The Bias in the Machine

Denise Warren is fifty-nine years old, and she cannot get an interview.

Over six months, she applied for 247 software engineering positions. She has thirty years of experience, regularly updated technical skills, strong references, and a track record of delivering complex projects on time. She received three callbacks. The rest disappeared into a silence so complete she began to wonder if her applications were being received at all.

They were received. They were also filtered out before any human saw them. The applicant tracking systems used by most companies she applied to include AI-powered screening that scores candidates based on patterns in historical hiring data. Those patterns reflect decades of bias against older workers. Denise’s graduation year, her years of experience, even her email domain (she still uses AOL) triggered flags in algorithms designed to identify candidates who “fit the profile” of past successful hires. The profile skews young.

Denise does not know this. She cannot know this. The algorithms that rejected her operate invisibly, at scale, without explanation or appeal. She only knows that her phone does not ring.

How Algorithmic Bias Works

#

Machine learning systems learn patterns from training data. If that data reflects historical discrimination, the algorithm learns to discriminate. This is not malfunction. It is the system working exactly as designed, reproducing the biases encoded in the information it was trained on.

Age discrimination in hiring is not new. What is new is the scale, speed, and invisibility with which AI enables it. A human recruiter reviewing a hundred resumes might bring bias to the task, but they also bring judgment, context, and the possibility of recognizing an unusual candidate. An algorithm screening ten thousand applications in minutes applies its biases uniformly, invisibly, and without deviation. Every candidate over fifty who triggers the wrong signals is filtered out before human judgment can intervene.

The technical mechanisms are straightforward. Algorithms use proxies for age even when they do not ask for age directly. Graduation year correlates almost perfectly with age. Years of experience exceeds what younger candidates could have. Technology skills listed may include older platforms that signal career stage. Employment gaps that correlate with caregiving, health events, or economic disruptions disproportionately affect older workers. The algorithm does not need to know your birthdate to effectively discriminate based on age.

The feedback loop compounds the problem. If AI-assisted hiring favors younger candidates, the resulting workforce skews younger. That workforce generates the performance data that trains the next iteration of the algorithm. The bias reinforces itself, each cycle producing training data that makes the pattern more pronounced.

Where Age Bias Appears

#

The applications extend far beyond hiring, though hiring remains the most consequential domain.

In resume screening, AI systems score candidates based on features that predict “culture fit” or “success probability.” A 2025 Stanford study found that AI resume-screening tools gave older male candidates higher ratings than both female candidates and young candidates when the resumes contained identical qualifications. The algorithms had learned that older men “look like” successful employees because the historical data reflected workplaces that favored them. The same study found systematic bias against older women across multiple AI platforms, reflecting and amplifying existing patterns of intersectional discrimination.

In job advertising, AI determines who sees which opportunities. Facebook settled a lawsuit over allegations that its advertising platform allowed employers to exclude older users from seeing job postings. The mechanism was not explicitly age-based; it used targeting parameters that correlated with age. The effect was the same: older workers never knew the jobs existed.

In healthcare, risk models that allocate resources may incorporate age in ways that disadvantage older patients beyond what clinical evidence supports. An algorithm determining who receives a transplant, who gets intensive care, or who qualifies for certain treatments may weight age in ways that are neither medically justified nor transparent to patients or providers. The Optum algorithm exposed in 2019 showed racial bias in healthcare allocation; similar audits for age bias remain rare.

In insurance and credit, age factors into pricing and decisions both explicitly and through proxy variables. Length of credit history, income stability, employment tenure: these factors correlate with age and shape lending decisions, insurance premiums, and access to financial products. Older entrepreneurs face documented lending bias, compounding the challenges covered in Series 6.

In fraud detection, algorithms may flag transactions by older adults as suspicious based on patterns that assume older people should not be making certain purchases or financial moves. Account freezes and false accusations of fraud disproportionately affect older users who deviate from algorithmic expectations of how they should behave.

The Regulatory Gap

#

Existing civil rights law provides uneven protection against algorithmic age discrimination.

The Age Discrimination in Employment Act protects workers forty and older, but its application to AI-assisted decisions remains uncertain. The EEOC issued guidance in 2023 clarifying that employers are liable for discriminatory outcomes from AI hiring tools, regardless of whether they intended to discriminate. Enforcement has been minimal. The guidance creates theoretical liability without practical deterrence.

New York City’s Local Law 144, which took effect in 2023, requires employers to conduct annual bias audits of automated employment decision tools and publicly report the results. It is the first law of its kind. Illinois requires notification when AI analyzes video interviews. Colorado’s AI Act, effective in 2026, will require developers and users of AI hiring tools to use reasonable care to prevent algorithmic discrimination.

These laws represent progress. They also represent a tiny fraction of the regulatory framework that would be needed to address algorithmic bias comprehensively. Most jurisdictions require nothing. Most audits are not conducted. Most discrimination operates invisibly and with impunity.

The class action lawsuit Mobley v. Workday, proceeding through federal court as of early 2026, alleges that Workday’s AI hiring tools discriminate based on race, age, and disability. A federal judge granted preliminary certification allowing the case to proceed as a collective action, ruling that claims based on disparate impact from algorithmic systems are suitable for collective treatment. The case may establish important precedents for how AI vendors can be held accountable for biased outputs.

The Research Gap

#

Academic and policy attention to AI fairness has focused heavily on race and gender bias. Age bias receives less attention despite comparable harm.

The Amazon hiring tool scrapped in 2018 for showing bias against women became a landmark case in algorithmic discrimination. Similar analysis for age bias is rarely conducted or published. Healthcare allocation algorithms have been audited for racial bias; audits for age bias are almost nonexistent. The research infrastructure that identifies and quantifies bias in AI has not prioritized age as a category of concern.

This gap is not accidental. Age discrimination is normalized in ways that race and gender discrimination are not. The assumption that older workers are less valuable, less adaptable, less worth hiring is widely shared, even by people who would never voice similar assumptions about other protected categories. When bias is assumed to be reasonable, it does not attract the scrutiny that exposes its unreasonableness.

The growing attention to algorithmic fairness creates an opportunity. The frameworks developed to identify and mitigate race and gender bias in AI can be applied to age. The auditing methodologies, the fairness metrics, the regulatory approaches: all can be extended if there is political will and research attention. Whether that extension occurs depends on whether older adults and their advocates demand inclusion in conversations that have largely proceeded without them.

What Would Change This

#

Technical solutions exist but require implementation. Training data can be examined for age distribution and adjusted to include adequate representation of older workers. Algorithms can be audited for disparate impact on older applicants using the same methodologies applied to other protected categories. Proxy variables that correlate with age can be identified and removed or weighted differently.

Regulatory solutions are emerging. Mandatory bias audits that include age as a protected category, transparency requirements that disclose when AI is used in consequential decisions, and accountability frameworks that hold both AI developers and deployers responsible for discriminatory outcomes would all reduce harm.

Design principles matter. Age-inclusive AI development would involve older adults in design, testing, and evaluation. Voice assistants trained primarily on younger voices fail to understand older users; image recognition systems trained on younger faces misidentify older ones. These technical failures compound the discrimination embedded in decision-making algorithms.

Advocacy is beginning to engage these issues. Organizations like AARP and Justice in Aging are incorporating algorithmic fairness into their policy agendas. The intersection of technology policy and aging policy is newly visible, though the work of building coalitions and changing law remains in early stages.

What the Machine Reflects

#

The algorithm does not hate Denise Warren for being fifty-nine years old. It simply learned from a world that does.

And now it scales that learning to millions of decisions per day. Every application filtered, every ad targeted, every risk score calculated reflects patterns absorbed from data that encoded decades of discrimination. The machine does not create bias; it amplifies and accelerates it, removing the friction of individual judgment that might occasionally override the pattern.

The bias in the machine is not a glitch to be fixed. It is a reflection of the society that built it. The data is biased because the world is biased. The algorithms discriminate because the humans who generated the training data discriminated. Fixing the machine requires fixing the data it learns from, the people who design it, and the oversight that governs it.

None of that is happening fast enough. Denise Warren and millions of older workers like her continue to disappear into algorithmic silence, rejected by systems they cannot see, cannot challenge, and cannot understand. The machine has made discrimination frictionless, invisible, and effectively unchallengeable.

That is not inevitable. It is a choice. The same technology that scales bias could be designed to detect and correct it. The question is whether the will exists to make that choice. For now, the machine reflects the values embedded in its training data. Those values include the assumption that older workers are worth less. Until that assumption changes, the algorithms will keep acting on it.