Key Takeaways
Treat every automated hiring score as a decision, not a neutral suggestion
A score can cost a job. In the book’s opening case, makeup artist Lizzie was laid off after being told she had received a zero on a HireVue assessment. When she obtained her employment records, she found that the interview had never been scored. The mistake went unnoticed during her appeal, and she lost her job anyway. The episode captures the central warning: software does not have to be intelligent or intentionally unfair to cause serious harm.
Scale changes the stakes. The book reports that nearly all Fortune 500 companies use algorithms or AI in hiring, and that software often screens applications before a person reviews them. A single faulty process can affect thousands of people, while applicants may not even know a machine influenced the outcome.
The case is a reminder that an automated score is not merely information when it controls access to income. In decision theory, the distinction between advice and authority matters: a tool used to rank candidates can quietly become the decision-maker if staff routinely follow its output. Lizzie’s experience also shows why appeals need more than a polite review. A meaningful process must let applicants see the underlying record, identify errors, and obtain a human reconsideration before a consequential decision becomes final.
Demand job-related evidence, because correlation can turn names and hobbies into filters
Patterns are not explanations. Résumé screeners often compare applicants with the profiles of employees already labeled successful. In one tool examined by an organizational psychologist, names such as Thomas and Elsie, places including Seattle and Syria, and hobbies such as football or biking appeared to predict success. These details may simply have appeared in the training résumés. They do not establish that a candidate can do the job.
Proxies can discriminate. Amazon abandoned a résumé tool that downgraded résumés containing phrases such as women’s chess club because its training data reflected a male-heavy workforce. Employment lawyer Ken Willner found potentially troubling variables in about a quarter of the tools he reviewed. Even a seemingly neutral factor, such as shoe size or postal code, can disadvantage protected groups without measuring job ability.
The crucial question is not whether a variable predicts an outcome in a dataset, but why it predicts it and whether that relationship is defensible for this job. This is a problem of construct validity: the system may measure a feature accurately while mistaking it for the ability an employer actually needs. Asking vendors to explain every feature is therefore not a technical nicety. It is a way to expose when historical patterns, social privilege, or accidental quirks are being repackaged as merit.
Efficiency filters can erase qualified applicants before a human ever sees them
The funnel can be too narrow. Software meant to reduce application volume often rejects people for missing exact keywords, lacking preferred credentials, or having a long employment gap. A Harvard Business School and Accenture survey found that 88 percent of employers agreed that qualified high-skill candidates were screened out for failing to match exact criteria. Entry-level retail job descriptions in the research averaged 31 listed skills, even though many were not essential.
Applicants feel the result. Sophie, a Black Navy veteran with graduate training in data science, submitted 146 applications before finding work. She often advanced only after contacting recruiters directly. In another finding, nearly half of surveyed tools automatically excluded applicants with employment gaps longer than six months, regardless of the reason. These filters save recruiters time partly by discarding people they might have wanted to meet.
This is a classic optimization trap: a system can improve the metric it was given, such as speed or reduced applicant volume, while worsening the real goal, finding capable employees. The labor market then mistakes a smaller shortlist for a better one. A useful safeguard is to track not only time-to-hire, but also who was excluded, whether hires succeed, and whether people with adjacent skills were ever considered. Otherwise, efficiency becomes a flattering name for lost opportunity.
Do not confuse digital traces with a reliable personality diagnosis
Online behavior is noisy evidence. Some hiring products claim to infer personality by scanning social media, sometimes without a candidate’s knowledge. One background report flagged posts that a worker had merely liked, including ordinary references to alcohol and a candle holder. The software struggled to distinguish a joke, a quotation, or a political statement from a genuine warning sign.
Predictions disagreed. In the author’s tests, the same person received sharply different personality descriptions from the same vendor’s analysis of LinkedIn and Twitter. In a study involving 94 students, personality outputs also changed depending on the platform or document format, and two vendors often disagreed. The book connects this confidence in outward clues to older, discredited attempts to infer character from skull shape or handwriting.
A large volume of behavioral data can create an illusion of intimacy: the system seems to know a person because it has collected many traces of them. But abundance does not guarantee that the data captures a stable trait, especially when platform conventions shape what people post. Personality research can be useful in limited settings, yet prediction should not be mistaken for diagnosis. The greater the consequence, the more important it is to separate a plausible story about a person from a validated measure.
Make assessments prove they predict the actual work, not merely test-taking style
A game is not a job. Candidates have been asked to solve timed puzzles, inflate virtual balloons, or record one-way video interviews as a way to reveal potential. Martin Burch, an experienced data analyst, was rejected from a Bloomberg role after a Plum assessment. He could not see how a shape puzzle predicted his ability to scrape data from websites. The assessment provider said employers could use cut scores, which can automatically screen people out.
A polished score can hide weak evidence. In tests described in the book, one video-interview tool assigned a strong job match even when applicants spoke German or Chinese and the generated English transcript was gibberish. HireVue eventually removed facial-expression and voice analysis after concluding those features added little. The book’s point is not that every assessment is useless, but that each needs evidence connecting its measurements to the duties of the specific job.
Assessment science offers a useful discipline here: validity belongs to an interpretation and use, not to a test in the abstract. A puzzle might predict performance for one role and be irrelevant for another. Even a modestly predictive measure can be valuable when combined with other evidence, but it should not become an automatic gate simply because it produces precise-looking numbers. Clear explanations, job-specific validation, and more than one source of evidence help distinguish a useful signal from a theatrical one.
Build accessibility and accommodation into the first screen, not after rejection
One test can exclude many minds. People with disabilities make up an estimated 10 to 25 percent of the US population, yet hiring tools are often designed around a narrow idea of normal performance. A game that rewards fast keyboard presses may measure dexterity as much as reasoning. Henry Claypool, who has a spinal cord injury, found the timed games stressful and difficult to connect with the work being assessed.
Access barriers accumulate. Deaf and hard-of-hearing applicants working with a vocational counselor sometimes could complete only three or four questions before timed assessments ended. Video interviews could be unusable for people who do not speak, while accommodation requests went unanswered. Disability law requires reasonable accommodations, but applicants may fear disclosure or be screened out before they can explain what they can do with support.
Accessibility is not an optional layer added after a tool is built. It changes what a system can validly claim to measure. A task that depends on speed, vision, hearing, or motor control may be relevant to a particular job, but that relationship must be established rather than assumed. Individual accommodation also resists the false comfort of group averages: a person’s capability cannot always be inferred from how closely they resemble a training sample. Fair design starts by making room for difference.
Use workforce analytics to open career paths, not label people as destiny
Matching can reveal overlooked skills. Some companies use internal talent platforms to identify employees who could move into new roles or learn missing skills. Vodafone used Eightfold to match staff with internal opportunities and recommend learning. A separate marketplace at Unilever reportedly matched employees with projects and created 300,000 hours of internal project work. Skill adjacencies can help people move across industries, such as flight attendants applying customer-service and logistics skills elsewhere.
Predictions still need context. Vendors also offer tools that estimate who may leave, burn out, or advance quickly. One company executive described an employee whose records were wrong because her ethnicity had been misclassified. At Cushman & Wakefield, an internal model could predict resignations accurately at a broad group level, but individual accuracy fell below 50 percent. A manager’s high turnover could also reflect a healthy transfer policy, not a toxic team.
The strongest use of workforce analytics may be expanding options rather than ranking people. A skills map can make hidden routes visible, especially for workers whose experience does not fit conventional credential patterns. That benefit depends on separating recommendation from destiny. Once a manager sees a flight-risk label or a slow-career flag, the label can influence training, pay, and promotion, creating the very outcome it predicts. Systems should support employee agency, correct bad data, and make their inferences contestable.
Measure results, not keystrokes: surveillance can raise stress while lowering performance
Activity is not achievement. Medical coder Emily Smith was monitored through mouse activity, idle time, websites visited, and daily productivity rankings. She moved her mouse while reading charts so the system would not mark her idle, and timed bathroom breaks to protect her numbers. The monitoring made her anxious and demeaning, while her later job, with a human manager and more trust, felt far better.
The evidence challenges the sales pitch. A review of 76 studies involving more than 21,000 participants found that monitoring increased stress and negative attitudes without improving performance. More intense monitoring was associated with worse performance. Microsoft reported that workers spent roughly an hour a day on productivity theater, such as sending messages or creating activity signals to appear busy. Those behaviors generate data, but not necessarily useful work.
Surveillance changes the behavior it claims merely to observe. This is a version of Goodhart’s law: when a measure becomes a target, people optimize the measure, not necessarily the underlying purpose. In work that requires judgment or concentration, visible activity can be especially misleading. Monitoring may be defensible for a narrow safety or security need, but the rationale should be specific, proportionate, and disclosed. A system that rewards mouse motion may be measuring compliance with the system rather than contribution to the organization.
Keep health and emotional data out of employment decisions unless protections are real
Wellness can become surveillance. Workplace programs may combine fitness tracking, insurance claims, benefits records, and digital behavior. One vendor demonstrated how dropping a spouse from a health plan, taking loans from a retirement account, or adding a baby could trigger personalized messages about mental health, finances, or care. The book notes that the vendor described a dataset containing more than 33 billion data points on 100 million Americans.
Sensitive inferences travel. Fitness data could reveal health conditions, while voice-analysis products claim to detect signs of anxiety or depression. These tools may be offered as support, but the same information could affect hiring, promotion, or insurance decisions. The author found no double-blind, placebo-controlled study establishing the vocal biomarker products discussed, and no such tool had received FDA approval.
The ethical question is not simply whether a wellness intervention might help. It is whether people can refuse it without penalty, who can access the data, and whether the information can be repurposed. In privacy engineering, purpose limitation means collecting data for a clearly defined use and preventing drift into unrelated decisions. That principle matters especially for health signals, where a nudge toward care can sit beside an incentive to reduce costs. Support is more credible when it does not require surrendering control.
Require independent validation and a human appeal before algorithms shape livelihoods
Prediction has hard limits. In a study using thousands of variables on more than 4,000 families, 160 research teams tried to forecast outcomes such as grades, eviction, layoffs, and material hardship. The best machine-learning predictions were weak: the top score was .23 for material hardship, .19 for grades, and .07 for several other outcomes. Models using only four conventional variables performed about as well as many complex systems.
Accountability must match the stakes. Workers such as Amazon Flex drivers Claire Grove and Vicky Graham received termination notices they believed were triggered by faulty app records, and neither could get a clear explanation. The book calls for transparent decision records, independent testing, ongoing audits, and a practical way to challenge decisions. A vendor’s claim of fairness or accuracy is not a substitute for evidence that the specific tool works for the specific role.
Prediction is hardest when the target itself is unstable. Job success depends not only on an individual, but on managers, teams, health, family circumstances, and changing organizational needs. A complex model can obscure this uncertainty while giving decisions an undeserved air of inevitability. Independent trials can test whether a tool adds value beyond simpler methods or chance, while appeal rights address the cases no model can anticipate. The goal is not to ban computation, but to prevent uncertainty from being transferred entirely to the person being judged.
Analysis
The book’s strongest contribution is its shift in the question from whether AI is biased to what these systems are actually doing, and whether that activity is fit for a high-stakes decision. The tools examined are not one unified technology. They include keyword filters, ranking systems, personality inferences, games, speech analysis, predictive dashboards, and automated management rules. Each has different failure modes. Treating them all as “AI” makes scrutiny harder, while vendors’ smooth scores can make weak evidence look comparable across products.
A useful way to evaluate any such tool is to separate four tests. Does it measure the quality it claims to measure? Does that quality relate to the work? Does the tool produce comparable results across groups and circumstances? Can affected people understand and challenge the outcome? Passing one test does not guarantee the others. A system can predict an outcome while relying on a proxy that is unfair, or distribute outcomes evenly while selecting people poorly.
The recurring institutional failure is not simply bad code. It is incentive design. Companies reward hiring speed, reduced labor costs, and measurable activity, so tools optimized for those goals can succeed commercially while failing workers and even employers. When a score is used to reject people, the organization may never learn whether the rejected applicants would have excelled. That creates a feedback loop: the system validates the workforce it helped produce, then treats that workforce as evidence of what success looks like.
The book’s caution should not be mistaken for nostalgia about human judgment. Human hiring has long reproduced discrimination and inconsistency. The better comparison is between accountable and unaccountable decision processes. Automation can help when its purpose is narrow, its evidence is independently tested, and humans retain both the authority and obligation to question it. Transparency, accessibility, and appeal are not obstacles to innovation. They are conditions for claiming that a system deserves trust.
Review Summary
The Algorithm receives mostly positive reviews for its thorough exploration of AI's impact on hiring and workplace practices. Readers praise Schellmann's research and accessible writing style, highlighting the book's revelations about bias in AI tools and their potential harm to job seekers. Critics note the book's narrow focus and repetitive themes. Many readers find the content eye-opening and recommend it to those interested in AI's role in employment. Some suggest the book could benefit from more solutions and a broader scope.
People Also Read
Glossary
adverse impact
Unequal selection outcomesA hiring or employment practice has adverse impact when it selects people from a protected group at a substantially lower rate than others. The practice may not have been designed to discriminate, but employers may still need to show that it is related to the job and necessary for the business. The term is central to evaluating automated screens.
four-fifths rule
A selection-rate comparisonA commonly used US guideline for screening selection rates. A group’s selection rate is compared with the rate of the group selected most often. If the lower rate is less than four-fifths of the higher rate, the difference may signal adverse impact and warrant closer examination. It is a screening rule, not proof that a system is fair.
face validity
Apparent relevance to workFace validity describes whether a test appears, to the people taking it, to measure something relevant to the job. A puzzle-based assessment for a data analyst might lack face validity if candidates cannot see how solving puzzles relates to the work. Appearance alone does not establish scientific validity, but a test that seems irrelevant can undermine trust.
data exhaust
Traces left by digital activityData exhaust is the information people generate as a by-product of using digital services, including clicks, likes, searches, posts, and interactions. The book describes how employers and vendors may use these traces to infer preferences, skills, personality, or risk. Such inferences can affect employment even when people did not create the data for hiring.
overfitting
Learning accidental patternsOverfitting occurs when a machine-learning model captures patterns specific to its training data rather than the meaningful relationship it is supposed to learn. A model trained to distinguish wolves from huskies may rely on snow in the background if that feature happened to correlate with wolves in its examples. The resulting prediction can fail on new cases.
quiet hiring
Filling needs with existing staffQuiet hiring refers to meeting organizational needs by finding or developing skills among current employees, rather than relying only on external recruitment. Internal talent platforms can suggest roles, projects, and training based on employees’ skills and possible skill adjacencies. The practice can create mobility, though its recommendations and employee data still require scrutiny.
flight risk
Predicted likelihood of leavingA flight-risk score estimates whether an employee may leave an organization within a given period. Models may use information such as engagement, absence, pay, promotions, performance reviews, or work patterns. The score is uncertain, especially for individuals, and can influence how managers allocate raises, development opportunities, or attention.
vocal biomarkers
Voice-based health signalsVocal biomarkers are measurable features of speech, such as pauses, clarity, pitch, or speaking rate, that developers analyze for possible signs of health conditions or emotional states. Companies described in the book claim they may help identify anxiety or depression. The technology remains difficult to validate, and its use can expose sensitive information.
predictive society
Life shaped by behavioral forecastsThe book uses predictive society to describe a condition in which organizations use accumulated personal data to assess and forecast people’s behavior across many areas of life. In employment, digital traces may shape assumptions about personality, career potential, or risk. The concern is that predictions can narrow people’s opportunities and autonomy, even when the predictions are uncertain.
FAQ
What's The Algorithm about?
- AI in Employment: The Algorithm by Hilke Schellmann delves into the growing use of artificial intelligence in hiring, monitoring, promoting, and firing employees, highlighting its impact on workplace fairness and discrimination.
- Real-World Implications: Through case studies and personal stories, the book illustrates the real-world consequences of AI-driven decisions in employment, emphasizing the need for awareness and action.
- Call for Change: Schellmann advocates for transparency and accountability in AI systems, urging society to challenge their use in employment decisions.
Why should I read The Algorithm?
- Understanding AI's Impact: The book provides insights into how AI technologies affect employment, making it essential for job seekers and HR professionals to understand potential biases.
- Awareness of Discrimination: It highlights how AI can reinforce existing biases against marginalized groups, emphasizing the importance of being informed.
- Practical Advice: Schellmann offers practical tips for navigating AI-driven hiring processes, making it a valuable resource for job seekers.
What are the key takeaways of The Algorithm?
- AI's Prevalence in Hiring: The book reveals that AI is used by 99% of Fortune 500 companies for hiring, yet many job seekers are unaware of its operation.
- Bias in Algorithms: Schellmann discusses how algorithms can discriminate based on race, gender, and other factors due to biased training data.
- Need for Oversight: The author argues for regulations to ensure fair and transparent use of AI in hiring, to prevent systemic discrimination.
How does The Algorithm address bias in AI hiring tools?
- Real Stories Highlight Bias: The book shares personal accounts of individuals facing discrimination due to flawed AI systems, illustrating the impact of these biases.
- Critique of Training Data: Schellmann emphasizes that biased training data can lead to discriminatory outcomes, particularly concerning in high-stakes hiring.
- Call for Transparency: The author advocates for transparency in AI tool development and usage, urging companies to disclose methodologies and training data.
What are some examples of AI tools discussed in The Algorithm?
- Résumé Screeners: These tools analyze résumés for suitability but can be biased, as seen in Amazon's failed screener that discriminated against women.
- Video Interviewing Software: AI analyzes candidates' expressions and tone in video interviews, criticized for lack of scientific validity and potential misinterpretation.
- Gamified Assessments: Tools like Pymetrics use games to assess traits, aiming to reduce bias but raising concerns about relevance and effectiveness.
What are the best quotes from The Algorithm and what do they mean?
- Promises vs. Reality: “The ‘revolution’ and fair decision-making that AI vendors promise were definitely missing in her case.” This highlights the gap between AI vendors' promises and the reality faced by individuals.
- Discrimination Warning: “Discrimination could be rampant in these automated systems.” This serves as a warning about unchecked algorithm use in high-stakes decisions.
- Fairness in Hiring: “We need to ensure that employment opportunities are based on merit and qualifications.” This encapsulates the book's message about fairness in hiring.
How does AI affect hiring decisions according to The Algorithm?
- Automated Screening: AI is used to screen resumes, potentially overlooking qualified candidates due to biases.
- Opaque Decision-Making: Candidates often don't know how their applications are evaluated or why they were rejected, highlighting a lack of transparency.
- Potential Discrimination: AI tools can unintentionally discriminate, particularly against those with disabilities, by not accounting for individual needs.
What are the ethical concerns raised in The Algorithm?
- Privacy Issues: Concerns about personal data collection and use by AI systems, often without individuals' consent or knowledge.
- Bias in Algorithms: Using biased data to train AI systems can lead to unfair treatment based on race, gender, or disability.
- Impact on Rights: AI in hiring and monitoring can undermine workers' rights, particularly regarding transparency and accountability.
How does The Algorithm address the issue of disability in the workplace?
- Disability Discrimination: AI hiring tools can discriminate against individuals with disabilities, often failing to accommodate their needs.
- Real-Life Challenges: Narratives from vocational counselors and job seekers illustrate the impact of AI-driven hiring on those with disabilities.
- Advocacy for Inclusivity: The author calls for more inclusive hiring practices that consider candidates' abilities beyond automated assessments.
What recommendations does The Algorithm make for companies using AI?
- Conduct Bias Audits: Regularly audit AI tools for bias and effectiveness to prevent discrimination.
- Human Interaction: Maintain human oversight in hiring to allow meaningful conversations about candidates' qualifications.
- Transparent Practices: Be transparent about AI tool usage and provide candidates with clear evaluation information.
How can job seekers navigate AI-driven hiring processes as discussed in The Algorithm?
- Optimize Resumes: Create machine-readable resumes with common templates and keywords to pass automated screenings.
- Network and Follow Up: Increase visibility and demonstrate interest by networking and following up with recruiters.
- Prepare for AI Assessments: Familiarize with AI assessment tools and prepare, understanding they may not fully capture abilities.
How does The Algorithm suggest we can improve AI hiring practices?
- Diverse Data: Use diverse and representative data sets to train AI systems, ensuring inclusivity and reducing bias.
- Engage Stakeholders: Involve employees and advocacy groups in discussions about AI tool implementation in hiring.
- Legislative Action: Call for stronger regulations and oversight of AI technologies to protect workers' rights and ensure fair practices.
Download PDF
Download EPUB
.epub digital book format is ideal for reading ebooks on phones, tablets, and e-readers.