AI, robotics and crypto, in context Deutsch

AI

Screening Job Applicants With AI: Fairness You Can Prove

Software now sorts, ranks and sometimes interviews applicants, and the EU treats such systems as high-risk. What it takes to make AI screening fair, checkable and clearly overseen by recruiters.

Screening Job Applicants With AI: Fairness You Can Prove

A popular job posting can draw hundreds of applications within days. No recruiter reads all of them with equal care. Software has filled the gap: it extracts data from resumes, matches profiles to job requirements, ranks candidates and books interviews. Some tools also run online tests or chat with applicants before a human gets involved.

These systems promise speed and consistency, and they can deliver both. They also make decisions that shape people's working lives, often without the applicant ever learning that software was involved. A rejected candidate rarely knows whether a person or an algorithm said no.

European lawmakers have drawn a clear line here. The EU AI Act lists AI systems used to recruit and select people as high-risk, for example tools that place targeted job ads, filter applications or evaluate candidates. That classification brings obligations. It also sets a useful standard for any employer: if a tool influences who gets hired, it must be possible to show that it works fairly and that people stay in control.

Where software enters the hiring process

AI in recruitment is not one product. It is a set of functions that appear at different stages, each with its own risks.

  • Job ads: text generators draft postings, and ad platforms decide who sees them, which already shapes the applicant pool.
  • Resume parsing: software turns documents into structured fields such as job titles, dates and skills.
  • Matching and ranking: a model scores how well each applicant fits the requirements and sorts the list.
  • Chatbots: assistants answer questions, collect missing information and schedule interviews.
  • Assessments: online tests, games or recorded video interviews are scored automatically.

The earlier a tool acts, the less visible its effect. If an ad never reaches certain groups, or a parser misreads an unusual resume format, those candidates drop out before anyone looks at them.

How bias gets into a ranking

Most ranking models learn from examples. A common source is a company's own history: who was hired, who was promoted, who left early. If that history reflects old preferences, the model learns them too. It does not need to see gender or ethnicity to do so. Proxies are enough.

A career gap may stand for parental leave or illness. The name of a school or a sports club can correlate with social background. Even word choice in a resume differs between groups. A model that rewards the patterns of past hires can quietly penalize anyone who looks different from them.

Technical errors add to the problem. Parsers struggle with two-column layouts, foreign degrees or unconventional job titles. A strong candidate can rank low simply because the software misread the document.

What high-risk status means in practice

Under the AI Act, the provider of a high-risk system, usually the vendor, must run a risk management process, use suitable training data, keep technical documentation and logs, and design the system so that humans can oversee it. The employer using the system, called the deployer, has duties as well. It must use the tool according to its instructions, assign competent people to oversee it, keep the logs it controls and inform workers and their representatives where required.

The act also bans AI systems that infer emotions in the workplace, apart from narrow medical or safety reasons. That rules out tools claiming to read enthusiasm or honesty from a candidate's face in a video interview. Separately, EU data protection law already restricts decisions based solely on automated processing when they significantly affect a person.

Employers do not need to wait for every detail of enforcement to settle. The core expectations are clear enough to build processes around them today.

If nobody can explain why a candidate was rejected, the company cannot defend the decision either.

Making screening checkable

Fairness in hiring is hard to prove after the fact. It becomes manageable if the evidence is collected along the way. Several practices help.

Keep records of which model version and which criteria were used for each vacancy, what score each applicant received and where a recruiter overruled the ranking. Compare outcomes across groups regularly, as far as data protection law allows. If one group passes the screen at a markedly lower rate, find out why. Test the tool with matched pairs of resumes that differ only in a name, an age signal or a career gap.

Ask the vendor for documentation in plain language: what the model was trained on, what it optimizes and how it was tested for bias. A vendor that answers only with marketing claims leaves the employer carrying the risk alone.

Assessment tools deserve the same scrutiny. A game-based test or an automated interview score should measure something linked to the job, and the vendor should be able to show that link. Candidates who need accommodations, for example because of a disability, must have a simple way to ask for them.

The recruiter's job changes, but it does not vanish

Human oversight is easy to write into a policy and hard to practice. A recruiter facing three hundred ranked profiles will naturally start at the top and stop when the shortlist is full. In effect, the software has made the decision. This tendency to follow automated suggestions is known as automation bias.

Real oversight requires time and authority. Recruiters should regularly look below the cutoff, understand which factors drive a score and feel free to disagree. They also handle the part software does badly: judging motivation, explaining the role honestly and treating candidates with respect, including those who are turned down.

Used this way, AI can take over sorting and scheduling, while people concentrate on judgment. That division only holds if the organization measures it. A useful sign is how often recruiters actually change the ranking. If the answer is almost never, oversight may exist only on paper.

Questions and answers

Is it legal to reject applicants automatically?

In the EU, data protection law restricts decisions based solely on automated processing that significantly affect a person, and a job rejection usually qualifies. In most cases, a human must play a genuine part in the decision. Clicking to confirm a machine's verdict is generally not enough.

Does the EU AI Act ban AI in hiring?

No. It classifies AI systems used for recruitment and selection as high-risk, which means stricter requirements rather than a ban. One related practice is prohibited outright: systems that infer emotions in the workplace, except for narrow medical or safety purposes.

Does removing names and photos make AI screening fair?

It helps, but it is not enough. Models can pick up the same information from proxies such as schools, career gaps or word choice. Regular checks of outcomes across groups remain necessary.

What can applicants do about AI screening?

Clear resumes in a common format are parsed more reliably than elaborate layouts. Applicants may ask an employer whether automated tools were used and how decisions were made. In the EU, they also have rights to information about how their personal data is processed.

Who is responsible if a hiring tool discriminates?

The employer remains responsible for its hiring decisions, even if a vendor built the tool. The vendor has its own duties for the product. In practice, both sides need contracts and documentation that make clear who checks what.