The Credential Vacuum

When Proof of Competence Becomes Cheap to Synthesize, Trust Shifts Toward What Is Harder to Fake

The following scene is illustrative, not reported. Imagine two browser windows open side by side on a technical recruiter’s monitor, each holding a candidate profile for a mid-level distributed systems engineering role.

Both resumes are immaculate. Their formatting parses cleanly through the applicant tracking system and hits every designated keyword. Both cover letters balance technical command with corporate polish, citing containerization challenges, concurrency bottlenecks and microservice decoupling in flawless phrasing. Both link to GitHub accounts with active contribution heatmaps, tidy repositories, automated test suites and thorough documentation. Under the evaluation norms of 2018, both candidates would be uncontroversial finalists.

Now suppose the requisition has drawn 1,480 submissions over four weeks, and more than 450 meet every listed technical requirement. These figures are invented for the scene. The hiring team still cannot tell which polished documents reflect the work of the people who submitted them.

That difficulty is the subject of this article. Generative software can draft an engineering memo in seconds and customize hundreds of resumes in an hour. The problem is not that AI destroys proof. It is that AI dilutes the information contained in proof. The artifact still exists, but it says less than it used to about the person behind it.

For decades the university degree served as a blunt, standardized proxy for baseline capability, diligence and compliance. In recent years, degree requirements began declining in specific sectors and jurisdictions, and many employers announced a pivot toward skills-based hiring. Removing a visible credential did not remove the need to choose among candidates, and the replacement evidence arrived just as generative tools lowered the cost of producing it.

This article argues that employers facing diluted evidence are not abandoning screening. Research on AI-assisted applications suggests that evaluators ask for costlier signals, and signals whose credibility rests on provenance, verification and institutional scarcity may gain relative value. The shift moves hiring from a visible credential, the degree, toward a distributed one: a set of informal signals spread across employers, networks and verification tools. A distributed credential is harder for outsiders to see and to plan for.

Act I: Why the Degree Gate Weakened

The retreat from the university degree accelerated in specific sectors and jurisdictions as employers and policymakers began removing degree requirements. The share of 25- to 29-year-olds holding a bachelor’s degree or higher rose from 23 percent in 1990 to 40 percent in 2022, and employers began attaching degree requirements to jobs that had not previously demanded one. A 2017 study by Accenture, Grads of Life and Harvard Business School, drawing on an analysis of more than 26 million job postings, called the pattern degree inflation. It found that in 2015, 67 percent of production supervisor postings required a bachelor’s degree, while only 16 percent of workers already in that role held one.

This inflation created what Opportunity@Work calls a “paper ceiling” for the roughly half of the U.S. workforce that developed its skills outside a bachelor’s degree. These workers acquired valuable technical proficiency through community colleges, military service, apprenticeships or on-the-job training.

The shift was especially visible in public-sector hiring, where states began formally removing degree requirements from government roles. By 2024, twenty-five U.S. states had enacted measures of this kind. A National Bureau of Economic Research study of the rollout found that degree requirements in affected listings fell by 2.5 percentage points for each additional year of policy exposure.

AI-related technical hiring showed a related pattern. A 2025 study of approximately 11 million UK job vacancies found that demand for AI roles rose 21 percent, while mentions of university requirements for AI roles fell 15 percent. AI skills carried a 23 percent wage premium in the dataset. The authors treat the link between these trends as correlative, not causal.

By these measures, the formal architecture of hiring was changing. The visible gate, standardized and institutional, was weakening in many of the job descriptions where the shift was occurring.

Act II: Why Removing It Did Not Transform Hiring

Hiring data suggest that removing the visible requirement was easier than replacing the selection process behind it. Removing a sentence from a job description costs little. Restructuring how hiring managers rank applicants is much harder.

In 2024, the Harvard Business School Project on Managing the Future of Work, in collaboration with the Burning Glass Institute, published Skills-Based Hiring: The Long Road from Pronouncements to Practice. The researchers examined hiring data across 11,300 roles at major U.S. corporations where employers had publicly committed to dropping degree requirements.

Metric or scope	Finding	Reading
Direct policy impact (within reset roles)	Non-degree hiring rose by 3.5 percentage points	The change was measurable but small.
Occupational scope (share of roles reset)	Only 3.6 percent of corporate roles underwent genuine credential resets	Few roles had actually been reset.
Economy-wide aggregate impact	An incremental hiring shift of 0.14 percent	Across roughly 77 million annual hires, about 97,000 additional non-degree hires per year.

In aggregate, the hiring changes measured across the study amounted to fewer than one in seven hundred hiring decisions. Removing the credential from the text of a job posting did not produce a comparable transformation in who was hired. The finding suggests that changing the visible requirement is easier than rebuilding the underlying selection process.

One plausible reason is that recruiters and managers still need a fast way to rank large applicant pools, which makes familiar credentials attractive as a risk-reduction shortcut. The study does not establish this mechanism directly, and it should be read as an interpretation.

The data do not support dismissing skills-first hiring. The same analysis identified a cohort of “Skills-Based Hiring Leaders,” the 37 percent of firms studied that changed their actual hiring after dropping requirements. On average these firms hired 18 percent more non-degreed workers into the affected roles, and the authors describe tangible changes to their internal hiring processes. At these firms, non-degreed hires had a two-year retention rate 10 percentage points higher than their college-educated coworkers, 58 percent against 48 percent. Across the study, workers without a degree who stepped into roles that previously required one saw an average salary increase of about 25 percent.

The evidence suggests that skills-first hiring can work when organizations invest in measuring capability directly. That investment is expensive and labor-intensive. And while employers were still building it, a separate shock was reaching the replacement evidence itself.

Act III: Why AI Damaged the Replacement Signals

The premise of the skills-based movement was to discard the proxy and evaluate the work. Instead of asking where an applicant spent four years between eighteen and twenty-two, employers would look at what the applicant could build. The resume would yield to the portfolio, the commit history, the take-home sprint and the writing sample.

That approach rests on an idea formalized by Nobel laureate Michael Spence in his 1973 paper Job Market Signaling. In Spence’s model, a signal separates applicants only when it costs different types of applicant different amounts to produce or acquire. A signal that is meaningfully more costly or difficult for lower-ability workers than for higher-ability workers can sort them: the higher-ability candidates find the investment worthwhile, the others do not, and the employer can infer something from who sends it. When those cost differences shrink, the signal carries less information.

Generative AI can shrink them. The mechanism has been tested in a study by Bo Cowgill, Pablo Hernández-Lagos and Nataliya Wright, now published in Management Science. The experiments involved more than 1,100 participants who produced professional business proposals and technical pitches, and more than 800 professional evaluators and investors who assessed them. The study reported three findings that matter here:

1. Signal compression. With access to generative models, candidate submissions became more compressed and homogenized, narrowing the visible gap between expert and non-expert work.
2. Lower screening accuracy. Evaluators’ ability to tell experts from non-experts fell by 4 to 9 percent.
3. Demand for costlier signals. Evaluators responded by increasing their demand for other, more costly signals by roughly 6 to 9 percent.

The third finding is the empirical bridge for the article: when AI lowers the information value of one signal, evaluators do not simply stop screening. They become more willing to pay for other, costlier forms of evidence. That finding offers a mechanism for understanding why hiring systems may shift weight from the artifact itself toward provenance and verification.

The effect is not uniform. It depends on whether AI helps experts and non-experts equally, and in some settings it improved screening accuracy. Among senders from non-English-speaking countries, for example, AI tools improved screening accuracy in the study, apparently because they helped experts more than nonexperts.

The portfolio has not become useless. It has become less self-authenticating. When capable coding and writing tools can produce a repository or a research dossier quickly, the existence of the artifact says less about whether the candidate could produce it unaided.

Act IV: How Employers Reacted

Employers report that the loss of signal reliability has changed how they screen. Generative tools make polished applications easy to produce, and survey data show the burden landing on hiring teams.

United States. In March 2026, staffing firm Robert Half published a survey of HR leaders. Sixty-seven percent said AI-generated applications were slowing hiring, and 65 percent said AI-enhanced resumes made skills harder to verify. Forty-two percent reported spending more time reviewing applications, and 38 percent had increased the number of interview rounds. These figures are consistent with verification inflation: more effort goes into checking claims that once required less scrutiny.

United Kingdom. The Institute of Student Employers’ 2025 early-career survey found that only 15 percent of employers had never suspected or identified candidate cheating, down from 22 percent the year before. The share reporting frequent cheating more than doubled, from 7 percent to 15 percent. The most commonly reported form of misconduct, cited by 61 percent of employers, was candidates using AI during interviews without disclosure or permission. The same research shows the limits of the concern: 49 percent of employers had no problem with some candidate AI use, particularly on CVs, cover letters and application questions. A 2026 peer-reviewed study of more than 9,000 job candidates across two consecutive cohorts also found that self-reported use of generative AI in standardized pre-hire employment assessments rose between 2024 and 2025.

When employers cannot tell an honest engineer using a language model as a productivity tool from an applicant reading generated answers, the cost and complexity of verification can rise for employers and candidates.

The extreme case. At the edge of remote hiring, the problem extends from synthetic documents to synthetic identities. In June 2025, security firm Pindrop reported a vendor estimate that about one in six applicants for remote engineering roles at its own company showed anomalies consistent with identity fraud, including deepfake video filters, synthesized voices and proxy interviewees. Because the figure comes from a vendor describing its own applicant pool in a high-security sector, it indicates direction of travel rather than a market-wide rate.

Act V: The Second-Order Problem of Automated Screening

The verification contest now runs on both sides. Candidates use AI to produce evidence, and employers use AI to screen it. LinkedIn’s 2025 Future of Recruiting report found that 37 percent of organizations were actively integrating or experimenting with generative AI in hiring, and teams using it reported saving about 20 percent of their workweek.

Automated screening adds a second layer of evidence, and that layer has to earn trust too. Asynchronous video interviews illustrate the point. A 2026 study of 626 mock-interview participants and 523 real applicants found that features of the video-interview setting could affect completion and evaluation, while standardization reduced some bias but not across all characteristics. The study shows that standardization can reduce some sources of bias without eliminating all variation in how candidates are evaluated.

Regulators have started to respond, through separate provisions. The EU AI Act classifies specified AI systems used for employment, recruitment and worker evaluation as high-risk, which brings requirements such as human oversight and data governance. A 2026 EU Digital Omnibus regulation delayed application of the high-risk provisions for Annex III systems to December 2, 2027. The Act’s Article 50 sets transparency obligations for AI systems that interact with people and for AI-generated or manipulated content. Those are distinct obligations, and Article 50 should not be read as an employment-screening disclosure regime. In the United States, New York City’s Local Law 144 requires a bias audit before an automated employment decision tool is used, public availability of audit information, and notices to candidates and employees.

Automated evaluation is not inherently regressive. A 2025 randomized field experiment, posted as an arXiv preprint, assigned about 37,000 applicants for a junior front-end developer role to one of two pipelines. The traditional pipeline used resume screening followed by human selection. The AI-assisted pipeline began with a structured, AI-led video interview before human review. Both tracks ended in the same final human interview, conducted by interviewers who did not know which pipeline a candidate had come through.

• Interview performance. Among the 35 finalists from each track, 54 percent of those from the AI-assisted pipeline passed the final human interview, compared with 34 percent from the traditional pipeline.
• Subsequent employment. Five months later, LinkedIn profiles of a subsample of top candidates showed that 23 percent of the AI-assisted group had found new jobs, compared with 18 percent of the traditional group.
• Different selection profile. The AI-assisted pipeline tended to select younger applicants with less experience and fewer advanced credentials.

The trial suggests that structured tasks and uniform rubrics can provide a different and potentially more informative signal than unstructured application text. It is a single preprint and should be read as early evidence rather than settled. When an automated applicant submits a synthetic resume to an automated recruiter running keyword matching, neither side learns much. This article’s inference is that meaningful signal is more likely when evaluation moves from unverified text toward controlled, standardized performance.

The broader response therefore has two directions: more regulation around automated assessment, and more sophisticated attempts by employers and recruiting platforms to use automation without abandoning verification. On September 29, 2026, LinkedIn announced Hiring Assistant 2, a next-generation recruiting agent. Hiring Assistant 2 brings verification, ATS insights, prescreening responses and relevant work experience together as what LinkedIn calls “richer, more trusted signals.” The rollout is scheduled to begin in November. The recruiting market is responding to synthetic abundance with increasingly AI-assisted screening, and LinkedIn’s newest product illustrates the direction. Yet even this recruiting agent surfaces verification status alongside its other signals. The platform is asking which parts of a candidate’s story can be trusted as well as whether the candidate matches the job.

Act VI: The Shadow Stack and the Return of Provenance

When no single visible credential does the work, employers still have to choose. This article uses the term Shadow Stack for the informal, distributed set of signals they lean on. The term is an analytical label, not an established category in the research literature.

The path to the Shadow Stack can be described in seven stages, offered here as a proposed analytical framework rather than an established sequence:

1. Credential inflation. The university degree becomes a default filter, even for jobs where many incumbent workers do not hold one.
2. The skills-first pivot. Employers publicly strip degree requirements, and the visible, standardized gate is removed from many postings.
3. Signal substitution. Employers seek alternative proof: GitHub profiles, project portfolios and take-home assignments.
4. Synthetic abundance. Generative AI lowers the cost of producing some forms of proof, and those signals become less informative.
5. Verification inflation. Overwhelmed by polished applications, employers add interviews, proctored tests and identity checks.
6. The provenance premium. Employers may place greater weight on signals tied to institutions, referrals and live observation.
7. The shadow credential. Informal, fragmented signals supplement or replace the degree.

The defining idea behind the Shadow Stack is that provenance becomes part of the evidence itself, alongside the artifact. The question shifts to where a claim originates, who verified it previously, and what constraints keep it from being a digital fabrication.

The signals in the stack

The right-hand column states a potential advantage. It is an inference about who each signal may favor, not a measured outcome.

Signal	What it verifies	Potential advantage
Corporate lineage	Prior institutional vetting	Favors candidates who already have a recognized employer stamp
Referral	Social endorsement	Favors candidates with strong professional networks
Live assessment	Observed performance	Reduces reliance on unverified artifacts
Elite academic brand	Selectivity and institutional reputation	Favors candidates with access to scarce institutions
Uncompensated trial	Capacity to invest time	Can disadvantage candidates with limited financial runway

Several of these signals have documented counterparts.

Referral. A 2026 Journal of Labor Economics paper models what happens when labor markets rely heavily on employee referrals. The model shows why referrals can improve matching and productivity while simultaneously disadvantaging workers with few connections. With homophilous networks, meaning networks in which people connect mostly with similar peers, the same mechanism can reinforce inequality and reduce mobility. A referral can also work as an information intermediary: an employee who knows both the candidate and the firm can transmit information about the candidate’s likely productivity.

Institutional lineage. A regression-discontinuity study of university admissions in Chile, by Bordón and Braga in Economics of Education Review, compared students who barely cleared the cutoff for the nation’s two most elite universities with those who barely missed it. In the first year in the labor market, the elite credential carried a 13 percent wage premium. For workers with six or more years of labor market experience, the premium had contracted to 4 percent. The pattern is consistent with an employer-learning mechanism: prestige matters more when employers have less direct information about individual performance. If some early-career signals have become less informative, employers may lean harder on scarce, recognizable signals such as an elite university or a tier-one employer. The article’s inference is that such signals may become more valuable when employers have less direct information about individual performance.

Verification. LinkedIn’s Hiring Assistant 2, described in Act V, is the most current example: candidate verification status placed among the core signals of a recruiting product, in a launch announcement that calls identity and skill verification critical to hiring.

Skills-first hiring does not inevitably produce a shadow gate. Organizations that build structured assessments can change who advances. The Skills-Based Hiring Leaders and the 37,000-applicant trial both illustrate that possibility. The risk lies where employers drop the visible credential without building that measurement infrastructure.

Act VII: Who Benefits from the Shadow Stack

A selection process can create socioeconomic filtering without asking about income. If completing it takes large amounts of unpaid time, candidates with greater financial flexibility can participate more easily. This article calls that effect the Runway Tax. It is a structural inference, not a measured finding about current hiring.

The inference has precedent. Research published in 2025 in Taylor & Francis Online raises concern about financial barriers to placements and unpaid work. A hiring process that grows from a resume review and two interviews into take-home assignments, psychometric profiles, technical challenges and multi-day trials resembles a compressed unpaid placement.

Consider two hypothetical candidates. One works forty hours a week in an hourly service job or cares for dependents, and cannot spend twenty unpaid hours over a weekend on a sample go-to-market plan. The other has several months of living expenses saved and can absorb three weeks of take-home sprints and multi-round loops. The process never asks for a bank balance, yet the two candidates face different costs of completing it. Whether real employers’ processes produce this pattern at scale is a question for further research.

The old credentialing system was expensive and exclusionary for many people. It did have one civic virtue: a visible admissions process, published requirements and a named credential. An ambitious outsider could see roughly what the path demanded. The path was steep, but the map was public.

A distributed credential has no equivalent map. Social capital, corporate lineage, communicative polish and the capacity to absorb uncompensated vetting time can all count, and few of them are written down. The advantage did not vanish with the degree. It may have shifted toward those who already hold the resources to produce the newer, unwritten signals.

Act VIII: The Credential Vacuum

What changed is where the evidence for competence comes from. When the degree was the main visible credential, applicants could see the requirement and plan for it. As generative tools make some artifacts of competence cheaper to produce, evaluators have been shown to ask for costlier signals, and employers are building verification into their processes. LinkedIn’s newest recruiting agent brings candidate verification status, screening responses and work evidence into one view, consistent with the shift toward evidence that recruiters can trust.

The signals most likely to gain value under these conditions are tied to provenance: who vouched for a claim, which institution selected the person, whether performance was observed directly. If this shift continues without validated assessment, candidates with recognized employers, strong referrals, elite institutional ties and the financial runway for long assessments may find the new signals easier to produce than candidates without them. Where employers invest in structured, validated assessment, the evidence reviewed here suggests that measured performance can take the place of the old gate and broaden access.

When competence becomes cheap to simulate, provenance can become a more valuable form of labor-market capital.

Sources and Methodology

This article draws on the following sources. Statistics are reported as the sources state them, and survey results are attributed to the surveyed groups.

• Bachelor’s degree attainment among 25- to 29-year-olds: NCES, Digest of Education Statistics Table 104.20 for 1990, and Condition of Education for 2022.
• Degree inflation: Accenture, Grads of Life and Harvard Business School, Dismissed by Degrees (2017), as described in Harvard Business School Working Knowledge. The posting figures describe the postings analyzed.
• Paper ceiling: Opportunity@Work. The “paper ceiling” term and the estimate of roughly half the U.S. workforce are this advocacy organization’s own.
• Degree requirements in U.S. public-sector hiring: NBER working paper.
• UK vacancies, AI roles and degree mentions: Technological Forecasting and Social Change, 2025. The authors describe the hiring evidence as correlative.
• Skills-based hiring in practice: Harvard Business School Project on Managing the Future of Work and Burning Glass Institute, 2024.
• Signaling theory: Spence, Job Market Signaling, 1973.
• Generative AI and screening accuracy: Cowgill, Hernández-Lagos and Wright, Management Science.
• Employer reaction in the United States: Robert Half, March 2026. A survey of HR leaders, so results reflect self-report.
• Employer reaction in the United Kingdom: ISE 2025 survey and its summary of trends. Misconduct figures are employer-reported.
• Candidate self-reported generative AI use in standardized pre-hire employment assessments: Robie, Wingate, Baytalskaya and Butera, International Journal of Selection and Assessment, 2026. Use figures are self-reported.
• Identity fraud in remote hiring: Pindrop, June 2025. A vendor-reported estimate drawn from one company’s applicant pool.
• Asynchronous video interviews: 2026 study of 626 mock-interview participants and 523 real applicants.
• Structured AI-assisted screening: Aka, Palikot, Ansari and Yazdani, Better Together: Quantifying the Benefits of AI-Assisted Recruitment, a 2025 preprint of a randomized field experiment. It has not been presented here as settled evidence, and two of the four authors are affiliated with micro1, the company whose AI interviewing platform was tested.
• Employer-side AI adoption: LinkedIn, Future of Recruiting 2025.
• Regulation: EU AI Act, including Article 50, as amended by the Digital Omnibus on AI, Regulation (EU) 2026/1744; New York City Local Law 144.
• Referrals: Journal of Labor Economics, 2026. A model of referral markets, so its findings are model-based.
• Elite university premium: Bordón and Braga, Economics of Education Review, 2020, a regression-discontinuity study of university admissions in Chile.
• Financial barriers to placements: Taylor & Francis Online, 2025.
• LinkedIn Hiring Assistant 2: LinkedIn announcement, September 29, 2026.

The opening scene is illustrative. “Shadow Stack” and “Runway Tax” are this article’s own labels. Their implications for who benefits are structural inferences, and the evidence reviewed here does not measure them directly.