of its algorithms, having already announced in 2020 that it had stopped using candidates' facial expressions in its scoring.[3]
The arrival of large language models changed the field again. ChatGPT made it easy for both job applicants and recruiters to draft resumes, cover letters, screening rubrics, and interview questions in seconds. Vendors retrofitted generative models into their products: HireVue, Paradox, Workday, LinkedIn, and Eightfold all shipped major AI agent features in 2024. By 2025, several large employers were openly using AI as a justification for hiring freezes, while several others quietly walked back earlier claims that AI had eliminated entire job categories.
AI applicant tracking systems
Applicant tracking systems handle the requisition, application, screening, and interview scheduling workflow. The market is fragmented, with most large employers running an enterprise HCM platform alongside one or more specialized tools. The table below covers vendors that have publicly described AI features and their approximate launch or AI release dates.
| Vendor | Founded | Owner or status | AI features |
|---|---|---|---|
| Workday Recruiting | 2005 | Public (NASDAQ: WDAY) | Candidate skills cloud, generative AI job description and screening assistants, requisition matching, and embedded recommendations across the HCM suite |
| Taleo | 1999 | Acquired by Oracle 2012 | Oracle Recruiting Cloud now integrates Eightfold agentic interview intelligence |
| iCIMS | 2000 | Vista Equity Partners | Talent Cloud AI, Frontline AI for high volume hiring (2026), generative job description writer, candidate matching |
| Greenhouse | 2012 | TPG (majority stake, 2021) | Greenhouse AI Recruiting (2024), automated sourcing assistant, structured interview kits |
| SmartRecruiters | 2010 | SAP (acquired September 2025)[51] | Winston AI assistant (2024), automated job posting distribution, candidate ranking |
| Lever | 2012 | Employ Inc. | Nurture campaigns, automated requisition matching |
| Jobvite | 2006 | Employ Inc. | Resume matching, predictive analytics, Brian AI assistant |
| Eightfold AI | 2016 | Private, more than $410 million raised through 2021, $2.1 billion valuation[32][47] | Talent intelligence platform built on a 1.6 billion profile dataset, AI Interviewer, Recruiter Agent, Career Hub |
| Beamery | 2014 | Private, over $1 billion valuation after a $50 million Series D (December 2022)[33] | Talent graph, Ray workforce planning agent, skills based matching |
| Phenom | 2010 | Private | Career site personalization, talent CRM with predictive scoring |
| LinkedIn Recruiter | 2008 | Microsoft | Hiring Assistant AI agent (October 29, 2024), candidate sourcing, InMail drafting[27] |
| Paradox | 2016 | Workday (acquired October 2025)[50] | Olivia conversational assistant for high volume hiring, used by McDonald's, Chipotle, and General Motors[28] |
| ZipRecruiter | 2010 | Public (NYSE: ZIP) | Phil AI assistant for candidates, Smart Matching for employers |
| Indeed | 2004 | Recruit Holdings | Smart Sourcing AI candidate matching, automated outreach |
| HireVue | 2004 | Carlyle Group | Video interview scoring, structured interview templates, technical coding assessments. Acquired Modern Hire May 9, 2023[29] |
| Modern Hire | 2019 (merger of Shaker International and Montage) | Acquired by HireVue May 2023[29] | Pre-hire science assessments, virtual job tryouts |
| Pymetrics | 2013 | Acquired by Harver August 2022[30] | Neuroscience based behavioral games scored against company specific performance benchmarks |
| Harver | 2015 | Private | Situational judgement tests, cognitive assessments, integrated Pymetrics games |
| HireAbility | 2001 | Private | Resume parsing and structured data extraction |
| Textkernel | 2001 | Acquired Sovren 2021,[31] owned by Main Capital | Resume parsing in 29 languages, LLM Parser (2023) using GPT-3.5 |
| Sovren | 1996 | Acquired by Textkernel 2021[31] | Resume parsing and semantic matching components licensed to ATS vendors |
| MyInterview | 2015 | Private | Asynchronous video interviews with AI scoring |
| Vervoe | 2016 | Private | Skills based job auditions graded by AI |
| Plum | 2012 | Private | Behavioral assessments with predictive matching |
Workday is the most consequential vendor for class action exposure because the same platform is licensed by thousands of employers and produces a consistent candidate recommendation across all of them. That common architecture is the central allegation in Mobley v. Workday, discussed below.[9]
Video interview AI
Video interview platforms record candidates answering pre-set questions and produce numerical scores intended to predict job performance. The largest vendor is HireVue, which said when it acquired Modern Hire in May 2023 that the combined company would serve more than 1,150 customers, including more than half of the Fortune 100.[29] Other players include MyInterview, Vervoe, and the assessment products that Modern Hire brought into HireVue through the May 2023 acquisition.[29]
The scoring methods have changed under public pressure. HireVue's original product analyzed facial expressions in addition to speech and word choice. The Electronic Privacy Information Center filed a complaint with the Federal Trade Commission on November 6, 2019, alleging that HireVue's facial analysis was "biased, unprovable, and not replicable."[2] HireVue commissioned an external audit by O'Neil Risk Consulting and Algorithmic Auditing, and released the audit results in January 2021; by then HireVue had already announced, in 2020, that it had stopped using facial expressions as a factor in its algorithms.[3] The current product scores on speech content, intonation, and behavioral signals.
Hilke Schellmann, an investigative journalist and journalism professor at New York University, ran a widely cited test of MyInterview in 2021.[20] Schellmann answered interview questions in German while reading a Wikipedia entry about psychometrics. The platform produced a transcription that was nonsensical English and still rated her a 73 percent match for an English language role.[20] The vendor told her the score was driven by intonation, not the substance of her answers.[20] The episode has become a standard example of why audio based scoring is hard to validate.
MyInterview, HireVue, and several smaller players continue to be used for high volume entry level hiring, often for retail, hospitality, and call center positions where employers receive hundreds of applications per opening. Civil rights groups have repeatedly raised concerns that asynchronous video interviews disadvantage candidates with disabilities, candidates who do not speak the assessment language as a first language, and candidates without reliable broadband access.
AI chatbots and applicant matching
Conversational AI tools handle the volume problem at the top of the funnel. The largest example is Olivia, the conversational assistant built by Paradox, a Scottsdale Arizona vendor founded in 2016 by Aaron Matos.[28] Olivia screens candidates through text and SMS, schedules interviews against recruiter calendars, sends reminders, and answers basic questions about pay and shift schedules.[28]
A 2021 Aptitude Research case study reported that conversational AI had cut McDonald's hiring process from 10-14 days to one to two days.[28] Chipotle, which launched a Paradox-built assistant called Ava in 2024, said it expected the tool to reduce time to hire by up to 75 percent.[49] Other Paradox customers have included General Motors and 7-Eleven. Paradox does not publish pricing. Workday completed its acquisition of Paradox on October 1, 2025.[50]
LinkedIn announced its Hiring Assistant at the Talent Connect conference in Phoenix on October 29, 2024.[27] The product is described as LinkedIn's first AI agent, built to read a job description, suggest candidates from the platform's roughly 1 billion member profile database, draft outreach messages, and ask screening questions.[27] Initial customers included AMS, Canva, Siemens, and Zurich Insurance.[27] LinkedIn published metrics in 2025 claiming that Hiring Assistant users review 81 percent fewer profiles to find a qualified match and save about 1.5 hours per role.
ZipRecruiter built a candidate facing assistant called Phil that recommends roles based on conversational profiling. Indeed's Smart Sourcing tool surfaces candidates against employer job posts using a 24 percent positive response rate benchmark that the company has published.
The specialized talent intelligence vendors compete on the size and freshness of their candidate graphs. Eightfold AI, founded in 2016 by former Google engineer Ashutosh Garg, raised $220 million in a June 2021 Series E round led by SoftBank Vision Fund 2 at a $2.1 billion valuation, bringing its total funding to more than $410 million.[32][47] The company markets a 1.6 billion profile dataset.[32] Beamery, a London based competitor, raised $50 million in December 2022 at a $1 billion valuation.[33] Phenom raised more than $100 million across multiple rounds and serves enterprise customers including General Motors and AbbVie.
Workforce monitoring and productivity surveillance
The other side of the AI in employment story is software that monitors workers after they are hired. The category expanded dramatically during the 2020 pandemic as employers tried to manage remote workforces. Most of the major vendors offer some combination of keystroke logging, screenshot capture at random intervals, application and URL tracking, idle time detection, and behavioral analytics built on top of those signals.
| Vendor | Founded | Primary features |
|---|---|---|
| Hubstaff | 2012 | Time tracking, screenshots, app and URL tracking, GPS for mobile teams |
| Time Doctor | 2012 | Time tracking, screenshots, idle detection, payroll integration |
| Teramind | 2014 | Keystroke logging, screen recording, behavioral analytics, insider threat detection |
| ActivTrak | 2009 | Passive monitoring, focus and burnout analytics, no keystroke logging by default |
| Veriato | 1998 | Insider risk monitoring, screen capture, user behavior analytics |
| Sneek | 2017 | Always-on webcam tiles for distributed teams |
| InterGuard | 2002 | Cloud and on-premises monitoring with policy enforcement |
| RemoteDesk | 2017 | AI based identity verification and continuous webcam monitoring for remote agents |
| Crossover WorkSmart | 2010 | Random screenshots and webcam captures every 10 minutes |
| Prodoscore | 2016 | Composite productivity score across email, calendar, and CRM activity |
The most prominent recent controversy was around Microsoft's Productivity Score, launched in October 2020.[15] The original release allowed administrators to see per-user metrics across 73 indicators including how often individual workers turned on their cameras in meetings, sent emails with @ mentions, used Word, Excel, or Teams, and contributed to shared documents.[15] Privacy researcher Wolfie Christl publicized the feature on November 24, 2020, and the criticism intensified through that week.[44] Microsoft announced on December 1, 2020 that it would remove user names from the dashboards and aggregate communication, meeting, content collaboration, teamwork, and mobility metrics only at the organization level.[14][15]
The broader workforce surveillance market continued to grow regardless. A 2021 ExpressVPN survey of 2,000 US employers found that 78 percent used monitoring software to track employee performance or online activity. The Center for Democracy and Technology has repeatedly raised concerns about the discriminatory effects of always-on monitoring on workers with disabilities and caregivers.
Bias and notable incidents
Several high profile audits and journalism investigations have established a pattern. AI tools that are sold as bias mitigators frequently amplify the disparities in their training data, and generative models that are not specifically tuned for hiring produce ranked outputs that vary with candidate name or demographic.
Amazon 2018
The most famous example is the Amazon resume screening tool that the company scrapped in 2017. Reuters journalist Jeffrey Dastin first reported the story on October 10, 2018.[1] The system, built by an Amazon ML team in Edinburgh from 2014 onward, scored candidates from one to five stars for software engineering and other technical roles.[1] By 2015 the team realized the model was not gender neutral. The system penalized resumes containing the word "women's," downgraded graduates of two all women's colleges, and favored candidates who used verbs that appeared more frequently on male engineer resumes such as "executed" and "captured."[1] Amazon disbanded the team by the start of 2017. According to people familiar with the project, recruiters looked at the tool's recommendations when searching for new hires but never relied solely on them; Amazon said the tool "was never used by Amazon recruiters to evaluate candidates."[1] The story is now a standard example in machine learning fairness coursework.
HireVue facial analysis
The Electronic Privacy Information Center complaint of November 6, 2019 argued that HireVue's facial analysis produced results that were "biased, unprovable, and not replicable" and pointed out that facial recognition systems have well documented racial accuracy gaps.[2] HireVue commissioned an audit from O'Neil Risk Consulting and Algorithmic Auditing and released the results in January 2021, after it had already stopped using facial analysis (a change announced in 2020).[3] The decision did not affect HireVue's continued use of voice and speech based scoring.
Bloomberg ChatGPT resume study
Bloomberg's data investigations team published a study on March 8, 2024 titled "OpenAI's GPT Is a Recruiter's Dream Tool. Tests Show There's Racial Bias."[18] The reporters derived 800 demographically distinct names, built from the most common first names and most distinct last names associated with white, Black, Hispanic, and Asian Americans in North Carolina voter records and US census data, and assigned them to eight otherwise near-identical resumes.[18] They asked GPT-3.5 and GPT-4 to rank the resumes 1,000 times for each of four real job descriptions from Fortune 500 companies.[18] Both models produced demographic disparities large enough to fail standard EEOC adverse impact thresholds across nearly every job category tested.[18] The full dataset and code were published on GitHub.[18]
University of Washington study
Researchers at the University of Washington Information School, led by Kyra Wilson and Aylin Caliskan, published "Gender, Race, and Intersectional Bias in Resume Screening via Language Model Retrieval" at the AAAI ACM Conference on AI, Ethics, and Society in October 2024.[19] The team ran more than three million comparisons across three open-weight text embedding models derived from LLMs: E5-mistral-7b-instruct, SFR-Embedding-Mistral from Salesforce, and GritLM-7B from Contextual AI, all fine-tuned from Mistral AI's Mistral-7B.[19] The models favored resumes with white associated names in 85.1 percent of cases. The paper originally reported that female associated names were favored in only 11.1 percent of cases, but an August 2026 erratum states that a code bug inverted the gender-only results, so that figure describes male associated names; the race and intersectional results were unaffected and were replicated by researchers at the University of Salzburg.[19]
Other documented incidents
A 2021 Harvard Business School and Accenture study, "Hidden Workers: Untapped Talent," estimated that the US had more than 27 million "hidden workers" and found that automated screening systems often filtered out candidates with non-traditional career paths, such as gaps in employment. The Center for Democracy and Technology, the Lawyers' Committee for Civil Rights Under Law, and Upturn have published several detailed reports cataloguing similar concerns.
Lawsuits and audits
Litigation around AI hiring tools moved slowly through US courts for several years before picking up sharply in 2023 and 2024. Two cases dominate the current docket.
Mobley v. Workday
Derek L. Mobley, an African American man over forty with anxiety and depression that he characterizes as disabilities under the ADA Amendments Act, filed suit against Workday on February 21, 2023 in the US District Court for the Northern District of California, case number 3:23-cv-00770.[9] Mobley alleges that he applied for more than 100 positions through employers that use Workday's recruiting platform between 2017 and 2022 and was rejected from every one, and that the rejections were driven by Workday's algorithmic candidate scoring rather than the underlying employers' independent judgment.[9] He brought claims under Title VII of the Civil Rights Act of 1964, the Civil Rights Act of 1866, the Age Discrimination in Employment Act of 1967, and the ADA Amendments Act of 2008.[9]
The case was reassigned to District Judge Rita F. Lin on November 27, 2023.[13] Judge Lin granted Workday's first motion to dismiss with leave to amend on January 19, 2024.[13] Mobley filed an amended complaint on February 20, 2024 adding allegations that Workday acted as an employment agency, an indirect employer, and an agent, plus a state law aiding and abetting claim under California's Fair Employment and Housing Act.[13]
The EEOC filed an amicus brief on April 9, 2024 supporting Mobley's agent theory.[12] On July 12, 2024, Judge Lin denied Workday's motion to dismiss the disparate impact claims, holding that an AI vendor whose tools "perform a traditional hiring function" can be liable as an agent of the employer under federal civil rights law.[10]
On May 16, 2025, Judge Lin granted preliminary collective certification under the ADEA.[11] Workday represented in filings that "1.1 billion applications were rejected" using its software tools during the relevant period, and the court noted Workday's estimate that the collective could number in the "hundreds of millions" of people.[11][53] In July 2025 Judge Lin expanded the collective to include applicants scored with Workday's HiredScore AI features.[54] Court-authorized notice went out in early 2026 with an opt-in deadline of March 7, 2026, and as of March 2026 there had been no ruling on liability.[55]
EEOC v. iTutorGroup
The EEOC's first AI hiring case settled in 2023. The agency sued iTutorGroup, a Shanghai based remote English tutoring company, in the US District Court for the Eastern District of New York on May 5, 2022, case number 1:22-cv-02565.[13] The complaint alleged that iTutorGroup's online application system was programmed to automatically reject female applicants aged 55 or older and male applicants aged 60 or older for tutoring roles, in violation of the Age Discrimination in Employment Act.[7][13]
The parties filed notice of settlement on August 9, 2023, and the consent decree was entered on September 8, 2023.[7] iTutorGroup agreed to pay $365,000 to more than 200 applicants who had been screened out.[7] The decree also required new anti-discrimination policies, training, and EEOC monitoring for five years.[7] Although the underlying technology was a simple date-of-birth filter rather than a machine learning model, the case is widely cited as the first federal enforcement action against an algorithmic hiring tool.
Other actions
The Federal Trade Commission has not brought a public enforcement action against an AI hiring vendor as of 2025, although EPIC filed a complaint against HireVue with the agency in 2019.[2] The New York City Department of Consumer and Worker Protection received only two complaints under Local Law 144 between July 2023 and June 2025 and closed both without a penalty, according to a December 2025 state audit.[34] State attorneys general have signaled interest, and the California Civil Rights Department issued employer guidance in 2024 on automated decision systems.
Regulation
The regulatory landscape is a patchwork. The most consequential rules are city level (New York), state level (Illinois, California, Colorado), federal guidance (EEOC), and supranational (EU AI Act). No US federal statute specifically governs AI in hiring as of 2025.
New York City Local Law 144
The New York City Council passed Local Law 144 of 2021 on November 10, 2021, and it became law on December 11, 2021 without the mayor's signature.[5] The law governs "automated employment decision tools" or AEDTs, defined as computational processes derived from machine learning, statistical modeling, data analytics, or AI that issue simplified outputs and that substantially assist or replace discretionary human decision making for employment decisions.[5] The law took technical effect on January 1, 2023, and the Department of Consumer and Worker Protection began enforcement on July 5, 2023 after publishing final implementing rules in April 2023.[5][6]
The core requirements are that any employer or employment agency using an AEDT for a position located in New York City must commission an independent bias audit no more than one year before use, publish a summary of the audit results on the employer's website, and provide candidates with at least ten business days' notice that an AEDT will be used.[5][6] The audit must report selection or scoring rates and impact ratios broken out by sex, race or ethnicity, and intersectional categories.[6] The DCWP can impose civil penalties between $500 and $1,500 per day per violation.[5] DCWP does not approve or list auditors; selection is the employer's responsibility.[6]
A 2024 paper by Cornell researcher Lucas Wright and colleagues, in which 155 student investigators examined 391 employers, found that 18 had posted bias audit reports and 13 had posted transparency notices, though the authors cautioned that because the law lets employers decide whether their tools are in scope, a missing audit cannot be read as non-compliance.[35] A December 2, 2025 audit by the Office of the New York State Comptroller found DCWP's enforcement ineffective: DCWP surveyed 32 companies and identified one issue of non-compliance, while state auditors reviewing the same companies identified at least 17 instances of potential non-compliance.[34]
Illinois Artificial Intelligence Video Interview Act
Illinois enacted Public Act 101-0260 in August 2019, and the law took effect on January 1, 2020.[4] The Artificial Intelligence Video Interview Act (820 ILCS 42) applies to any employer that uses AI to analyze video interviews of applicants for Illinois based positions.[4] Employers must give applicants notice that AI may be used, explain how the AI works and what general types of characteristics it evaluates, obtain consent, restrict sharing of the video to those whose expertise is necessary to evaluate the candidate, and delete the video within 30 days of an applicant's request.[4]
A 2021 amendment added a reporting requirement: employers that rely solely on AI analysis to decide which applicants receive in-person interviews must collect race and ethnicity data on the candidates who were and were not selected, and report the data annually to the Illinois Department of Commerce and Economic Opportunity.[4] A separate Illinois law amending the Illinois Human Rights Act, which took effect January 1, 2026, broadens AI hiring obligations across all employers in the state.
Colorado AI Act
Colorado Governor Jared Polis signed Senate Bill 24-205, the Colorado Artificial Intelligence Act, on May 17, 2024.[17] The law was the first comprehensive US state statute regulating high risk AI systems, and it explicitly covers automated decisions about employment, education, financial services, government services, healthcare, housing, insurance, and legal services.[17] The original effective date was February 1, 2026, later pushed to June 30, 2026 by Senate Bill 4 signed on August 28, 2025. A federal magistrate judge stayed enforcement in April 2026, and Governor Polis signed Senate Bill 26-189 on May 14, 2026, which repeals and reenacts the law as a narrower, notice-based framework for automated decision-making technology that takes effect January 1, 2027.[45]
The original CAIA imposed duties on both developers and deployers of high risk AI to use reasonable care to avoid algorithmic discrimination, to maintain documentation, to conduct impact assessments, and to notify affected consumers of adverse decisions.[17]
EU AI Act
The European Union's Artificial Intelligence Act, Regulation (EU) 2024/1689, was adopted on June 13, 2024 and entered into force on August 1, 2024.[16] Most provisions phase in over two to three years. The high risk obligations that apply to most employment uses were originally scheduled to begin on August 2, 2026, but Regulation (EU) 2026/1744, which entered into force on July 27, 2026, deferred the obligations for stand-alone Annex III high risk systems, including employment uses, to December 2, 2027.[16][46]
Annex III of the AI Act classifies the following employment uses as high risk: AI systems used for the recruitment or selection of candidates, including job advertising targeting, application filtering, and candidate evaluation; AI systems used to make decisions about work related contractual relationships, promotion, or termination; AI systems used to allocate tasks based on individual behavior or personal traits; and AI systems used to monitor and evaluate performance and behavior of workers.[16] Providers of those systems must carry out conformity assessments, maintain risk management documentation, and register the system in the EU database, while deployers must assign human oversight, keep the logs the system generates, and inform workers and their representatives before putting such a system into use at the workplace.[16]
Penalties scale with infringement type. Use of prohibited AI practices can trigger fines of up to 35 million euros or 7 percent of global annual turnover.[16] Violations of high risk obligations can trigger fines of up to 15 million euros or 3 percent of turnover.[16] Several EU member states began appointing national supervisory authorities in 2025.
EEOC guidance and the federal landscape
The US Equal Employment Opportunity Commission launched its Artificial Intelligence and Algorithmic Fairness Initiative under Chair Charlotte A. Burrows in October 2021. The agency issued technical assistance documents on the Americans with Disabilities Act and AI in May 2022 and on Title VII adverse impact and AI in May 2023.[8] The EEOC's Strategic Enforcement Plan for fiscal years 2024 through 2028, adopted September 21, 2023, explicitly prioritizes employer use of AI in recruitment, hiring, and other employment decisions. The agency settled iTutorGroup in August 2023 and filed an amicus brief in Mobley v. Workday in April 2024.[7][12]
Federal action stalled after the change in administration in January 2025. After Commissioner Andrea Lucas became acting chair, the EEOC removed its May 2022 and May 2023 technical assistance documents on AI from its website, and the Labor Department's Office of Federal Contract Compliance Programs removed its April 2024 AI guidance for federal contractors; the underlying anti-discrimination statutes did not change.[52] The result is that state and local rules and private litigation are the main pressure points on AI hiring vendors.
AI driven layoffs and corporate communications
The loudest public claims that AI has displaced workers have come from CEOs of large public companies. Several of those claims have later been adjusted.
IBM
IBM Chief Executive Arvind Krishna told Bloomberg on May 1, 2023 that the company would pause hiring for roles that AI could replace, particularly in back office and human resources functions.[22] He estimated that roughly 30 percent of IBM's approximately 26,000 non customer facing roles, or about 7,800 positions, could be replaced by AI and automation over five years.[22][37] The reduction was expected to come largely through attrition rather than direct layoffs.[22] IBM subsequently launched its watsonx generative AI platform in May 2023 and continued aggressive hiring in AI engineering and consulting.
Dropbox
Dropbox Chief Executive Drew Houston announced on April 27, 2023 that the company would lay off 500 employees, or 16 percent of staff.[23] Houston cited slowing growth and an explicit need to shift the workforce toward AI and early stage product development.[23] The company laid off an additional 528 employees, or 20 percent of staff, in October 2024 as part of a further reorganization tied to the company's Dash AI search product.
Klarna
Klarna, a Stockholm based buy now pay later fintech, became the most widely cited example of AI driven workforce reduction in 2024. The company published a press release on February 27, 2024 announcing that its OpenAI powered customer service assistant had handled 2.3 million conversations in its first month and was doing "the equivalent work of 700 full time agents."[24] Chief Executive Sebastian Siemiatkowski said publicly that Klarna would not hire any new employees through 2024 because of AI productivity gains. Klarna projected a $40 million profit improvement from the AI deployment.[24]
The reversal came in 2025. Siemiatkowski told Bloomberg in May 2025 that the company had "gone too far" with AI replacement, that service quality had degraded, and that Klarna was hiring human agents again on a flexible remote contractor basis.[25] The episode is now frequently cited by labor economists as evidence that AI productivity claims often outpace measured customer outcomes.
Salesforce
Salesforce Chief Executive Marc Benioff told the 20VC podcast with Harry Stebbings, in an episode released in December 2024, that Salesforce would not add any new software engineers in 2025.[26] Benioff said internal engineering productivity had risen more than 30 percent after deployment of Agentforce, the company's autonomous agent product, and that the company expected to add 1,000 to 2,000 sales staff rather than engineers to support customer adoption of Agentforce.[26] Benioff told Fortune in July 2025 that he disagreed with predictions that AI would cause a huge displacement of white collar workers, saying AI agents were not accurate enough to remove "the human in the loop" and that augmentation could bring "a lot more employment."[38] In a podcast interview published in early September 2025, he said Salesforce had reduced its support headcount "from 9,000 heads to about 5,000" as AI agents handled a growing share of customer conversations.[48]
Industry pattern and applicant inflation
The broader trend is harder to measure. LinkedIn data reported in June 2025 showed that the number of applications submitted on the platform had grown 45 percent year over year, reaching about 11,000 applications per minute. A significant share of that growth is attributed to applicants using ChatGPT and similar tools to tailor resumes and to bulk apply through automated job application services. Employers have responded by leaning more heavily on AI screening, producing an arms race in which both sides of the hiring interaction are increasingly automated. The 2025 Society for Human Resource Management benchmark survey reported a median nonexecutive cost per hire of $1,200 and a median time to fill of about a month and a half for both executive and nonexecutive positions.[36] The survey was not designed to isolate an effect of generative AI adoption on either measure.
Gartner forecast in October 2024 that through 2027 generative AI would spawn new roles in software engineering and operations, requiring 80 percent of the engineering workforce to upskill, and its analyst Philip Walsh said that "human expertise and creativity will always be essential to delivering complex, innovative software."[56]
The AI Layoff Trap model
In June 2026, Brett Hemenway Falk of the University of Pennsylvania and Gerry Tsoukalas of Boston University released version 3 of "The AI Layoff Trap," a 63-page theoretical economics preprint.[39] The paper is a conditional model, not a peer-reviewed empirical study or a forecast that AI will collapse the economy. Its authors state that the proposed empirical signature would require displacement at a scale and speed beyond what had occurred, and describe the contribution as identifying a structural vulnerability rather than diagnosing an active crisis.[39]
The baseline model has one sector, one period, and at least two symmetric firms. Each firm starts with a fixed number of tasks, and its automation rate also equals the share of workers it displaces because every automated task replaces one worker. AI and human labor are perfect substitutes and produce the same fixed output in the baseline. Wages are exogenous, owners' spending on the sector is normalized to zero, and some displaced wage income may be restored through new employment, transfers, or other sources. A quadratic integration cost makes additional tasks progressively harder to automate.[39] These assumptions isolate one mechanism: an automating firm receives its full labor-cost saving, but competitive revenue sharing makes it bear only a fraction of the demand loss created when its displaced workers lose income.
Under the paper's active interior conditions, the Nash automation rate exceeds the common rate that maximizes aggregate firm profit. The gap is ell(1 - 1/N)/k, where ell is demand lost per displaced worker, N is the number of firms, and k represents integration friction.[39] In the model, the gap widens with competition and demand loss, narrows as integration becomes harder, and vanishes for a monopolist. It also vanishes if displaced workers fully recover their income. If reemployment raises their income above its prior level, the sign reverses and the model can imply too little automation. In a more restrictive frictionless case, full automation is a dominant strategy even though collective restraint raises every firm's profit, but that Prisoner's Dilemma result requires a specific ordering of cost savings and demand loss.[39]
The paper evaluates policies only against this demand-externality margin. Retraining and targeted income replacement narrow the modeled gap. A general universal basic income is represented as a constant increase in autonomous demand, so it does not change the marginal incentive to automate. The authors explicitly say this construction is not a judgment about every basic-income design.[39] A per-task Pigouvian tax equal to the uninternalized demand loss implements the cooperative rate in the maintained interior setup. The analysis does not estimate a tax rate for an actual economy or evaluate administration, labor-market distortions, and political feasibility. It also notes that in a frictionless general equilibrium, interest-rate adjustment can redirect saving and remove the rationale for a corrective tax. The model's tax conclusion instead depends on frictions such as constrained interest rates or incomplete financial markets.[39]
The results are analytic comparative statics, not a calibrated macroeconomic simulation. The paper provides no causal estimate, dataset, or public code repository. Its figures illustrate assumed parameter values, and a numerical grid used in the free-entry extension is a sensitivity exercise rather than observed evidence.[39] Earlier task-based research by Daron Acemoglu and Pascual Restrepo distinguishes automation's displacement effect from the reinstatement effect created by new labor-intensive tasks.[40][41] Workplace evidence also shows heterogeneous outcomes. A study of 5,172 customer-support agents found that access to a generative AI assistant increased issues resolved per hour by 15 percent on average, mostly through gains among less experienced workers, but the authors said their single-firm study could not identify aggregate employment or wage effects.[42] French firm-level research associated AI adoption from 2017 to 2020 with higher employment and sales overall while finding negative employment growth for some administrative uses.[43] Neither study tests the cross-firm demand loop proposed in "The AI Layoff Trap."
Task economics, adoption and employment effects
The price of an AI response, the cost of completing a work task and the effect on employment are distinct measurements. A low token price can make experimentation cheaper without showing that an entire occupation can be automated. A task may also require source retrieval, tools, verification, exception handling and a person's approval. AI task cost includes these inputs and the cost of retries when measuring an accepted outcome.
A task-level comparison can record baseline human time, time with assistance, accepted output quality, review time, error recovery and the full cost of operating the system. For example, if a hypothetical task falls from 20 minutes to 12 minutes but requires five minutes of review, the measured saving is three minutes, before setup and training time. It is not an eight-minute saving or evidence of a corresponding headcount reduction. The value of time saved also depends on whether it can be used for additional work and whether demand changes.
Three common kinds of evidence answer different questions:
| Evidence | What it can measure | What it does not establish by itself |
|---|---|---|
| Capability or skill-exposure assessment | Whether a model may perform parts of a task under stated conditions | Actual adoption, displacement or net job creation |
| Product usage data | Activities of the observed product's users | Representative use across employers or the labor market |
| Workplace outcome study | Productivity or quality in the studied setting | Effects in every occupation or aggregate employment |
Indeed Hiring Lab's September 2025 Skill Transformation Index evaluated almost 2,900 skills, combining model assessments of problem-solving and physical requirements. It describes potential transformation rather than observed replacement, and states that realized effects depend on business adoption and integration.[44] Anthropic's September 2025 Economic Index analyzed Claude.ai conversations and first-party API traffic. Those populations describe use of a particular supplier's products; their task shares do not directly measure the share of jobs automated across the economy.[45]
The customer-support study discussed above found heterogeneous productivity effects in one firm.[42] Such evidence can guide tests of a similar workflow, but the relevant comparison still needs its own quality and review measures. Employment outcomes additionally depend on demand, organizational choices, complementary investments and the creation of new tasks. Neither a benchmark cost per token nor a product's task-usage share supplies those missing measurements.
Continue reading
- Artificial IntelligenceArtificial intelligence (AI) is the field of computer science dedicated to creating systems that perform tasks normally requiring human intelligence, such as learning, reasoning, recognizing patterns…Mentioned 6 times above
- Generative AIGenerative AI is a class of artificial intelligence systems that produces new data instances, such as text, software code, images, audio, video, molecular structures, or other representations, by learning…Mentioned 6 times above
- EU AI ActThe EU AI Act is the common name for Regulation (EU) 2024/1689, a directly applicable European Union law that sets horizontal rules for artificial intelligence.Mentioned 3 times above
- ChatGPTChatGPT is a conversational artificial intelligence chatbot developed by OpenAI and built on the company's GPT series of large language models.Mentioned 3 times above
- AI AgentsAn AI agent is a software system that selects and performs actions in an environment in pursuit of an objective.Mentioned twice above
- AI ethicsAI ethics is the field that studies the moral principles, values, and frameworks governing how artificial intelligence systems are designed, built, deployed, and used, and the obligations that developers and…Listed in See also
More to read
See also
- Artificial intelligence
- AI ethics
- AI regulation
- AI bias
- EU AI Act
- Generative AI
- Large language model
- ChatGPT
- Microsoft Copilot
- Facial recognition
- AI agent
References
- ^1 ^2 ^3 ^4 ^5 ^6Dastin, Jeffrey. "Amazon scraps secret AI recruiting tool that showed bias against women." Reuters, October 10, 2018. reuters.com/...bs-automation-insight-idUSKCN1MK08G
- ^1 ^2 ^3 ^4Electronic Privacy Information Center. "In re HireVue" complaint to the Federal Trade Commission, November 6, 2019. epic.org/...in-re-hirevue
- ^1 ^2 ^3Fortune. "HireVue stops using facial expressions to assess job candidates amid audit of its AI algorithms." January 19, 2021. fortune.com/...monitoring-amid-a-i-algorithm-audit
- ^1 ^2 ^3 ^4 ^5Illinois General Assembly. Public Act 101-0260, Artificial Intelligence Video Interview Act, 820 ILCS 42. Effective January 1, 2020. ilga.gov/...101-0260
- ^1 ^2 ^3 ^4 ^5 ^6New York City Council. Local Law 144 of 2021, Automated Employment Decision Tools. Enforcement began July 5, 2023. legistar.council.nyc.gov/LegislationDetail
- ^1 ^2 ^3 ^4New York City Department of Consumer and Worker Protection. "Final Rules Regarding Automated Employment Decision Tools." April 6, 2023. rules.cityofnewyork.us/...t-decision-tools-updated
- ^1 ^2 ^3 ^4 ^5 ^6US Equal Employment Opportunity Commission. "iTutorGroup to Pay $365,000 to Settle EEOC Discriminatory Hiring Suit." Press release, September 11, 2023. eeoc.gov/...settle-eeoc-discriminatory-hiring-suit
- ^US Equal Employment Opportunity Commission. "Assessing Adverse Impact in Software, Algorithms, and Artificial Intelligence Used in Employment Selection Procedures Under Title VII of the Civil Rights Act of 1964." May 18, 2023. Archived copy: data.aclum.org/...re-algorithms-and-artificial.pdf
- ^1 ^2 ^3 ^4US District Court for the Northern District of California. Mobley v. Workday, Inc., case number 3:23-cv-00770. Complaint filed February 21, 2023. clearinghouse.net/...137252
- ^1 ^2US District Court for the Northern District of California. Order on Motion to Dismiss, Mobley v. Workday, July 12, 2024, Judge Rita F. Lin. storage.courtlistener.com/....cand.408645.80.0.pdf
- ^1 ^2 ^3US District Court for the Northern District of California. Order Granting Preliminary Collective Certification, Mobley v. Workday, May 16, 2025. storage.courtlistener.com/...cand.408645.128.0.pdf
- ^1 ^2US Equal Employment Opportunity Commission. Amicus brief filed in Mobley v. Workday, April 9, 2024. eeoc.gov/...mobley-v-workday-inc
- ^1 ^2 ^3 ^4 ^5Civil Rights Litigation Clearinghouse. Mobley v. Workday and EEOC v. iTutorGroup case records. clearinghouse.net/...44074 and clearinghouse.net/...44258
- ^Microsoft. "Our commitment to privacy in Microsoft Productivity Score." December 1, 2020. microsoft.com/...y-in-microsoft-productivity-score
- ^1 ^2 ^3The Register. "Microsoft 365 axes per-user productivity monitoring after privacy backlash." December 1, 2020. theregister.com/...484010
- ^1 ^2 ^3 ^4 ^5 ^6 ^7European Union. Regulation (EU) 2024/1689 on Artificial Intelligence. Entered into force August 1, 2024. eur-lex.europa.eu/...oj
- ^1 ^2 ^3Colorado General Assembly. Senate Bill 24-205, Colorado Artificial Intelligence Act. Signed May 17, 2024. leg.colorado.gov/...SB24-205
- ^1 ^2 ^3 ^4 ^5 ^6Yin, Leon, Davey Alba, and Leonardo Nicoletti. "OpenAI's GPT Is a Recruiter's Dream Tool. Tests Show There's Racial Bias." Bloomberg, March 8, 2024. bloomberg.com/...-gpt-hiring-racial-discrimination
- ^1 ^2 ^3 ^4Wilson, Kyra, and Aylin Caliskan. "Gender, Race, and Intersectional Bias in Resume Screening via Language Model Retrieval." Proceedings of the AAAI/ACM Conference on AI, Ethics, and Society, October 2024. arxiv.org/...2407.20371
- ^1 ^2 ^3Schellmann, Hilke. "The Algorithm: How AI Decides Who Gets Hired, Monitored, Promoted, and Fired and Why We Need to Fight Back Now." Hachette, January 2024. hachettebookgroup.com/...9780306827341
- ^MIT Technology Review. "The AI hiring industry is under scrutiny, but it'll be hard to fix." November 7, 2019. technologyreview.com/...scrimination-ftc-epic-bias
- ^1 ^2 ^3Bloomberg. "IBM to Pause Hiring for Back-Office Jobs That AI Could Kill." May 1, 2023. bloomberg.com/...ck-office-jobs-that-ai-could-kill
- ^1 ^2TechCrunch. "Dropbox lays off 500 employees, 16 percent of staff." April 27, 2023. techcrunch.com/...slowing-growth-and-the-era-of-ai
- ^1 ^2Klarna. "Klarna AI assistant handles two-thirds of customer service chats in its first month." Press release, February 27, 2024. klarna.com/...mer-service-chats-in-its-first-month
- ^Daly, Charles. "Klarna Slows AI-Driven Job Cuts With Call for Real People." Bloomberg, May 8, 2025. bloomberg.com/...i-to-real-person-customer-service
- ^1 ^2Salesforce Ben. "Salesforce Will Hire No More Software Engineers in 2025, Says Marc Benioff." December 18, 2024. salesforceben.com/...ers-in-2025-says-marc-benioff
- ^1 ^2 ^3 ^4LinkedIn. "Introducing Hiring Assistant." Talent Connect, October 29, 2024. linkedin.com/...introducing-hiring-assistant
- ^1 ^2 ^3 ^4Aptitude Research and Paradox. "The 2021 Conversational AI Report," including the McDonald's case study. 2021. info.paradox.ai/...onversational%20AI%20Report.pdf
- ^1 ^2 ^3 ^4HireVue. "HireVue Acquires Modern Hire." Press release, May 9, 2023. hirevue.com/...modernhire
- ^Harver. "Harver Acquires pymetrics." Press release, August 11, 2022. harver.com/...harver-acquires-pymetrics
- ^1 ^2 ^3Textkernel. "Textkernel acquires US-based software company Sovren." November 30, 2021. textkernel.com/...i-powered-recruitment-technology
- ^1 ^2 ^3Eightfold AI. "Eightfold AI raises $220M, doubles valuation in six months." Series E funding announcement, June 10, 2021. eightfold.ai/...eightfold-ai-raises-220m
- ^1 ^2Beamery. "Beamery Raises $50 Million in Series D Funding." December 13, 2022. beamery.com/...ises-50-million-in-series-d-funding
- ^1 ^2Office of the New York State Comptroller. "Enforcement of Local Law 144: Automated Employment Decision Tools." December 2, 2025. osc.ny.gov/...-automated-employment-decision-tools
- ^Wright, Lucas, et al. "Null Compliance: NYC Local Law 144 and the Challenges of Algorithm Accountability." Proceedings of the ACM Conference on Fairness, Accountability, and Transparency, 2024. arxiv.org/...2406.01399
- ^Society for Human Resource Management. "2025 Recruiting Executives Benchmarking." 2025. shrm.org/...2025-recruiting-benchmarking-report.pdf
- ^The Washington Post. "IBM could replace 7,800 jobs with artificial intelligence, CEO says." May 2, 2023. washingtonpost.com/...ai-jobs-takeover-ibm
- ^Fortune. "Why Salesforce CEO Marc Benioff doesn't see a white-collar jobs apocalypse despite big gains from AI agents." July 30, 2025. fortune.com/...oductivity-unemployment-future-work
- ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8Hemenway Falk, Brett, and Gerry Tsoukalas. "The AI Layoff Trap." arXiv:2603.20617v3, June 3, 2026. arxiv.org/...2603.20617
- ^Acemoglu, Daron, and Pascual Restrepo. "The Race between Man and Machine: Implications of Technology for Growth, Factor Shares, and Employment." American Economic Review 108, no. 6, 2018. aeaweb.org/articles
- ^Acemoglu, Daron, and Pascual Restrepo. "Automation and New Tasks: How Technology Displaces and Reinstates Labor." Journal of Economic Perspectives 33, no. 2, 2019. aeaweb.org/articles
- ^1 ^2Brynjolfsson, Erik, Danielle Li, and Lindsey Raymond. "Generative AI at Work." Quarterly Journal of Economics 140, no. 2, 2025. academic.oup.com/...7990658
- ^Aghion, Philippe, et al. "How Different Uses of AI Shape Labor Demand: Evidence from France." AEA Papers and Proceedings 115, 2025. aeaweb.org/articles
- ^1 ^2Speed, Richard. "Privacy campaigner flags concerns about Microsoft's creepy Productivity Score." The Register, November 26, 2020. theregister.com/...productivity_score
- ^1 ^2McDermott Will & Schulte. "Colorado AI law in flux: Comprehensive replacement bill signed after federal court blocks predecessor's enforcement." JD Supra, May 2026. jdsupra.com/...i-law-in-flux-comprehensive-2107294
- ^1 ^2European Union. Regulation (EU) 2026/1744 amending Regulation (EU) 2024/1689 (Digital Omnibus on AI). Entered into force July 27, 2026. eur-lex.europa.eu/...TXT
- ^1 ^2TechCrunch. "AI startup Eightfold valued at $2.1B in SoftBank-led $220M funding." June 10, 2021. techcrunch.com/...-1b-in-softbank-led-220m-funding
- ^Salesforce Ben. "AI Agents Drive 4,000 Job Cuts in Salesforce Support Division." September 3, 2025. salesforceben.com/...n-salesforce-support-division
- ^Paradox. "Chipotle's fresh new way to reduce time to hire by up to 75%." November 1, 2024. paradox.ai/...y-to-reduce-time-to-hire-by-up-to-75
- ^1 ^2Workday. "Workday Completes Acquisition of Paradox." Press release, October 1, 2025. newsroom.workday.com/...tes-Acquisition-of-Paradox
- ^SAP. "SAP Completes Acquisition of SmartRecruiters." September 11, 2025. news.sap.com/...pletes-smartrecruiters-acquisition
- ^Cooley. "Gone but Not Forgotten: Federal Laws Still Apply Despite AI Guidance Disappearance Act." February 21, 2025. cooley.com/...y-despite-guidance-disappearance-act
- ^Proskauer Rose. "AI Bias Lawsuit Against Workday Reaches Next Stage as Court Grants Conditional Certification of ADEA Claim." June 11, 2025. proskauer.com/...ional-certification-of-adea-claim
- ^PPC Land. "Workday faces bigger discrimination lawsuit." August 2025. ppc.land/workday-faces-bigger-discrimination-lawsuit
- ^House, Mark. "Mobley vs Workday: an update on the case that is challenging AI assessment." RM Compare blog, March 27, 2026. compare.rm.com/...hat-is-challenging-ai-assessment
- ^Gartner. "Gartner Says Generative AI will Require 80% of Engineering Workforce to Upskill Through 2027." Press release, October 3, 2024. gartner.com/...g-workforce-to-upskill-through-2027
- ^1 ^2Hering, Annina, and Arcenis Rojas. "AI at Work Report 2025: How GenAI is Rewiring the DNA of Jobs." Indeed Hiring Lab, September 23, 2025. hiringlab.indeed.com/...s-rewiring-the-dna-of-jobs
- ^1 ^2Anthropic. "Anthropic Economic Index Report: Uneven Geographic and Enterprise AI Adoption." September 15, 2025. anthropic.com/...nomic-index-september-2025-report
Improve this article
Add missing citations, update stale details, or write a clearer explanation. Anyone can edit: changes go live when you publish, and every edit is checked by an editor.
10 revisions · v10 · 7,844 words · full history
Fact-checks are independent of edits: a reviewer re-verifies the article against its sources and stamps the date. How we verify
Research and drafting on this wiki are AI-assisted, under named human editorial standards. How AI is used here
Reviewer note: Independent fact-check of the full article against its cited sources (150 claims, 50 sources fetched); 22 material + 22 minor fixed
Cite this page: AI Wiki. "Employment." aiwiki.ai, updated 17 Aug 2026, fact-checked 4 Nov 2026. CC BY 4.0. https://aiwiki.ai/wiki/employment

