Can You Trust an AI Resume Screening Score? What the Dutch Regulator Found, and a Test You Can Run in 10 Minutes

What did the Dutch regulator find about AI resume screening?

It found weak transparency, and scores that can depend on irrelevant details.

In March 2026 the Autoriteit Persoonsgegevens (AP), the Dutch data protection authority, released the sixth edition of its Rapportage AI & Algoritmes Nederland, with a cover dated February 2026. Chapter 3, which starts on p. 35, follows recruitment step by step, from the vacancy text to the final choice.

Many applicant tracking systems now offer CV screening that analyses CVs for fit with the vacancy, the AP notes. The employer then sees the algorithm's selection, often with advice on whether to proceed, rather than every applicant. Such a system can reject candidates "without a person having looked" (p. 38, our translation), and the matching score can depend on technical criteria that may be insignificant or arbitrary, such as whether the CV contains a LinkedIn link, or its file format. The AP questions whether an organisation needs to use such a tool without human assessment (p. 38).

The same page describes applicants hiding transparent or white text in their CVs and cover letters, aimed at AI screening tools: a sign, in the AP's reading, that candidates increasingly write for the system rather than from their own experience (p. 38). A CV score can say as much about the document as about the person.

The AP's own test

The AP tested an online assessment for a fictitious sales manager role with three fictitious candidates (box 3.4, p. 42). Each candidate received a five page personal report that did not refer to the vacancy; one showed an above average score on cognitive skills. The employer received a 26 page report on that same candidate with a profile score of 53%, plus a ranking of all candidates. The candidate's above average result did not match the employer's conclusion, and the employer could not trace how the tool reached the profile score or how each part of the test contributed (pp. 41 to 42).

That was an online assessment, not a CV parser, but the problem is the same: a number without its reasoning. The AP's key messages say explainability and transparency in recruitment fall short on many points (p. 3). It says employers should weigh whether a tool is suitable and necessary, as some selection tools have no clear link with the job requirements, and that it will look at both providers and employers in 2026 (p. 44).

What the report does not show. It is a set of observations and one small test, not an enforcement decision, and it does not say how any named tool performs.

Why would a file format or a LinkedIn link change a score?

Because some tools react to how a CV is packaged, not only to what it says.

The research behind the AP's point is an audit by Rhea and colleagues (AP footnote 157 cites their 2022 conference paper; we use the extended journal version, Rhea et al., 2022). It examined two commercial personality prediction tools sold for hiring, using CVs and LinkedIn profiles from 94 graduate students at New York University, tested between November 2020 and September 2021.

The study measured personality scores, not CV matching or hiring decisions, for US graduate students rather than frontline applicants. Its value is the method: change something that should not matter and see whether the output moves. Two tools behaved very differently, which is why you test your own.

Two other findings point the same way. In a lab simulation of CV retrieval with text embedding models, across nine occupations and more than 500 CVs, the models significantly favoured white associated names in 85.1% of cases (Wilson and Caliskan, 2024); it tested embedding models in a simulation, not a commercial screening product. And in 2018 Reuters reported that Amazon had dropped an experimental CV rating tool, begun in 2014, which learned from ten years of mostly male applications and marked down CVs that included the word "women's" (Dastin, 2018). That account rested on five unnamed sources.

These studies do not prove that every screening tool fails. They do put the burden of proof on the tool. The researchers who revised the best known estimates of how well hiring methods predict performance wrote in 2023 that they were still waiting for a body of validity evidence to build up on newer predictors, such as gamified assessments and automated scoring of video interviews or essays (Sackett et al., 2023).

Why is a CV score still a claim?

Because the score is calculated from what the candidate wrote.

Radius Hire sorts the evidence behind a hiring decision into five kinds: verified (checked against a document or a third party), demonstrated (a task observed or tested), structured (the same question for every candidate, rated against criteria written in advance), claimed (the candidate's own statement) and not covered. Our Signal Gap article uses the same five levels to audit a whole hiring process; here we apply them to one tool output.

A CV sits at claimed. Parsing it, weighting it and turning it into a number does not move it up a level. The software can add influences the candidate never controlled: the file format, a profile link, patterns learned from past hires. A score moves up only when it rests on something stronger, such as structured answers about the job, a demonstrated task or a fact verified at the source.

This matters most in frontline hiring. When applicants for warehouse, cleaning, security or hospitality roles apply from a phone or through a job board, the format of their CV can depend on the device or the platform rather than on them. The requirements that decide whether someone can do the job, such as a valid certificate, night shifts or a 05:30 start, are often not on the CV at all. That is why we read a CV as a source of questions (We Hate CVs).

See requirements and evidence side by side: Match reads each interview against the requirements you wrote and shows the answer behind each one. Explore Match

How do you test an AI resume screening tool in 10 minutes?

Score one CV in several versions and check whether the result moves.

Good practice, not a legal requirement. Radius Hire recommends this check. No law requires this specific test, and passing it does not show that a tool is fair, valid or lawful.

At about a minute per upload, the check takes around 10 minutes if your tool allows direct uploads.

Illustrative example: a worksheet for one fictitious CV. The values are invented to show how to record results and do not come from any real tool.

What the check cannot show. Whether the tool treats groups of candidates differently, which needs outcome data and often legal advice (see No Hiring Tool Removes Bias); whether the score predicts job performance; or whether its inputs serve any requirement, which the supplier questions cover.

What should you ask your CV screening software supplier?

Three things: what each input serves, how stable it is, and who reviews rejections.

For the wider buying decision, How to Choose AI Recruitment Software lists ten questions for any provider, and The Tool Should Never Be the One Saying No explains meaningful review.

What does this mean for staffing agencies and frontline employers?

Volume makes a score tempting, and it makes every error larger.

An agency screening for many clients runs one tool across many vacancies. If a score moves with file format, that error repeats on every role, and it falls hardest on candidates with the least control over how their CV was produced. Three habits help, for agencies and employers alike: write each client's or site's requirements down at intake, so every input can be tied to one; treat a missing answer as a reason for a call, not a rejection; and record, at the time, which requirement a candidate did not meet.

What does the law say about automated rejection?

Two EU laws are relevant, and how they apply depends on the facts.

Under the EU AI Act, AI systems intended to analyse and filter job applications or evaluate candidates are listed as high risk in Annex III (Regulation (EU) 2024/1689). The Act's rules for high risk AI in employment and recruitment apply from 2 December 2027, a date set by the Digital Omnibus on AI (Regulation (EU) 2026/1744), which moved it from 2 August 2026. Separately, the GDPR does not ban automated screening. It gives candidates the right not to be subject to a decision based solely on automated processing that significantly affects them, such as an automatic rejection, unless an exception applies: necessity for a contract, authorisation by law, or explicit consent (GDPR, Article 22). Where an employer relies on the contract or consent exception, candidates must at least be able to get human intervention, give their view and contest the decision; where a law authorises the decision, that law must lay down suitable safeguards. The AP adds that organisations can already be expected to meet requirements on explanation, transparency and meaningful human intervention, because these follow from the GDPR (p. 44). Whether a particular tool or configuration falls within these rules depends on how it is used, so take legal advice. Our guide to the EU AI Act and recruitment covers the dates and duties.

So, can you trust an AI resume screening score?

Trust it as far as you have tested it, and no further.

A score built from a CV is a summary of claims, and some tools add noise from how the CV is packaged. The Dutch regulator has now said so about tools used in the Netherlands. You do not need to abandon screening software to act on that: run the stability check this week, ask your supplier the three questions, and make sure a person reviews everyone the tool would turn away. Then put your effort where evidence comes from: structured questions about the job, demonstrated tasks and checks at the source.

How Radius helps

Radius Hire is the AI hiring platform for frontline roles. Candidates take a short, structured voice interview on their phone, whenever it suits them, and the recruiter sees the evidence behind each requirement.

Hold us to the same standard as any other tool: bring the three supplier questions to your demo, and see how the AI Interviewer gathers answers.