A recruiter opens a role on Monday. By Friday, there are hundreds of applications in the ATS. For some positions, the number is already in the thousands.<\/span><\/p>\n At first glance, that looks like a good problem to have. A large talent pool should mean more choice, better competition and a better chance of finding the right person. In practice, many recruiting teams are experiencing the opposite. More applications are coming in, recruiters are spending more time reviewing them, and confidence in what those applications actually tell them is falling.<\/span><\/p>\n The numbers make the change hard to ignore. According to<\/span> Greenhouse’s 2026 recruiting benchmarks<\/span><\/a>, based on data from more than 6,000 companies and over 640 million applications, the average number of applications per job increased from 116 in 2022 to 244 in 2025. At the same time, the average number of recruiters per organization in the dataset fell by 56%, while time to fill increased by 37%.<\/span><\/p>\n There are now more applications, fewer recruiters available to process them and longer hiring cycles.<\/span><\/p>\n The problem is not simply that more people are looking for work. The application itself has changed.<\/span><\/p>\n Until recently, applying for a job required at least some effort. A candidate had to find the role, read the description, update a resume, perhaps write a cover letter, complete a form and submit the application.<\/span><\/p>\n That process created friction. Friction limited volume.<\/span><\/p>\n Now much of it can be automated. AI tools can search job boards, compare vacancies with a candidate’s profile, adapt the resume, fill application forms and submit applications. In some cases, candidates are no longer making an individual decision about every vacancy before applying.<\/span><\/p>\n This is no longer a niche scenario. A<\/span> July 2026 report from The Irish Times<\/span><\/a> describes applications becoming a partly or fully automated process in which AI systems can find vacancies and send a candidate’s CV when the system decides there is a match. The same report gives an example of a business analyst vacancy that might previously have attracted 100 to 150 applications but received more than 2,000.<\/span><\/p>\n From the candidate’s point of view, using automation can be perfectly rational. When response rates are low and one job search can last for months, increasing the number of applications feels like improving the odds.<\/span><\/p>\n For the recruiter, however, it changes what an application means.<\/span><\/p>\n Someone clicking Apply used to provide at least a small signal of intent. Today, the candidate may have carefully chosen the company and role, or an agent may have selected the vacancy as one of hundreds of acceptable matches. Both applications arrive in the same inbox.<\/span><\/p>\n The ATS cannot tell you how interested the person really is.<\/span><\/p>\n The second problem begins after the application arrives.<\/span><\/p>\n Recruiting teams facing hundreds or thousands of CVs naturally look for ways to screen them faster. Keyword filters, matching scores and AI-assisted resume review all promise to reduce the manual work.<\/span><\/p>\n Candidates know this too.<\/span><\/p>\n There is now an entire market built around helping applicants adapt resumes to job descriptions, identify missing keywords and rewrite experience in language that is more likely to match the role.<\/span><\/p>\n Used honestly, there is nothing wrong with that. A good engineer can have a badly written resume. A strong salesperson can describe their results poorly. Helping a candidate communicate genuine experience more clearly is not the problem.<\/span><\/p>\n The problem begins when the document is optimized for the screening system rather than written to describe reality.<\/span><\/p>\n For years, one common trick has been to copy keywords, or even the entire job description, into a CV and hide the text by making it white or very small. The idea is that a human recruiter will not see it, but a machine will.<\/span><\/p>\n It is important to be precise here. Simple white-text keyword stuffing does not automatically fool every modern ATS. In an<\/span> Associated Press report on AI and job search<\/span><\/a>, Greenhouse CEO Daniel Chait argued that modern systems have moved beyond the idea that one hidden keyword can guarantee success.<\/span><\/p>\n But LLM-based screening has created a more serious version of the same problem: prompt injection.<\/span><\/p>\n Instead of only hiding keywords, a candidate can add instructions intended for the model reading the CV. The instruction may be hidden in white text, tiny text or another part of the document that a person is unlikely to notice. The content can tell the model to rate the candidate highly, ignore previous instructions or treat the profile as a strong match.<\/span><\/p>\n This is not just an internet trick discussed on forums.<\/span><\/p>\n A<\/span> 2026 study of approximately 200,000 real resumes<\/span><\/a>, published for the USENIX Security Symposium 2026, found hidden prompt injections in approximately 1% of the resumes studied. The researchers also found that their prevalence had increased noticeably over the previous one to two years.<\/span><\/p>\n Another<\/span> 2026 study published in the Findings of ACL<\/span><\/a> tested prompt injection in automated resume screening. The researchers found that manipulation could improve candidate rankings in some conditions and, in certain cases, allow a lower-quality candidate to rank above a stronger one.<\/span><\/p>\n This does not mean every AI screening system can be defeated with white text. It means recruiters should be very careful about treating an AI-generated match score as objective evidence of candidate quality.<\/span><\/p>\n If the candidate knows what the system is looking for, and the system is ranking documents largely on the information inside those documents, the score may partly measure how well the application was prepared for the model.<\/span><\/p>\n That is not the same as measuring whether the person can do the job.<\/span><\/p>\n We are reaching an odd point in online recruiting.<\/span><\/p>\n On one side, a candidate uses AI to find jobs, adapt a resume and apply. On the other, the employer uses AI to screen, summarize and rank the applications.<\/span><\/p>\n The candidate may not read the full vacancy. The recruiter may not read the full resume.<\/span><\/p>\n One system writes the application. Another system reviews it.<\/span><\/p>\n Meanwhile, the recruiter still has the same responsibility: find a real person who can do the work.<\/span><\/p>\n Robert Half’s<\/span> March 2026 survey of more than 2,000 U.S. hiring managers<\/span><\/a> shows what this looks like inside hiring teams. Sixty-seven percent of HR leaders said reviewing AI-generated applications was slowing hiring, and 20% reported delays longer than two weeks. Sixty-five percent of hiring managers said AI-enhanced applications had made candidate skills harder to verify.<\/span><\/p>\n The same survey found that companies were responding by spending more time on reviews, adding interviews and changing job descriptions.<\/span><\/p>\n This is the paradox. Automation was supposed to reduce the work around hiring, but when both sides optimize against each other, the work moves somewhere else. Less time is spent writing and submitting applications, but more time is spent checking whether the application is trustworthy.<\/span><\/p>\n One reason this problem is easy to miss is that many recruiting metrics still improve.<\/span><\/p>\n Applications per vacancy go up. Cost per application goes down. The top of the funnel looks full.<\/span><\/p>\n None of those numbers tells you how long it takes to find one candidate worth interviewing.<\/span><\/p>\n A recruiting team can collect 1,000 applications and still have a sourcing problem. The volume hides the cost of finding useful signal inside the pile.<\/span><\/p>\n Gem’s<\/span> 2026 Recruiting Benchmarks Report<\/span><\/a>, based on more than 165 million applications and 1.2 million hires, shows the difference between volume and yield. Job boards and company marketing channels generated roughly 90% of applications but only around half of hires. Direct sourcing produced 11% of hires from just 2.6% of applications, while sourced candidates were nearly eight times more likely to be hired than inbound applicants.<\/span><\/p>\nCandidates no longer need to apply to every job themselves<\/b><\/h2>\n
The application can now be optimized for the system reviewing it<\/b><\/h2>\n
Hiring is becoming a conversation between two machines<\/b><\/h2>\n
The dashboard can still look healthy<\/b><\/h2>\n