Borrowed Competence: When AI-Driven HR Builds Startups’ People Capability and When It Hollows It Out

May combine two claims.

Hiring is close to a textbook wicked learning environment: outcomes for rejected candidates are never observed, outcomes for those hired are confounded by onboarding, manager quality, and team composition, and feedback arrives 6 to 18 months late.

“Hiring is close to a textbook wicked learning environment.Outcomes for rejected candidates are never observed, outcomes for those hired are confounded by onboarding, manager quality, and team composition, and feedback arrives 6 to 18 months late [41,42].”

• cites [41] R M Hogarth, 2015, The Two Settings of Kind and Wicked Learning Environments, Current Directions in Psychological Science sourcePartly supported

  Split into 2 atoms — 0 confirmed, 2 partially supported, 0 disputed, 0 misattributed, 0 unverifiable. Expand the claim to see per-atom verdicts.

  Experimental — accuracy not yet measured
• cites [42] P Tambe, 2019, Artificial Intelligence in Human Resources Management: Challenges and a Path Forward, California Management Review sourceNot checked — no full text of the source

May combine two claims.

The result of the system was narrower pools and dissatisfied hiring managers.

“The result was narrower pools and dissatisfied hiring managers.”

• cites [35] E Van Den Broek, 2025, Is There Fairness in AI?, New Research on AI and Fairness in Hiring sourceSupported

Classified as probably not a substantial claim. May combine two claims.

The buyer runs the whole-process case: an AI tool checks candidates' licenses and certifications, tests their English, and conducts a professional interview; only those above a threshold reach a recruiter for a short conversation about availability and loose ends.

“An AI tool checks the candidates' licenses and certifications, tests their English, and conducts a professional interview.Only those above a threshold reach a recruiter for a short conversation about availability and loose ends.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Our vendor informants described the same gap from the other side of the transaction.

“Our vendor informants described the same gap from the other side of the transaction.”

Classified as probably not a substantial claim.

What was never assessed before--social fit and a meeting with the future manager--stays out now, because the manager has no way to check it.

“What was never assessed before--social fit and a meeting with the future manager--stays out now, because, as the manager puts it, he has no way to check it.”

Classified as probably not a substantial claim. May combine two claims.

Validation is scheduled every six months, and the manager follows up with early leavers, but when asked whether the validation rested on data, he gave no numbers and instead described the people who had passed and joined the firm.

“Validation is scheduled every six months, and he follows up with early leavers.However, when asked whether the validation rested on data, he gave no numbers.Instead, he described the people who had passed and joined the firm.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

Firms examine their hiring only after something happens, usually a wave of departures, and do so with measures that are late and indirect.

“Firms examine their hiring only after something happens, usually a wave of departures, and does so with measures that are late and indirect.”

May combine two claims.

The vendor maintains that a person has to stay in the loop, but the person does not settle who answers for a decision whose basis the system cannot state: the recruiter, the manager, or whoever selected the tool.

“The vendor maintains that a person has to stay in the loop.However, the person does not settle who answers for a decision whose basis the system cannot state: the recruiter, the manager, or whoever selected the tool.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Handing over the entire hiring process to a tool is a phenomenon a vendor sees mainly in blue-collar and mass hiring; elsewhere she sees parts of the process automated, particularly screening and interviewing first.

“Handing over the entire hiring process to a tool is a phenomenon she sees mainly in blue-collar and mass hiring.Elsewhere, she sees parts of the process automated, particularly screening and interviewing first.”

Classified as probably not a substantial claim. May combine two claims.

In startups, the founder's assumptions about who fits enter the process early, remain through the growth stage, and do not leave as the firm grows.

“In startups, she added, the founder's assumptions about who fits enter the process early.They remain through the growth stage, and do not leave as the firm grows.”

Classified as probably not a substantial claim.

Most hiring processes a vendor informant encounters are broken somewhere: in the role definition, the criteria, a stage of the screening, or the way the decision is made.

“Most hiring processes she encounters are broken somewhere: in the role definition, the criteria, a stage of the screening, or the way the decision is made.”

Classified as probably not a substantial claim.

The tool applies the criteria the organization gave it.

“The tool applies the criteria the organization gave it.”

Probably not a substantial claim: may not pass S1 (asserted), S3 (load-bearing), S4 (based). May combine two claims.

According to this interviewee, a tool with almost no biases that saves time results in better hiring and more retention.

“According to this interviewee, a tool with almost no biases that saves time results in better hiring and more retention.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Sara automated sourcing, scheduling, and the drafting of job descriptions from the start, and has never employed a full-time recruiter either.

“Sara automated sourcing, scheduling, and the drafting of job descriptions from the start, and has never employed a full-time recruiter either.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Sara's logistics software company reached 120 employees on much the same stack Maya used.

“Sara's logistics software company reached 120 employees on much the same stack Maya used.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

Two activities are tracked across four stages.

“Two activities are tracked across four stages.”

May combine two claims.

Retained recruiters have long rented some judgment about whom to hire for senior roles, but a recruiter presents a shortlist with reasons and leaves the decision, and the argument for it, with the founder.

“Retained recruiters have long rented some of that judgment for senior roles.Nevertheless, a recruiter presents a shortlist with reasons and leaves the decision, and the argument for it, with the founder.”

Probably not a substantial claim: may not pass S2 (checkable). May combine two claims.

A firm in borrowed competence is not failing to build capability; it is building a different one.

“A firm in this condition is not failing to build capability.It is building a different one.”

May combine two claims.

Retraining presumes a baseline, documentation presumes knowledge the firm once held, and reinsourcing presumes an internal standard for judging the vendor; borrowed competence supplies none of the three, so it is far cheaper to prevent than to build after the fact.

“Retraining presumes a baseline, documentation presumes knowledge the firm once held [18], and reinsourcing presumes an internal standard for judging the vendor.Borrowed competence supplies none of the three, so it is far cheaper to prevent than to build after the fact.”

• cites [18] P M De Holan, 2004, Remembrance of Things Past? The Dynamics of Organizational Forgetting, Management Science sourceNot checked — no full text of the source

Probably not a substantial claim: may not pass S2 (checkable).

A certificate covering a vendor's model across all its customers is not a record of the firm's configuration.

“A certificate covering a vendor's model across all its customers is not a record of the firm's configuration.”

May combine two claims.

The endoscopists detected precancerous growths at a measurably lower rate after exposure to AI assistance than before, and the capability they lost was demonstrably present beforehand.

“For example, the endoscopists we will discuss later detected precancerous growths at a measurably lower rate after exposure to AI assistance than before, and the capability they lost was demonstrably present beforehand [16].”

• cites [16] K Budzyn, 2025, Endoscopist Deskilling Risk After Exposure to Artificial Intelligence in Colonoscopy: A Multicentre, Observational Study, The Lancet Gastroenterology and Hepatology sourceNot checked — no full text of the source

May combine two claims.

A professional employer organization rents administration (payroll, benefits enrollment, compliance filing) but does not rent judgments about whom to hire.

“However, a professional employer organization rents administration: payroll, benefits enrollment, compliance filing.It does not rent judgments about whom to hire.”

• cites [11] B S Klaas, 2000, Managing HR in the Small and Medium Enterprise: The Impact of Professional Employer Organizations, Entrepreneurship Theory and Practice sourceNot checked — no full text of the source

May combine two claims.

With 15 employees, a bad hire is visible within weeks, and the founder can see exactly what was missed.

“With 15 employees, a bad hire is visible within weeks, and the founder can see exactly what was missed.”

• AssumptionRests on an assumption: The window in which a founder can learn from people decisions is short and closes quickly.
• AssumptionRests on an assumption: Judgment forms using fast, clear, personally consequential feedback.

Probably not a substantial claim: may not pass S2 (checkable), S4 (based). May combine two claims.

Outsourcing moves execution outside the firm, while the ability to judge the work stays behind.

“Outsourcing moves execution outside the firm, while the ability to judge the work stays behind.”

May combine two claims.

Automation eliminates neither the hires nor their outcomes.

“However, automation eliminates neither the hires nor their outcomes.”

May combine two claims.

The people model is installed early and holds, but the demands on it keep moving.

“The people model is installed early and holds, but the demands on it keep moving.”

The quote only partly states this claim.

A founder who has never screened candidates unaided has no comparable baseline to deskilled endoscopists.

“A founder who has never screened candidates unaided has no comparable baseline.”

May combine two claims.

AI-enabled tools rent the decision itself, evaluating candidates, ranking them, and in many configurations rejecting them before anyone even looks at them.

“AI-enabled tools rent the decision itself, evaluating candidates, ranking them, and in many configurations rejecting them before anyone even looks at them.”

Probably not a substantial claim: may not pass S2 (checkable).

A defense the firm cannot make from its own records is not a defense.

“And a defense the firm cannot make from its own records is not a defense.”

May combine two claims.

Outsourcing execution leaves judgment where it was; outsourcing judgment moves it out of the firm.

“Outsourcing execution leaves judgment where it was; outsourcing judgment moves it out of the firm.”

May combine two claims.

Adapting the people model to what the firm has become requires making judgments about people, and it comes due years after the period in which those judgments would have been formed.

“Adapting it to what the firm has become requires making judgments about people, and it comes due years after the period in which those judgments would have been formed.”

Classified as probably not a substantial claim.

The classic warning about hollowing out concerns activities a firm once performed and could perform again.

“The classic warning about hollowing out concerns activities a firm once performed and could perform again [15].”

• cites [15] R A Bettis, 1992, Outsourcing and Industrial Decline sourceNot checked — no full text of the source

Classified as probably not a substantial claim. May combine two claims.

Founders who buy their HR function become genuinely skilled at scoping a need, running a trial, and switching vendors, but that skill cannot produce the standard by which a vendor's output is judged; therefore, the firm becomes better at borrowing but no better at the thing it is borrowing.

“Founders who buy their HR function become genuinely skilled at scoping a need, running a trial, and switching vendors.What that skill cannot produce is the standard by which a vendor's output is judged.Therefore, the firm becomes better at borrowing but no better at the thing it is borrowing.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

New York City requires annual independent bias audits of automated employment decision tools, public posting of the audit's summary, and at least 10 business days' notice to candidates.

“For example, New York City requires annual independent bias audits of automated employment decision tools, public posting of the audit's summary, and at least 10 business days' notice to candidates [13].”

• cites [13] 2023, Requires an annual independent bias audit, public posting of the audit summary, and at least ten business days' notice to candidatesNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Absorptive capacity explains why a founder cannot evaluate a vendor, but not how the founder came to be in that position while the firm was apparently doing the work.

“It explains why a founder cannot evaluate a vendor [17], but not how the founder came to be in that position while the firm was apparently doing the work.”

• cites [17] W M Cohen, 1990, Absorptive Capacity: A New Perspective on Learning and Innovation, Administrative Science Quarterly sourceNot checked — no full text of the source

May combine two claims.

A vendor's certificate describes a model across all its customers, but what a plaintiff asks for is what this firm did, and a firm that never articulated its criteria has no answer that originates with it.

“A vendor's certificate describes a model across all its customers.What a plaintiff asks for is what this firm did, and a firm that never articulated its criteria has no answer that originates with it.”

• AssumptionRests on an assumption: A vendor's certificate describes a model across all its customers.

Classified as probably not a substantial claim.

The authors use the erosion evidence for the mechanism, as a well-grounded risk to manage rather than a documented finding about startups.

“We use it for the mechanism, as a well-grounded risk to manage rather than a documented finding about startups.”

May combine two claims.

Currently, most firms are using AI for drafting purposes, with job descriptions ahead of resume screening.

“Currently, most firms are using AI for drafting purposes, with job descriptions ahead of resume screening [22].”

• cites [22] Shrm, 2025, Talent Trends: Artificial Intelligence in HR (Alexandria, VA: Society for Human Resource ManagementNot checked yet

  Not checked in the free run.

May combine two claims.

The classic validity estimates have been revised downward, although the relative standing of structured methods has survived the correction.

“The classic validity estimates have been revised downward, although the relative standing of structured methods have survived the correction [24].”

• cites [24] P R Sackett, 2022, Revisiting Meta-Analytic Estimates of Validity in Personnel Selection: Addressing Systematic Overcorrection for Restriction of Range, Journal of Applied Psychology sourceNot checked — no full text of the source

Probably not a substantial claim: may not pass S4 (based). May combine two claims.

In a runway-limited firm, time-to-hire is a survival metric, not an efficiency one; a key role vacant for four months can cost a funding milestone.

“In a runway-limited firm, time-to-hire is a survival metric, not an efficiency one.A key role vacant for four months can cost a funding milestone.”

May combine two claims.

The activity looks identical from outside and produces the same hire, but it creates a different firm.

“The activity looks identical from outside and produces the same hire.However, it creates a different firm.”

May combine two claims.

Both questions that buyers do ask are answerable without any knowledge of hiring; the two they do not ask are the ones that determine whether the tool builds judgment or absorbs it.

“Both questions that buyers do ask are answerable without any knowledge of hiring.The two they do not ask are the ones that determine whether the tool builds judgment or absorbs it.”

Probably not a substantial claim: may not pass S4 (based).

What the founder described near the end of the article has and Maya does not is a stated, tested, revisable account of whom the firm hires and why.

“What the founder we describe near the end of the article has and Maya does not is a stated, tested, revisable account of whom the firm hires and why.”

May combine two claims.

It is validation within the selected sample, which may reveal early failures among those hired but cannot show whom the threshold wrongly excluded.

“It is validation within the selected sample, which may reveal early failures among those hired but cannot show whom the threshold wrongly excluded.”

May combine two claims.

Accountability does not distribute; it asks who must answer, and that cannot be the system, which can neither face a rejected candidate nor change the practice that produced the rejection.

“Accountability does not distribute.It asks who must answer, and that cannot be the system, which can neither face a rejected candidate nor change the practice that produced the rejection [40].”

• cites [40] H Ma, Whom to Sue? Liability of Unaccountability in AI Decisions, Organizational DynamicsNot matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Classified as probably not a substantial claim. May combine two claims.

The interviews are not a sample from which the authors generalize; they are how the authors came to see the pattern and the source of the two composite firms.

“The interviews are not a sample from which we generalize.They are how we came to see the pattern, and they are the source of the two composite firms described below.”

Classified as probably not a substantial claim. May combine two claims.

Beane found that robotic surgery, which lets the attending surgeon perform an entire procedure alone, left most residents unable to operate; their skill was not eroded, but the occasions on which it would have formed were removed.

“Beane found that robotic surgery, which lets the attending surgeon perform an entire procedure alone, left most residents unable to operate.Their skill was not eroded.The occasions on which it would have formed were removed [19].”

• cites [19] M I Beane, 2019, Shadow Learning: Building Robotic Surgical Skill When Approved Means Fail, Administrative Science Quarterly sourceNot checked — no full text of the source

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

If reliable human judgment does not form from 20 hires a year regardless of automation, there is nothing to protect.

“If reliable human judgment does not form from 20 hires a year regardless of automation, there is nothing to protect.”

May combine two claims.

Hiring supplies repetition readily and feedback poorly, because outcomes arrive late and confounded, and rejected candidates are never observed.

“Hiring supplies repetition readily and feedback poorly, because outcomes arrive late and confounded, and rejected candidates are never observed.”

May combine two claims.

Fairness was not embedded in the code; it was negotiated by the people who built and deployed the system, and then frozen.

“Fairness was not embedded in the code.It was negotiated by the people who built and deployed the system, and then frozen.”

• cites [35] E Van Den Broek, 2025, Is There Fairness in AI?, New Research on AI and Fairness in Hiring sourceNot checked yet

  Checked before — its earlier verdict is held back until a new check runs.

May combine two claims.

A tool built on a broken hiring process does not run a sound process faster; it learns the one it is given, breaks included, and repeats it.

“A tool built on such a process does not run a sound process faster; it learns the one it is given, breaks included, and repeats it.”

Classified as probably not a substantial claim. May combine two claims.

Raisch and Krakowski distinguish automation, in which the machine takes over a task, from augmentation, in which human and machine capabilities are combined by design, and theorize the two as interdependent rather than mutually exclusive.

“Its vocabulary follows Raisch and Krakowski, who distinguish automation, in which the machine takes over a task, from augmentation, in which human and machine capabilities are combined by design, and theorize the two as interdependent rather than mutually exclusive [45].”

• cites [45] S Raisch, 2021, Artificial Intelligence and Management: The Automation-Augmentation Paradox, Academy of Management Review sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S2 (checkable).

The capability at stake here was never the founder's instinct.

“It is also why the capability at stake here was never the founder's instinct.”

May combine two claims.

In a startup, the founder is the person who was supposed to learn, and expertise hired later cannot retrieve decisions no one wrote down.

“In a startup, the founder is the person who was supposed to learn, and expertise hired later cannot retrieve decisions no one wrote down.”

May combine two claims.

In a large firm, a screening tool is one input into a function staffed by specialists with legal counsel to advise them; in a 30-person firm, the tool frequently is the function.

“In a large firm, a screening tool is one input into a function staffed by specialists with legal counsel to advise them.In a 30-person firm, the tool frequently is the function.”

May combine two claims.

The asymmetry is not about how many small firms adopt these tools; it is about what adoption means in a small firm.

“The asymmetry is not about how many small firms adopt these tools.It is about what adoption means in a small firm.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Pew puts the figure at about two-thirds: 66 percent of U.S. adults say they would not want to apply to an employer that uses AI to help make hiring decisions, as opposed to 32 percent who would.

“Pew puts the figure at about two-thirds: 66 percent of U.S. adults say they would not want to apply to an employer that uses AI to help make hiring decisions, as opposed to 32 percent who would [38].”

• cites [38] 2022, Survey of 11,004 U.S. adults conducted, AI in Hiring and Evaluating Workers: What Americans ThinkNot matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Probably not a substantial claim: may not pass S3 (load-bearing).

The colonoscopy study was observational.

“As the authors stress, the design of the study was observational.”

May combine two claims.

The founders described have no measured baseline, which makes the same erosion dynamic harder to detect rather than less consequential.

“These are cases of erosion from a measured baseline; the founders we describe have no such baseline, which makes the same dynamic harder to detect rather than less consequential.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Together Beane and Balasubramanian et al. establish the general point that a machine doing the work can prevent the firm from learning it.

“Together these two studies establish the general point that a machine doing the work can prevent the firm from learning it, but that point is not ours.”

May combine two claims.

Because the tool substitutes for the learning that would build assessment capacity, the deficit does not close with use and may widen.

“Given that the tool substitutes for the learning that would build assessment capacity, the deficit does not close with use and may indeed widen.”

May combine two claims.

Human beings reviewing recommendations they lack the expertise or incentive to challenge is not a safeguard but a ratification step, and ratification is the mechanism by which the decisions keep happening while the judgment behind them never forms.

“Human beings reviewing recommendations they lack the expertise or incentive to challenge is not a safeguard but a ratification step, and ratification is the mechanism by which the decisions keep happening while the judgment behind them never forms.”

• AssumptionRests on an assumption: Human review is the standard assurance offered when AI screening is questioned.

May combine two claims.

Absorptive capacity holds that a firm's ability to recognize and apply external knowledge depends on prior related knowledge in the same domain; one needs to know something about a field to evaluate expertise in it.

“Absorptive capacity holds that a firm's ability to recognize and apply external knowledge depends on prior related knowledge in the same domain [17].One needs to know something about a field to evaluate expertise in it.”

• cites [17] W M Cohen, 1990, Absorptive Capacity: A New Perspective on Learning and Innovation, Administrative Science Quarterly sourceNot checked — no full text of the source

May combine two claims.

Systems that share their reasoning can train rather than substitute, but that is a design and configuration question rather than a question about AI.

“Systems that share their reasoning can train rather than substitute.However, that is a design and configuration question rather than a question about AI.”

Classified as probably not a substantial claim.

The person who is answerable is the one whose part in the decision has narrowed to a summary and a short call.

“The person who is answerable is the one whose part in the decision has narrowed to a summary and a short call.”

Classified as probably not a substantial claim. May combine two claims.

One role is hired at volume, its requirements are largely codifiable, and a recruiting function exists; each of these is a low-risk condition in Table 3, and a first-time founder typically has none of them.

“One role is hired at volume, its requirements are largely codifiable, and a recruiting function exists.Each of these is a low-risk condition in Table 3, and a first-time founder typically has none of them.”

Probably not a substantial claim: may not pass S2 (checkable).

The objection strengthens rather than weakens the practical recommendation.

“What we would say is that the objection strengthens rather than weakens the practical recommendation.”

May combine two claims.

AI-assisted hiring probably produces better decisions than most founders do unassisted, unstructured interviewing by founders being among the weaker predictors of job performance in the selection literature.

“AI-assisted hiring probably produces better decisions than most founders do unassisted, unstructured interviewing by founders being among the weaker predictors of job performance in the selection literature.”

• cites [23] F L Schmidt, 1998, The Validity and Utility of Selection Methods in Personnel Psychology, Psychological Bulletin sourceNot checked — no full text of the source
• cites [24] P R Sackett, 2022, Revisiting Meta-Analytic Estimates of Validity in Personnel Selection: Addressing Systematic Overcorrection for Restriction of Range, Journal of Applied Psychology sourceNot checked — no full text of the source

Probably not a substantial claim: may not pass S2 (checkable), S4 (based).

These are not non-compliance rates, because the law grants firms a great deal of discretion over whether a system fails in scope.

“These are not non-compliance rates, because the law grants firms a great deal of discretion over whether a system fails in scope.”

• AssumptionRests on an assumption: The law grants firms a great deal of discretion over whether a system fails in scope.

May combine two claims.

People monitoring reliable systems under-monitor them and accept recommendations even when contradictory information is available; therefore, the design question is appropriate reliance rather than maximum trust.

“People monitoring reliable systems under-monitor them and accept recommendations even when contradictory information is available.Therefore, the design question is appropriate reliance rather than maximum trust [28,29].”

• cites [28] R Parasuraman, 2010, Complacency and Bias in Human Use of Automation: An Attentional Integration, Human Factors sourceNot checked — no full text of the source
• cites [29] J D Lee, 2004, Trust in Automation: Designing for Appropriate Reliance, Human Factors sourceNot checked — no full text of the source

Probably not a substantial claim: may not pass S4 (based). May combine two claims.

The gains from AI in HR are real, immediate, and correctly identified by the founders taking them, which is why nothing about the arrangement prompts a second look.

“The gains are real, immediate, and correctly identified by the founders taking them, which is why nothing about the arrangement prompts a second look.”

Probably not a substantial claim: may not pass S2 (checkable), S4 (based). May combine two claims.

The condition forms rather than strikes, and the sequence is driven by the tool working rather than failing.

“The condition forms rather than strikes, and the sequence is driven by the tool working rather than failing.”

• AssumptionRests on an assumption: The condition forms rather than strikes, and the sequence is driven by the tool working rather than failing.

Probably not a substantial claim: may not pass S2 (checkable).

Learning yield is not an arbitrary axis.

“Learning yield is not an arbitrary axis.”

Probably not a substantial claim: may not pass S2 (checkable). May combine two claims.

The boundary has an owner, but that owner expects it to move, which is the drift that Tool 3 is designed to catch.

“The boundary has an owner, but that owner expects it to move, which is the drift that Tool 3, below, is designed to catch.”

May combine two claims.

Algorithmic selection is generally seen as less fair than human or human-assisted processes, with the most consistent negative effects on perceived social presence and interpersonal treatment.

“Algorithmic selection is generally seen as less fair than human or human-assisted processes, with the most This preprint research paper has not been peer reviewed.Electronic copy available at: https://ssrn.com/abstract=7489654P r e p r i n t n o t p e e r r e v i e w e d consistent negative effects on perceived social presence and interpersonal treatment [37].”

• cites [37] A Hilliard, 2022, Robots Are Judging Me: Perceived Fairness of Algorithmic Recruitment Tools, Frontiers in Psychology sourceNot checked yet

  Checked before — its earlier verdict is held back until a new check runs.

May combine two claims.

The system hard-coded one definition of fairness--standardization under uniform rules--and progressively deprioritized managers' local judgment about role requirements, labor markets, and team needs.

“However, it hard-coded one definition of fairness--standardization under uniform rules--and progressively deprioritized managers' local judgment about role requirements, labor markets, and team needs [35].”

• cites [35] E Van Den Broek, 2025, Is There Fairness in AI?, New Research on AI and Fairness in Hiring sourceNot checked yet

  Checked before — its earlier verdict is held back until a new check runs.

May combine two claims.

A study of 391 employers, in which 155 student investigators recorded compliance as a prospective job applicant would encounter it, found that only 18 had posted a bias audit report and only 13 a transparency notice, roughly five and three percent, respectively.

“A study of 391 employers, in which 155 student investigators recorded compliance as a prospective job applicant would encounter it, found that only 18 had posted a bias audit report and only 13 a transparency notice, roughly five and three percent, respectively [1].”

• cites [1] L Wright, Null Compliance: NYC Local Law 144 and the Challenges of Algorithm Accountability, Proceedings of the 2024 ACM Conference on Fairness, Accountability, and Transparency (FAccT '24)Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Probably not a substantial claim: may not pass S2 (checkable).

The firms most exposed to the objection have the most to gain from the practices described below.

“The firms most exposed to the objection have the most to gain from the practices described below.”

May combine two claims.

A founder who ran hiring for 200 people elsewhere faces a different problem: the model may be stale or wrong for this firm, but at least the founder has a standard for evaluating what the tool returns.

“A founder who ran hiring for 200 people elsewhere faces a different problem.The model may be stale or wrong for this firm, but at least the founder has a standard for evaluating what the tool returns.”

May combine two claims.

Unstructured interviewing is among the weaker predictors of job performance, and founder-led hiring is unstructured in the extreme.

“Unstructured interviewing is among the weaker predictors of job performance, and founder-led hiring is unstructured in the extreme [23].”

• cites [23] F L Schmidt, 1998, The Validity and Utility of Selection Methods in Personnel Psychology, Psychological Bulletin sourceNot checked — no full text of the source

Probably not a substantial claim: may not pass S2 (checkable). May combine two claims.

If feedback in hiring is genuinely poor, then articulating the issues by writing down why, comparing decisions, and testing criteria against outcomes is not an optional supplement to experience; it is the only mechanism by which experience becomes capability at all.

“If feedback in hiring is genuinely poor, then articulating the issues by writing down why, comparing decisions, and testing criteria against outcomes is not an optional supplement to experience.It is the only mechanism by which experience becomes capability at all.”

Probably not a substantial claim: may not pass S2 (checkable).

Our framework turns that interdependence into a deliberate managerial boundary choice.

“Our framework turns that interdependence into a deliberate managerial boundary choice.”

May combine two claims.

When a tool makes the decision and a human ratifies the output, the repetitions still occur but nothing is articulated, and the mechanism by which repetition becomes capability is broken.

“When a tool makes the decision and a human ratifies the output, the repetitions still occur but nothing is articulated, and the mechanism by which repetition becomes capability is broken.”

Classified as probably not a substantial claim.

The study calls the condition null compliance: absence of evidence cannot be read as evidence of absence.

“The study calls the condition null compliance: absence of evidence cannot be read as evidence of absence.”

• cites [1] L Wright, Null Compliance: NYC Local Law 144 and the Challenges of Algorithm Accountability, Proceedings of the 2024 ACM Conference on Fairness, Accountability, and Transparency (FAccT '24)Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

May combine two claims.

Consistency improves on the real baseline, which is not AI against expert human judgment but AI against a distracted technical founder screening resumes at 11:00 PM between customer calls.

“Consistency improves on the real baseline.This is not AI against expert human judgment but AI against a distracted technical founder screening resumes at 11:00 PM between customer calls.”

May combine two claims.

The firm needs human judgment for the residual, and the objection offers no account of where it comes from.

“The firm needs human judgment for the residual, and the objection offers no account of where it comes from.”

Probably not a substantial claim: may not pass S3 (load-bearing).

A large technology firm's abandoned internal screening tool disadvantaged women after learning from a historical hiring record.

“A large technology firm's abandoned internal screening tool, which disadvantaged women after learning from a historical hiring record, illustrates the mechanism rather than being the evidence for it [34].”

• cites [34] J Dastin, 2018, Amazon Scraps Secret AI Recruiting Tool That Showed Bias Against Women, ReutersNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S2 (checkable), S4 (based). May combine two claims.

The upper-right quadrant is the heart of the problem; its position is not a coincidence but the structure of the situation.

“The upper-right quadrant is the heart of the problem.Its position is not a coincidence but the structure of the situation.”

May combine two claims.

The evidence of erosion is established in studies in medicine, aviation, and process control, but not yet demonstrated for HR judgment in entrepreneurial firms.

“The evidence in this section is evidence of erosion, established in studies in medicine, aviation, and process control, but not yet demonstrated for HR judgment in entrepreneurial firms.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Balasubramanian, Ye, and Xu argued that organizations learn from decisions through the people who make them; therefore, when machine learning substitutes for these people, the organization keeps deciding and stops learning.

“Balasubramanian, Ye, and Xu argued that organizations learn from decisions through the people who make them.Therefore, when machine learning substitutes for these people, the organization keeps deciding and stops learning [20].”

• cites [20] N Balasubramanian, Substituting Human Decision-Making with Machine Learning: Implications for Organizational Learning, Academy of Management Review sourceNot checked — no full text of the source

Classified as probably not a substantial claim. May combine two claims.

Well-trained and well-deployed AI systems reduce noise and enforce consistency.

“While conceding that well-trained and well-deployed systems reduce noise and enforce consistency, Chamorro-Premuzic's assessment is that AI has made hiring worse, an arms race of automation on both sides [25].”

• cites [25] T Chamorro-Premuzic, 2026, AI Has Made Hiring Worse -But It Can Still Help, Harvard Business ReviewNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S2 (checkable).

Approving a ranked list is not deciding.

“Approving a ranked list is not deciding.”

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing), S4 (based). May combine two claims.

It is the cheapest available substitute for the feedback that hiring does not naturally supply, and it is the cheapest item in this article.

“It is the cheapest available substitute for the feedback that hiring does not naturally supply, and it is the cheapest item in this article.”

Probably not a substantial claim: may not pass S2 (checkable). May combine two claims.

Retraining presumes a baseline, and reinsourcing presumes an internal standard.

“Retraining presumes a baseline, and reinsourcing presumes an internal standard.”

Probably not a substantial claim: may not pass S2 (checkable).

The automation boundary decision is the build-versus-buy decision applied to the organization.

“because it is that decision applied to the organization.”

Probably not a substantial claim: may not pass S2 (checkable), S4 (based). May combine two claims.

Moving sideways, from augment into automate, means an activity that builds human judgment is now being treated as though it does not, and no one decided that.

“Moving sideways, from augment into automate, means an activity that builds human judgment is now being treated as though it does not, and no one decided that.”

Probably not a substantial claim: may not pass S2 (checkable), S4 (based).

The upper-right quadrant is where borrowed competence is manufactured, because the pressure to hand off is maximal precisely where the learning lives.

“The upper-right quadrant is where borrowed competence is manufactured, because the pressure to hand off is maximal precisely where the learning lives.”

Probably not a substantial claim: may not pass S2 (checkable).

A date-based review will miss drift, because the boundary moves when the firm changes rather than when the calendar does.

“A date-based review will miss it, because the boundary moves when the firm changes rather than when the calendar does.”

Probably not a substantial claim: may not pass S4 (based). May combine two claims.

High-volume screening is both the most burdensome HR activity in a growing startup and the single richest source of judgment about who succeeds in the firm.

“High-volume screening is both the most burdensome HR activity in a growing startup and the single richest source of judgment about who succeeds in the firm.”

Probably not a substantial claim: may not pass S4 (based).

In most growing firms, the danger zone means high-volume screening.

“which in most growing firms means high-volume screening.”

Probably not a substantial claim: may not pass S4 (based). May combine two claims.

Calendar reviews get skipped; doublings are unmissable and mark the real structural breaks in how a firm operates.

“Calendar reviews get skipped; doublings are unmissable and mark the real structural breaks in how a firm operates.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

Nothing in stages 1 through 3 registers as a warning, because the tool is performing as promised.

“Nothing in stages 1 through 3 registers as a warning, because the tool is performing as promised.”

Probably not a substantial claim: may not pass S2 (checkable), S4 (based).

Stage 3 is the last cheap point of intervention.

“Stage 3 is the last cheap point of intervention.”

Probably not a substantial claim: may not pass S2 (checkable), S4 (based). May combine two claims.

Renting capability is a legitimate strategy; believing you own what you are renting is not.

“Renting capability is a legitimate strategy; believing you own what you are renting is not.”

Probably not a substantial claim: may not pass S2 (checkable), S4 (based).

Drift is easier to prevent than to detect.

“Drift is easier to prevent than to detect.”

Probably not a substantial claim: may not pass S2 (checkable), S4 (based). May combine two claims.

Borrowed competence is invisible when things are working, which is why it is usually discovered at the worst possible moment.

“Borrowed competence is invisible when things are working, which is why it is usually discovered at the worst possible moment.”

Probably not a substantial claim: may not pass S2 (checkable).

Founders lacking HR expertise need questions whose answers they can evaluate without it.

“Founders lacking HR expertise need questions whose answers they can evaluate without it.”

Probably not a substantial claim: may not pass S2 (checkable), S4 (based).

Question 4 is the load-bearing item, because it is the only one that distinguishes deliberate renting from borrowed competence.

“Question 4 is the load-bearing item, because it is the only one that distinguishes deliberate renting from borrowed competence.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

Every selection tool has false negatives.

“Every selection tool has false negatives.”

Probably not a substantial claim: may not pass S4 (based).

A capability held in reserve decays.

“A capability held in reserve decays, so whoever owns hiring should run one or two searches a year end to end, unaided, to keep the standard current.”

Classified as probably not a substantial claim.

The difference between Sara's and Maya's firms is that Sara knew she was renting capability and priced the difference at about two hours a month.

“The difference is that Sara knew she was renting it, and priced the difference at about two hours a month.”

• AssumptionRests on an assumption: Sara's firm implemented four inexpensive practices and achieved the described outcomes.

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing), S4 (based). May combine two claims.

Interview scheduling belongs in automate once volume makes it burdensome, and nothing is lost.

“Interview scheduling belongs there once volume makes it burdensome, and nothing is lost.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

Both activities rise as growth increases strain, and that movement is expected.

“Both rise as growth increases strain, and that movement is expected.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

The most reliable trigger for review is headcount doubling, because it is the point at which strain rises fastest and the point at which the founder is least likely to notice that it has.

“The most reliable trigger is headcount doubling, because it is the point at which strain rises fastest and the point at which the founder is least likely to notice that it has.”

Probably not a substantial claim: may not pass S2 (checkable), S4 (based).

Maya will do this while making the highest-stakes people decisions of the company's life.

“and she will do it while making the highest-stakes people decisions of the company's life.”

Probably not a substantial claim: may not pass S2 (checkable).

A boundary drawn thoughtfully at 15 people and never revisited will be wrong at 80 and dangerous at 200.

“a boundary drawn thoughtfully at 15 people and never revisited will be wrong at 80 and dangerous at 200.”

Probably not a substantial claim: may not pass S2 (checkable), S4 (based).

The boundary moves in one direction unless someone pushes back.

“the boundary moves in one direction unless someone pushes back.”

Probably not a substantial claim: may not pass S4 (based).

The danger zone sits where operational pressure and learning value are both greatest.

“It sits where operational pressure and learning value are both greatest”

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing), S4 (based).

The four tools are easier to adopt in sequence than all at once.

“The four tools above are easier to adopt in sequence than all at once.”

Probably not a substantial claim: may not pass S2 (checkable), S4 (based).

The most common failure is not over-automating but setting the boundary once and never revisiting it.

“The most common failure is not over-automating but setting the boundary once and never revisiting it.”

Probably not a substantial claim: may not pass S2 (checkable). May combine two claims.

Borrowed competence supplies neither a baseline nor an internal standard, which is why it must be prevented rather than corrected.

“Borrowed competence supplies neither, which is why it must be prevented rather than corrected.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

The seven questions in Table 5 are not presented as a validated instrument, and there is no threshold score.

“We do not present them as a validated instrument, and there is no threshold score.”

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing), S4 (based).

Vertical movement is not the problem.

“Vertical movement is not the problem.”

Probably not a substantial claim: may not pass S2 (checkable), S4 (based).

Capability can be rented.

“Capability can be rented.”

Probably not a substantial claim: may not pass S4 (based). May combine two claims.

Growth increases capability strain across the matrix, pushing every activity toward the top and creating steady pressure to move items into automate.

“Growth increases capability strain across the matrix, pushing every activity toward the top and creating steady pressure to move items into automate.”

Probably not a substantial claim: may not pass S2 (checkable).

The seven items in Table 7 do that work, because the shape of an answer is informative even when its technical content is not.

“The seven items in Table 7 do that work, because the shape of an answer is informative even when its technical content is not.”

Probably not a substantial claim: may not pass S3 (load-bearing).

The items respond directly to the finding that vendors' fairness claims are often unsubstantiated.

“In addition, the items respond directly to the finding that vendors' fairness claims are often unsubstantiated [31].”

• cites [31] M Raghavan, 2020, Mitigating Bias in Algorithmic Hiring: Evaluating Claims and Practices, Proceedings of the 2020 Conference on Fairness, Accountability, and Transparency sourceNot checked — no full text of the source

Probably not a substantial claim: may not pass S2 (checkable), S4 (based).

Maya will spend 18 months and considerable expense constructing the human judgment she believed she already possessed.

“she will spend 18 months and considerable expense constructing the human judgment she believed she already possessed”

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing), S4 (based).

If you cannot explain a criterion to a new manager in two minutes, automating it will lock in something nobody can later examine or defend.

“If you cannot explain it to a new manager in two minutes, automating it will lock in something nobody can later examine or defend.”

Probably not a substantial claim: may not pass S2 (checkable).

This single practice separates learning from processing.

“This single practice separates learning from processing.”

Probably not a substantial claim: may not pass S2 (checkable), S4 (based).

Founders who cannot state what their process selects for cannot tell whether it is still selecting for what makes the firm distinctive.

“Founders who cannot state what their process selects for cannot tell whether it is still selecting for what makes the firm distinctive.”

Probably not a substantial claim: may not pass S2 (checkable).

The automation boundary is a strategic decision rather than an operational one.

“the automation boundary is a strategic decision rather than an operational one”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

After stage 4, the cost rises with every hire, and after stage 7, it is paid all at once.

“After stage 4, the cost rises with every hire, and after stage 7, it is paid all at once.”

Probably not a substantial claim: may not pass S2 (checkable), S4 (based).

The cost here is not judgment but what judgment was supposed to protect.

“The cost here is not judgment but what judgment was supposed to protect.”