After more than two decades in senior fashion jobs, Stacey Duguid wanted what she describes as “one last job.” Instead, the 52-year-old spent 16 months submitting what she calls gazillions of CVs, receiving little more than automated replies. On the advice of career coaches, she began shaving the first 10 years of experience off her resume, a practice she describes as having had to “Botox” her CV.
“AI holds a mirror up to society. It’s a reflection of our bias,” Duguid told the BBC in August 2026. “I’m not saying this is just a women’s issue; I’ve had messages from men too. My point is about gendered ageism.”
Duguid’s experience points at a system most job seekers never see. Software, not people, screens the vast majority of applications at large employers, and a wave of new research shows these systems can disadvantage experienced, mid-career candidates in ways their designers never intended. Regulators have begun to respond, but slowly.
Software sees nearly every resume first
The scale of automated screening is no longer in dispute. Jobscan’s sixth annual ATS usage report, published in August 2026, found that a detectable applicant tracking system screened applications at 97.4% of Fortune 500 companies, 487 of the 500. Workday alone ran hiring for more than 40% of them, and Workday and SAP SuccessFactors together accounted for 52.4%.
An applicant tracking system (ATS) does not always make decisions with artificial intelligence, but the software’s first job is the same either way: to filter a stack of applications down to a shortlist. A recruitment platform is seeing your resume long before a recruiter is, and the evidence on what that means for candidates keeps growing.
Photo: Pexels Harvard Business School‘s Hidden Workers research, led by Joseph Fuller and Manjari Raman with Accenture, documented how these filters treat qualified people.
Harvard Business School‘s Hidden Workers research, led by Joseph Fuller and Manjari Raman with Accenture, documented how these filters treat qualified people. More than 90% of employers surveyed initially filter or rank candidates for middle-skills roles (94%) and high-skills roles (92%) using criteria such as skills, credentials, and years of experience. As the report put it, a recruiter may never see an application that would in fact satisfy every one of the employer’s requirements. Caregivers, veterans, and people with employment gaps are among the groups the filters systematically hide, which is precisely the profile of many mid-career applicants.
Future Skills & WorkThe Career Black Box: How AI Talent Marketplaces Are Collapsing Career Ladders
AI’s automation of 67% of sales tasks and 50% of market research has dismantled the on-the-job apprenticeships that historically built institutional expertise.
Read More →AI can invent its own bias
Until recently, most of the debate assumed machines inherited discrimination from historical data, and that cleaning the data would clean the model. Two studies published in 2026 undermine that assumption.
A team from Princeton University and the University of Chicago ran 15 large language models, including models from OpenAI, Anthropic, Google, Meta, and DeepSeek, through repeated rounds of a simulated hiring exercise with four fictional demographic groups designed to be equally likely to succeed. Even with no real differences to learn from, the models began steering different groups into different occupations based on early outcomes. The mechanism is an explore-exploit trade-off: the models stopped experimenting and locked onto whatever appeared to work in the first few rounds. The study, posted in November 2025 and accepted as an ICML 2026 oral, states bluntly that “LLMs can spontaneously develop novel social biases about artificial demographic groups even when no inherent differences exist,” and that such systems are “not merely passive mirrors of human social biases, but can actively create new ones from experience.”
Interventions mostly failed. Step-by-step reasoning prompts, added randomness, and shortened decision histories did little, according to SHRM’s October 2026 report on the research. What worked was explicitly incentivizing exploration and stating a diversity objective. Michelle Travis, a research professor at the University of San Francisco School of Law, told SHRM the findings “should still be a wake-up call for employers that incorporate AI tools into their hiring process.”
The second study looked at the real world. Stanford researchers analyzing a single hiring vendor’s data covered 3.4 million applicants who submitted 4 million applications to 156 employers between December 2018 and December 2022. In an analysis framed by Title VII employment-discrimination standards, 25.87% of applications from Black applicants and 14.74% from Asian applicants went to positions that showed adverse impact, meaning selection rates below four-fifths of the most-selected group’s rate. Because employers shared 42 models across the 156 firms, a rejection by one employer’s algorithm raised the odds of rejection by the next employer running the same tool, a dynamic the researchers call an algorithmic monoculture.
Why mid-career candidates should pay attention
For readers of every age, the monoculture finding matters as much as the bias finding. If a platform ranks you out because of a gap or an unconventional first job, the same vendor’s model can rank you out again at the next employer. In a monoculture, early rejections can compound across an entire market instead of washing out.
Why mid-career candidates should pay attention For readers of every age, the monoculture finding matters as much as the bias finding.
The age and gender evidence is sharpest for mid-career women. A study published in Nature, led by Stanford Graduate School of Business professor Douglas Guilbeault, prompted ChatGPT to generate more than 34,500 resumes across 54 occupations using typical male and female names. The resumes it produced for women depicted them as younger and less experienced, and when the model then graded its own output, it gave older men the highest ratings even when the underlying information was identical.
Future Skills & WorkThe Skill Inflation Paradox 2026: Why Mid-Career Professionals Are Overqualified for Their Roles—and How Companies Are Weaponizing Dynamic Skill Stacking
The report highlights that organizations are paying premiums for these capabilities, but it does not specify exact percentages (e.g.
Read More →
The BBC, reporting in the same week as the Nature-linked coverage, interviewed more than 60 women aged 40 to 65 who described months of automated rejections despite senior backgrounds. Koeyli Jaluka, a 49-year-old executive with a background in manufacturing, healthcare, and corporate roles who said she had been headhunted for a decade, applied to 442 jobs after a redundancy and came back empty. “I’ve never been in this position before,” she told the BBC. “Now I’m hunting for jobs and being relentlessly rejected.”
Concern about the pattern is not limited to the job seekers. The City of London Women Pivoting to Digital Taskforce found that 68% of more than 1,000 women surveyed said their employer had given them no opportunity to retrain or move into digital roles. Its chair, Caroline Haines, told the BBC that “there is a high level of concern that AI at initial recruitment level… doesn’t account for the skills of a vast body of women that could be very effective and therefore promote economic growth.”
Anonymized screening may even hide the problem. Laura Holden, an AI lawyer and founder of the consultancy Bonsai AI, told the BBC that some CV-screening tools grade resumes A to D and flag items such as a seven-year career gap as “a thing to note,” nudging recruiters before a human has weighed the application. She pointed to what she described as a lack of regulation globally that requires providers to show how their tools work.
The rules are starting, slowly
The counterweight is arriving in pieces. On January 1, 2026, House Bill 3773 took effect in Illinois, amending the Illinois Human Rights Act to prohibit AI tools that subject employees or applicants to discrimination on the basis of protected classes across recruitment, hiring, promotion, discipline, and the terms of employment. It also requires employers to tell workers and applicants when AI is being used in those decisions.
She pointed to what she described as a lack of regulation globally that requires providers to show how their tools work.
In the European Union, recruitment AI sits in the AI Act’s high-risk category, but companies got a reprieve: the EU’s Digital Omnibus package, given final approval by the Council in June 2026 and signed in early July, delays key high-risk obligations that had been due to take effect in August 2026, pushing them toward 2027. Compliance advocates read that as lost time. Industry groups argued the technical standards were not ready.
Future Skills & WorkAI-Generated Resumes Are Distorting Hiring Markets
A 2026 University of Maryland study found GPT-4o favors its own resumes 82% of the time in business roles.
Read More →What comes next is disclosure by default. If Illinois-style notices spread, every applicant will at least know when a machine read their resume first. Until then, the honest summary for candidates is uncomfortable: the filter that may have declined you is invisible, increasingly capable of generating its own preferences, and shared across so many employers that one early biased outcome can echo through a whole job market. The question the Princeton and Chicago researchers left open, how an employer audits a tool that develops new preferences over time, has no settled answer yet.
