hilighter Logohilighter Logo
EN
Get StartedSign In

ArticleFuture of work

AI Hiring Trust Is Not an Application Fraud Rate

Stepstone says 13% of German recruiters identified AI-assisted deception and 28% suspected it. That is 41% of recruiters, not of applications.

Christian Kosiek•Published •18 min read

A woman in a workshop adjusts a bicycle wheel while a man stands behind her holding a clipboard
Illustrative image supplied with this article: a person works on a bicycle in a workshop while another person watches with a clipboard. It is not a named candidate, a Stepstone respondent, a CIPD employer, or a hilighter screen. The file was not labelled as a photograph of a real person or as AI-generated. The scene resembles the hypothetical workshop later in this article. It is not a reported assessment, and it does not show the 13% or 28% figures.

A polished application is easy to produce. It is a weak proof of the work. On 16 November 2026 The Stepstone Group published a Hiring Trends Update from recruiters and candidates in Germany. The headline says four in ten recruiters (41%) have identified or suspected AI-assisted deception in recruitment over the past twelve months. The same release splits that 41%. 13% report confirmed cases. A further 28% suspect such activity. The two groups are separate. Adding them does not mean that 41% of applications are fraudulent. It means 41% of the recruiters in the survey say they identified or suspected at least one case in a year.1

The samples are in the release. The Stepstone Group surveyed 7,652 candidates and 1,227 recruiters in Germany between 24 September and 13 October 2026. The candidate survey is representative of Germany's working population by age, gender and education, and it is weighted. The recruiter data is unweighted. The release does not say how many applications those recruiters reviewed, how many reports were checked against records, or whether the same behaviour was more common before generative AI. A suspicion is not a confirmed case. A confirmed case, here, is what a recruiter reports as confirmed. It is not a court finding, and it is not a count of people hired on a false claim.1

Confirmed and suspected use different bases

The forms of deception are reported only among recruiters who say they identified or suspected something. The release says those findings rest on 505 recruiters. The chart base is 165 for self-reported detected attempts and 340 for suspected attempts. Multiple answers were possible, so the percentages in a column do not add to 100. A percentage in the detected column is not a percentage of all 1,227 recruiters.1

Form the release names Detected (n=165) Suspected (n=340)
Exaggerations in CVs 55% 71%
Fake or fabricated references 43% 23%
AI-generated answers in live interviews 33% 30%
Hidden instructions in documents (prompt injection) 26% 19%
A different person at the interview 22% 14%
Deepfakes in video interviews 8% 4%

The 22% line is easy to inflate. Among recruiters who report confirmed cases, 22% say someone other than the applicant attended an interview. That is not 22% of interviews in Germany, and it is not 22% of the full recruiter sample. The 8% deepfake line has the same limit. The release also says recruiters report hidden instructions in application documents, which it calls prompt injections, designed to influence automated selection. It does not describe how those instructions were written, and this article does not either.1

Dr Christina Langer, Senior Economist at The Stepstone Group, says in the release that AI can do more than optimise applications. It creates new opportunities to misrepresent qualifications or to manipulate recruitment processes. She also says a personal interview is about getting to know the person behind the application, and that an honest answer can reveal more about ability than a polished, generic response. Those sentences are her comments. They are not a measured rate of bad hires.1

Using a tool is not the same as exaggerating

On the candidate side, the release says around half of candidates use AI when applying for jobs, and that one in six admit using it to exaggerate their skills or experience. "Around half" is the release's wording. It does not print 50%. "One in six" is not printed as 16.7%, and it is not the recruiters' 13%. A candidate can use a tool to clarify true experience. A smaller share, on their own account, use it to overstate that experience. The survey does not show how often recruiters detect the second group, or how often they mistake the first group for the second.1

The same release says 59% of candidates prefer a recruitment process without AI, and 61% would not apply for a job if AI made the final hiring decision. Those are preferences. They are not a test of any hiring product, and they are not the 73% of recruiters who say they have become more sceptical when assessing applications. The release does not print a separate base for the 73%, the 59% or the 61%. The recruiter sample is 1,227 and the candidate sample is 7,652.1

A UK survey from the spring is a different object. The CIPD Resourcing and Talent Planning report, announced on 22 September 2026, is already discussed in why an application flood hides real skills. That article keeps the September figures in their own sample, including employers who reported more unsuitable candidates. Do not paste the German 13% onto that UK chart.

Suspicion has a cost, and so does a fluent claim

If every polished page is treated as deceit, employers can fall back on familiar names, elite credentials and personal referrals. That response can shut out career changers, people returning from care, candidates with disabilities, and people whose practical competence is stronger than their prose. The opposite error is to treat an AI-written statement as proof that someone can do the job. A person who can generate a project plan may still be unable to negotiate a supplier dispute, diagnose a complaint or spot a dangerous mistake.

The useful question is narrower. Can this person exercise the judgment the role requires, with the tools the role actually permits? Knowing how to use an assistant can be a workplace skill. Knowing when the output is wrong, what must stay confidential, and why a decision was made matters more than speed of text. That is an editorial judgment, not a Stepstone result. A related limit, after someone is hired, is who teaches judgment after the first draft.

A short work sample is not a surveillance programme

For a small employer, the next step need not be an assessment platform. Name two or three decisions that make someone effective in the role. Publish what AI assistance is allowed at each stage. Ask for a brief, job-relevant exercise, paid where that is practical, with the time limit and the scoring points written down. Judge the result against those points, not against prose style. Then ask the candidate what they checked, what they changed and what they would not trust.

Consider a hypothetical nonprofit hiring a grants coordinator. Instead of an essay about motivation, give a fictional project brief, an incomplete budget and a funder rule. Ask which information is missing, which questions come first, and why a budget line moves. If the job would allow an assistant, allow it in the exercise, and ask for the check. A repair workshop might ask a technician to inspect a safe, deliberately misadjusted bicycle part and explain the sequence. A service team might use an anonymised, simulated complaint with conflicting details. These scenes are illustrations. Stepstone did not run them. The image with this article is not one of those exercises.

The exercise should be short enough to avoid free client work, and plain enough to measure the task rather than the employer's private jargon. Unpaid samples weigh more heavily on people with little spare time. A shared rubric reduces arbitrary taste. Someone still has to check that the rubric is relevant.

A worse answer to fluent applications is an AI detector, continuous video monitoring, or an invasive identity check as the default. Detectors can mark honest assistance as cheating. Video scoring can create access barriers. Collecting sensitive data needs a legitimate reason, a minimum of what is kept, and a way for the candidate to challenge an adverse decision. That last sentence is editorial practice. This article does not interpret GDPR Article 22, because the official regulation text was not opened for this review. Data-protection duties already apply to candidate information.

A structured interview often shows more than a suspicion score. Ask every applicant comparable questions about a concrete decision: assumptions, alternatives and trade-offs. Follow up on one detail from the work sample. A candidate who used a tool legitimately should be able to explain the work and accept responsibility for its accuracy. Check employment and qualifications that matter to the decision, through proportionate channels and with consent. Say which AI uses are allowed. Unstated rules punish careful applicants and leave room for the others.

A UK expectation is not this German survey

The dates overlap, which makes a false merger tempting. The CIPD Autumn 2026 Labour Market Outlook, by J. Cockett and A. Ghazaryan, is a different survey in a different country. YouGov asked 2,051 senior HR professionals and decision-makers in the UK between 17 September and 11 October 2026. The figures are weighted to UK employment by organisation size, sector and industry. Do not add them to the Stepstone counts.2

More than four in five employers (83%) have adopted AI. Figure 10, on a base of all 2,051 employers, splits that use. 40% are at the ad hoc stage, experimenting with off-the-shelf tools and with no formal strategy or dedicated integration. The report's examples of that stage are AI assistants such as ChatGPT and data-analysis tools such as PowerBI. 35% are integrating AI into specific functions. 8% say it is deeply integrated into operations and strategy. 14% report no AI use that they are aware of. 3% do not know. The 40% is the largest single stage. It is not a statement that 40% have no AI at all. Those five shares add to 100.2

Private-sector SMEs are behind that average. 29% of them report no AI use, against 6% of large private-sector employers and 5% of public-sector employers. Deep integration is most common among large private-sector employers (15%). The bases are 935 private-sector SMEs with fewer than 250 employees, 515 large private-sector employers, and 427 public-sector employers.2

Headcount expectations are a second question, and the base changes the number. Among all employers, 21% believe AI will reduce headcount over the next 12 months, 61% believe it will make no difference, and 4% believe it will increase headcount. "Do not know" answers are in the base and are not shown, so 21, 61 and 4 do not add to 100. Among employers where AI is used (autumn 2026, n=1,641), 25% now expect a reduction, up from 21% in autumn 2025 (n=1,529). In that AI-using group, 59% still expect no impact, down from 61%, and 5% expect an increase, down from 8%. The report says the routing of these questions changed between surveys, which may slightly affect comparability. Do not describe this year's all-employer 21% as if it were the AI-using 25%. Do not describe last year's AI-using 21% as if it were this year's all-employer figure. They share a numeral and not a base.2

Where AI is deeply integrated, 44% expect a headcount reduction and 17% expect an increase. The base for that group is 146 employers, excluding people who do not know whether AI is used. At the ad hoc stage (n=851), the report's chart shows 15% expecting a reduction. Where AI is integrated in some functions (n=644), 32% expect a reduction. The report says the deeper-integration pattern can look like a story of significant displacement, and it names recent job-loss announcements at Microsoft and Barclays. It does not print those companies' figures. It says the pattern is more likely to reflect earlier CIPD research: organisations further along in adoption are more likely to be redesigning jobs, planning reskilling and building skills intelligence. An expectation of a headcount change may therefore reflect workforce planning as much as an anticipated cut. This article does not treat 44% as jobs already lost. What one bank's published figures do and do not prove is a separate piece: what the Barclays numbers do not prove.2

The same chapter separates size. Excluding employers who do not know whether AI is used, 12% of private-sector SMEs (n=911) believe headcount will fall because of AI, and 77% believe it will stay the same. Among large private-sector employers (n=504), 33% expect a reduction and 7% expect an increase. In the public sector (n=413), 21% expect a decrease. These are still expectations. They are not the German 13%.2

Job performance is a third question, on the AI-using base of 1,641. 62% believe performance has improved, 52% by a little and 10% by a lot. 28% believe it has made no difference. 3% say it worsened a little, and 0% say it worsened a lot. 7% do not know. Those shares add to 100. The body says that where AI is deeply integrated (n=146), nearly 35% believe performance improved a lot, against 4% at the ad hoc stage (n=851) and 11% where it is integrated in some functions (n=644). The key points print 35% for that deep-integration line. "Nearly 35%" is the body's wording. A performance report is not a headcount result, and neither number is a German fraud rate.2

A third CIPD figure belongs to September, not to this autumn report. The 22 September press release says 48% of employers had increased their efforts to develop talent in-house over the previous year. The notes say the sample was 1,014 employers with responsibility for HR decisions whose role includes resourcing and talent planning. Fieldwork was online, by YouGov, from 2 April to 2 May 2026, weighted by business size and sector. The notes also say the 2024 survey targeted senior HR or management responsibility, so comparisons with 2024 need that difference in view. The 48% is not a share of the 2,051 autumn respondents, and it is not a finding that external hiring should stop. When outside applications reveal little, evidence about current employees becomes more useful. Tenure is still not readiness. The September report's other hiring figures stay in the application-flood article.3

Five checks before the next vacancy

The list below is an editorial checklist. Stepstone and the CIPD did not publish it as a method, and this article does not treat it as a research result.

  1. Name the two or three decisions that make someone effective in the role.
  2. Publish which AI assistance is allowed at each stage, and apply that rule to every candidate.
  3. Use a short, job-relevant work sample with a written rubric. Pay for it where that is practical.
  4. Ask the candidate to explain the choices, and verify only claims that materially affect the decision.
  5. Review who is rejected, and whether the process disadvantages a group or a career path.

The point is not to catch people for using a tool. Polished language is no longer scarce. A hiring process that can show the work, keep access fair and leave a person accountable for the decision is a response to weak documents. It is not a claim that any of these surveys measured that process.

EU rules draw a line around automated selection

This is not legal advice. For an organisation hiring in the European Union, the frame is more specific than a general warning about AI in HR.

Annex III, point 4, is headed "Employment, workers' management and access to self-employment." Point 4(a) lists AI systems intended for the recruitment or selection of natural persons, in particular to place targeted job advertisements, to analyse and filter job applications, and to evaluate candidates. A short exercise that a person scores against a written rubric is not, by that description alone, such a system. Classification depends on the intended purpose and on the legal definition of an AI system. The Commission service desk says Annex III rules apply from 2 December 2027. The service desk says its summary is not legally binding, and that the text it displays follows the EUR-Lex consolidated version as at 27 July 2026. This review opened the service desk. EUR-Lex itself returned a network challenge and was not opened.4

Article 5(1)(f) prohibits placing on the market, putting into service for this specific purpose, or using AI systems to infer emotions of a natural person in the workplace and in education institutions, except where the system is intended for medical or safety reasons. The service desk says prohibited practices are enforceable from 2 August 2026. A tool that scores tone of voice, or enthusiasm, from a video can raise that question. A structured work sample does not, by the rubric alone.4

Article 26(7) says that before putting a high-risk system into service or using it at the workplace, deployers who are employers shall inform workers' representatives and the affected workers. That duty is about workers and a system that is high-risk. It is not a finding that every work sample is high-risk, and it does not by itself say how to inform an external candidate. Article 26(11) is the closer line for people who are not yet on the staff: deployers of high-risk systems referred to in Annex III that make decisions, or assist in making decisions, related to natural persons shall inform those persons that they are subject to the system. Both paragraphs apply when the system is in that high-risk class. This article does not decide whether any product is in that class.4

A closer look at oversight when software influences people at work is in who oversees the algorithmic manager.

How hilighter can help

hilighter does not screen external applicants, detect AI-written text, score interviews or make hiring decisions. It does not estimate a fraud rate, and it does not decide whether a hiring tool is high-risk under the AI Act.

When applications stop showing who can do the work, the practical question inside the organisation is which skills the role requires and what current people can already demonstrate. Teams can map skills by role, compare people with those requirements in skills and talent insights, record a check or a correction in skills evidence and recognition, and set personalized growth paths. A manager's review of that evidence sits in manager assessments. The AI job impact view is indicative. It is not a Stepstone finding, and it is not a prediction that a role will disappear. The Free plan covers up to three job roles. The number of employees is not capped.

Sources

  1. The Stepstone Group, AI-assisted deception in job applications: four in ten recruiters have detected or suspect its use, 16 November 2026. Hiring Trends Update, Germany. Fieldwork 24 September to 13 October 2026: 7,652 candidates, weighted and representative by age, gender and education; 1,227 recruiters, unweighted. 13% of recruiters report confirmed AI-assisted deception in the past 12 months; a further 28% suspect it; together 41% of recruiters, not of applications. Among detected cases (n=165) and suspected cases (n=340), multiple answers: CV exaggerations 55% and 71%; fabricated references 43% and 23%; AI-generated answers in live interviews 33% and 30%; hidden instructions (prompt injection) 26% and 19%; a different person at the interview 22% and 14%; deepfakes in video interviews 8% and 4%. The release says the form-of-deception findings rest on 505 recruiters. Around half of candidates use AI when applying; one in six admit using it to exaggerate skills or experience. 59% prefer a process without AI; 61% would not apply if AI made the final decision; 73% of recruiters say they are more sceptical. No separate base is printed for those three lines. Comments from Dr Christina Langer are her remarks in the release, not a rate of bad hires.
  2. Cockett, J. and Ghazaryan, A., CIPD, Labour Market Outlook: Autumn 2026, London: Chartered Institute of Personnel and Development. YouGov, 2,051 senior HR professionals and decision-makers in the UK, 17 September to 11 October 2026, weighted to UK employment by organisation size, sector and industry. Figure 10, all employers: 40% ad hoc, 35% integrated in some functions, 8% deeply integrated, 14% no AI use they are aware of, 3% do not know (83% adopted). Figure 11: private-sector SMEs (n=935) 29% no AI use; large private (n=515) and public (n=427) no-use rates 6% and 5%; deep integration 15% among large private employers. Figure 13, all employers (n=2,051): 21% expect AI to reduce headcount in the next 12 months, 61% no difference, 4% an increase; do-not-know answers are in the base and not shown. Figure 12, employers where AI is used: autumn 2026 n=1,641, 25% expect a reduction, 59% no impact, 5% an increase; autumn 2025 n=1,529, 21%, 61% and 8%. The report says question routing changed and may slightly affect comparability. Figure 14, excluding those who do not know if AI is used: deeply integrated n=146, 44% expect a reduction and 17% an increase; integrated in some functions n=644, 32% expect a reduction; ad hoc n=851, 15% expect a reduction. The report names Microsoft and Barclays announcements and does not print their figures. Figure 15: private-sector SMEs n=911, 12% expect a reduction and 77% no difference; large private n=504, 33% and 7%; public n=413, 21% expect a decrease. Figure 16, AI users n=1,641: performance improved a lot 10%, a little 52%, no difference 28%, worsened a little 3%, worsened a lot 0%, do not know 7%. Figure 17: nearly 35% of the deeply integrated group (n=146) say performance improved a lot, against 4% ad hoc (n=851) and 11% integrated in some functions (n=644). The key points print 35% for that line.
  3. CIPD, Almost half of employers are upskilling their workers in the wake of labour market and legislative challenges, press release, 22 September 2026. 48% of employers increased efforts to develop talent in-house over the previous year. Notes: YouGov, 1,014 employers with HR decision-making responsibility whose role includes resourcing and talent planning, 2 April to 2 May 2026, weighted by business size and sector. The 2024 survey targeted senior HR or management responsibility. This is not the Autumn 2026 Labour Market Outlook. Other figures from this survey are discussed in the 28 October article on this site and are not repeated here.
  4. European Commission, AI Act Service Desk, Annex III, point 4(a); Article 5, point (1)(f); Article 26, paragraphs 7 and 11; When does enforcement start?. The service desk says the displayed text follows the EUR-Lex consolidated version as at 27 July 2026, and that its summary is not legally binding. Annex III rules from 2 December 2027. Prohibited practices enforceable from 2 August 2026. EUR-Lex was not opened. Not a decision about any product. Not legal advice. GDPR Article 22 was not opened.

Sources last reviewed: 18 November 2026. The 41% figure is a share of German recruiters, not of applications. Detected percentages use n=165 and suspected percentages use n=340. "Around half" and "one in six" are the release's words and are not converted into other percentages. The CIPD 21% for all UK employers is not swapped with the 25% for AI-using employers, or with the 21% those AI-using employers reported in autumn 2025. The September 48% is a third sample. The nonprofit, the bicycle workshop and the five checks are editorial illustrations. EUR-Lex and GDPR Article 22 were not opened. Neither an AI-detector score nor the absence of Google penalties can be guaranteed.