The AI Application Flood Is Hiding Real Skills

A small employer can now receive a beautifully written application from someone who has never done the work. The same tools can help a capable candidate, unsure of written English, explain a career that a stiff form would have buried. Both documents can look equally polished. Someone still has to decide who can do the job.

That is the hiring problem created by cheap, abundant text. Generative AI makes it easier to apply. It does not automatically make ability easier to see. The useful response is not an arms race between automatic CV writers and automatic CV detectors. It is a hiring process that collects a small amount of comparable evidence about the work, and keeps a person responsible for the decision.

More applications are not more information

In its Resourcing and Talent Planning report of September 2026, the Chartered Institute of Personnel and Development (CIPD) found that 58% of employers reported an increase in unsuitable candidates over the previous year, while 33% reported an increase in suitable ones. Larger organisations were more likely than small and medium-sized employers to report the rise in unsuitable applicants (61% against 52%).¹

At the same time, 53% of organisations that had tried to recruit said attracting candidates was difficult, especially for skilled and senior roles. 18% also reported difficulty attracting candidates for lower-skilled roles. An employer can be flooded with applications and still be short of people who meet the actual need. These are not contradictory findings.

The survey was run online by YouGov from 2 April to 2 May 2026. 1,014 UK-based senior HR and management professionals with responsibility for resourcing and talent planning responded. The sample excludes sole traders and is weighted to reflect UK employment by organisation size and sector. This year the CIPD narrowed the sample to people whose role includes resourcing, so comparisons with the 2024 survey need care. The figures describe what employers reported. They are not an experiment that isolates what generative AI caused. Wages, advertising, economic uncertainty and a mismatch between job requirements and available experience can move the same numbers.¹

The report, produced with the recruitment firm Omni RMS, does treat candidate use of generative AI as part of the pressure on selection. Among organisations that attempted to recruit, 27% said candidates mostly or always made excessive use of AI in CVs and applications (Figure 34, base 791). Separately, 35% of employers said they try to monitor that use (42% of large organisations; 11% were unsure). Of those monitoring, 75% said they reject candidates for generative AI use at least some of the time. The rejection question has a chart base of 280. It is not a rejection rate for every applicant, or for every employer. About 21% of organisations said they had changed assessment methods to reduce candidate use of AI, and 10% did not know.¹

Treating the presence of AI assistance as proof of dishonesty is a real risk. One candidate may use a tool to translate a true account of their work, fix spelling or organise a complicated career history. Another may invent achievements. The tool used to prepare the page does not, by itself, distinguish the two. Screening for polished phrasing, or punishing it, can miss the point in opposite directions.

A US staffing survey says verification got harder

A separate survey, commissioned by the staffing firm Robert Half and published on 10 March 2026, points the same way without proving a single cause. An independent firm surveyed more than 2,000 US hiring managers in November 2025. 67% of HR leaders said reviewing AI-generated applications had slowed hiring, and 20% reported delays of more than two weeks. 65% of hiring managers said a surge in applications, many of them enhanced or generated by AI, had made skills harder to verify. 84% of HR leaders said their teams were carrying heavier workloads.²

The same release says employers have added steps that can stretch timelines further: 42% spend more time reviewing applications, 38% hold more interviews per candidate, and 32% have rewritten job descriptions to discourage generic AI-generated answers. Robert Half also notes that not every AI-assisted application is inaccurate. Many people use the tools to improve clarity or grammar. These are self-reported perceptions from a commercial recruitment provider, not a measured change in time-to-hire from administrative records.

The hiring signal was already getting weaker

The distinction matters because recruitment was already becoming a noisier market before generative AI. Revelio Labs’ US AI Labor Market Tracker for September 2026, released on 1 October 2026, finds that job postings per observed external hire have been rising for six years. The series links postings and hires at the same company, in the same state, at the same seniority and in the same role, counting hires from the posting month through the following two months. It is a trailing 12-month ratio for matched US Russell 3000 firms, not a literal vacancy-fill rate. A posting can match zero, one or several hires, and some postings or hires cannot be matched at all.³⁴

The same edition is a useful warning against dramatic headlines. Mentions of recruiter ghosting in reviews of interviews that did not lead to an offer are back near their October 2022 baseline: about 1% below it, reversing a July reading of 3% above. The page also says those mentions had risen by as much as 9% over the previous two years before falling back. This is an index of a mention share, smoothed over 12 months. A plotted value of 10% means the share of reviews using ghosting language is 10% higher than in October 2022. It does not mean that 10% of candidates were ignored.³⁴

Revelio’s own reading is that these matching problems predate ChatGPT, so they cannot be blamed entirely on generative AI. AI may still add noise, by making it easy to send more polished applications and weakening the signals employers use. That is a plausible mechanism, not a measured cause. Complaints about poor communication are a separate series, and the September page says they are rising even as complaints about slow processes fall.

None of this makes AI-written applications harmless. It means the evidence for a broader failure of hiring signals is stronger than the evidence for one technology as the sole cause. Posting a vacancy has become a weaker sign that a hire will follow. A polished application has become a weaker sign of independent ability.

For a five-person charity hiring a programme coordinator, a fluent cover letter is not the same as being able to prioritise beneficiaries, protect confidential information or resolve a last-minute scheduling conflict. For a small accounting firm, a list of software names is not the same as spotting an inconsistent invoice and explaining the next check. Those are illustrative scenarios, not accounts of particular employers. Their point is that the relevant skill can be observed in a short exercise that resembles the job. The same gap, between performing a task and having mastered the work behind it, is why AI can let apprentices stretch beyond their training without that stretch counting as mastery.

Structured questions can help. Opacity can undo the gain

There is evidence that a more consistent way of collecting information can improve hiring decisions, and it is narrower than the marketing language around "AI interviews". In a working paper first posted on 30 July 2026, Brian Jabarian and Luca Henkel report a field experiment with the recruitment firm PSG Global Solutions, part of Teleperformance. The setting is entry-level customer-service hiring in the Philippines, for US and European clients. The paper covers 70,884 applications. Applicants who passed an initial screen were assigned to a human interviewer, an AI voice agent, or a choice between the two. In every case, human recruiters reviewed the interview and made the hiring decision.⁵

In the assigned-interview comparison, 8.70% of applicants in the human-interviewer group received an offer (1,179 of 13,557), against 9.73% in the AI-interviewer group (3,904 of 40,103). That is 1.03 percentage points, which the authors describe as a 12% higher likelihood. They also report higher job starts and retention, and, for a subset of hires, no statistically significant or economically meaningful difference in average handling time, customer satisfaction or quality-assurance scores. They attribute part of the offer result to more structured, consistent interviews that still responded to the person. Recruiters knew whether an AI agent or a person had conducted the interview.⁵

This is a preprint, not a settled consensus, and it is one job family in one labour market. It is not evidence that software should make the final employment decision. It does not automatically transfer to nurses, school leaders, social workers or every small business. Language access, willingness to speak to a voice agent, the quality of the questions and the risk of unequal treatment all matter.

Candidate trust is a second constraint. In a Greenhouse survey published on 1 May 2026, 63% of US active job seekers said they had been interviewed by an AI, up 13 percentage points in six months. 38% said they had already left a process because it included an AI interview, and another 12% said they would. 70% said they had not been clearly told in advance that AI would evaluate them, and 21% found out only once the interview started. The survey included 2,950 people in the United States, the United Kingdom, Germany, Australia and Ireland. Unless the release says otherwise, the headline figures are the 1,200 US respondents. They describe perceptions, not the fairness or accuracy of any particular system. Greenhouse sells hiring software, so this is vendor research.⁶

The apparent paradox is useful. Standardised questions can improve what an employer learns. Hidden automation can destroy trust in how that learning happens. The practical question is not whether to replace every interviewer with a machine. It is whether each candidate gets a fair chance to show the same relevant abilities, understands the process and has a route to a human decision.

EU rules already draw a line around automated selection

For organisations hiring in the European Union, the legal frame is more specific than a general warning about "AI in HR". Under Annex III of the EU AI Act, AI systems intended for the recruitment or selection of people are listed as high-risk. The list covers, in particular, systems used to place targeted job advertisements, to analyse and filter applications, and to evaluate candidates. A short exercise scored by a person against a written rubric is not, by that description alone, such a system. Classification depends on what the technology actually does and on the legal definition of an AI system.⁷

The Commission’s AI Act Service Desk says the rules for Annex III high-risk systems apply from 2 December 2027. It also says that prohibited AI practices are enforceable from 2 August 2026. One of those prohibitions, as the same service desk explains it, covers AI systems used to infer emotions of a person in the workplace or in education institutions, except where the system is intended for medical or safety reasons.⁸⁹ A recruitment tool that scores tone of voice or "enthusiasm" from a video can therefore raise a different question from a structured work sample. This article is not legal advice. Data-protection duties already apply to candidate information, whatever the AI Act timetable says. For a closer look at human oversight when software influences people at work, see who oversees the algorithmic manager.

A better process starts with proof of the work

The simplest improvement is often low-tech. Before advertising a role, name three or four capabilities that actually predict success in it. Then design a short, proportionate way to observe them. A customer-facing employee might explain how they would handle a complaint. A project assistant might order a small set of competing requests. An analyst might look at a flawed table and say what cannot safely be concluded. The candidate should know what is being assessed. The exercise should resemble the work, not a puzzle that rewards interview coaching.

There are trade-offs. Unpaid work samples can burden people with caring responsibilities, disabilities or little spare time. They should be brief, accessible and genuinely diagnostic, not a request for free client work. A shared rubric reduces arbitrary impressions, but someone still has to check that the rubric itself is relevant and fair. Keep a person accountable for the outcome, and offer a reasonable alternative when the format creates a barrier.

Organisations also need a clear position on AI use that matches the job. If the role will use AI to draft routine messages, forbidding every candidate to use it may test an obsolete version of the work. A more revealing exercise can allow assistance, then ask the person to find a factual error, explain a judgment and say what they would verify themselves. Where independent reasoning is essential, a second exercise without assistance may be appropriate. The conditions should be explicit and the same for everyone.

This matters for people changing careers. A retail supervisor may have real conflict-resolution and scheduling skill without the job title a nonprofit expects. A parent returning to work may have recent experience organising complex tasks and an unconventional CV. A worker whose former tasks have been partly automated may be able to do the new work without the expected credential. None of them should be hired on an unverified claim. None of them should be excluded only because their experience is hard to summarise in standard recruitment language.

There is a workforce consequence beyond the vacancy. The CIPD survey found that 48% of employers had increased their efforts to develop talent in-house over the previous year, and that upskilling existing employees was the most common response to recruitment difficulty. That reaction is understandable: when external applications reveal little about real capability, knowing what current employees can demonstrate becomes more valuable. Internal promotion still needs evidence. Tenure is not the same as readiness.¹

The scarce resource is credible proof

AI has made professional-sounding text abundant. It has not removed the hard parts of work: handling uncertainty, noticing a mistake, adapting to a person, protecting information or taking responsibility for a result. A process that rewards presentation over those abilities makes the labour market noisier for everyone, including candidates who really can do the job.

The lesson for a small employer is not to buy another screening algorithm first. It is to ask what the current process actually learns about a person. If the answer is mostly how well that person can describe their skills, the process is exposed to the very technology it is trying to manage. A fair, short demonstration of the work may now be worth more than the most perfectly written application.

Questions employers actually ask

Does AI in an application prove the candidate is dishonest?

No. The CIPD figures show that many UK employers are seeing what they call excessive use, and that some reject candidates for it at least some of the time. Robert Half’s own release says many candidates use AI to improve clarity or grammar, while others may embellish or invent history. The assistance itself is not the evidence. What the person can do, under conditions you have explained, is the evidence.¹²

Did generative AI cause the rise in postings per hire?

The Revelio series has been rising for six years, which includes the years before ChatGPT. The company says the trend cannot be attributed entirely to generative AI. A polished application may add noise to a matching problem that already existed. That is a hypothesis, not a measured share of the increase.³⁴

Should a small employer ban AI in every test?

Only if the job itself forbids it. If people will use AI in the role, a total ban can hide the skill that will matter: checking an output, explaining a judgment and knowing what to verify. If the role depends on unaided reasoning, say so and assess that separately. Either way, tell every candidate the rule in advance.

How hilighter can help

hilighter does not screen external applicants, detect AI-written CVs or make hiring decisions. When applications stop revealing who can do the work, the practical question inside the organisation is what your current people can already demonstrate, and which skills the role now requires.

hilighter helps teams map skills by role, compare people with those requirements in Skills & Talent Insights, record proof of capability in Skills Evidence & Recognition and set personalized growth paths. Its AI Job Impact view is indicative, not a prediction that a role will disappear. A free plan covers up to three job roles, with unlimited employees. For a closer walkthrough, see how to run a free skills matrix.

Sources

1. Hogarth, A. and McCartney, C., CIPD, Resourcing and talent planning report 2026, September 2026.
2. Robert Half, Survey: 67% of HR leaders report AI-generated applications are slowing hiring, 10 March 2026.
3. Revelio Labs, AI Labor Market Tracker: September 2026, released 1 October 2026.
4. Revelio Labs, Appendix and methodology, September 2026, released 1 October 2026.
5. Jabarian, B. and Henkel, L., Voice AI in Firms: A Natural Field Experiment on Automated Job Interviews, arXiv:2607.28222, first posted 30 July 2026 (preprint, not peer reviewed).
6. Greenhouse, 63% of Job Seekers Have Faced an AI Interview, 1 May 2026.
7. European Commission, AI Act Service Desk, Annex III (high-risk employment uses, including recruitment and selection).
8. European Commission, AI Act Service Desk, When does enforcement start? (Annex III rules from 2 December 2027; prohibited practices enforceable from 2 August 2026).
9. European Commission, AI Act Service Desk, Is emotion recognition prohibited in workplace and education settings?.

Sources last reviewed: 29 October 2026. CIPD and Robert Half figures are employer surveys, not experiments. The Jabarian and Henkel paper is a preprint about one customer-service hiring setting. Revelio’s ghosting figure is an index of review language, not a rate of ignored candidates. EU dates follow the Commission service desk and are not legal advice.