NLP in Recruiting: How Text Analysis Cuts Bias – and Where It Creates New Risks
Two job ads, almost the same role, almost the same company. One is looking for "assertive go-getters for a demanding, competitive environment." The other wants "committed, reliable team players who want to take ownership." Two word clouds, two completely different applicant pools – without anyone at either company deliberately planning it that way.
That's exactly where a technology comes in that's currently playing two roles at once in the recruiting debate: NLP, or natural language processing, the machine analysis of language. On one hand, it promises to surface exactly these unconscious patterns and shrink bias in job ads, screening, and evaluation. On the other hand, the same technologies – especially once they're pointed at spoken language in interviews – carry the potential to create entirely new, more subtle forms of discrimination. Both are true, at the same time, and that's exactly what makes this topic relevant for TA teams in Switzerland, Germany, and Austria right now.
The Language Test Almost No Job Ad Passes
Just how big the problem really is shows up in a recent analysis by StepStone, which scanned 683,000 job ads across the DACH region. The result: 96 percent of ads contain gender-coded language. 50 percent of those lean masculine, 32 percent lean feminine, and only 17 percent are written in genuinely neutral terms. Typical masculine-coded adjectives include "independent," "individual," "challenging," or "career-oriented," while typical feminine-coded ones include "committed," "reliable," "responsible," or "motivating." The effect is especially pronounced at the leadership level: management positions are worded with masculine coding 62 percent of the time, compared to just 49 percent for entry-level roles.
So anyone specifically trying to attract women into leadership roles is working against themselves with ad language that pulls in the opposite direction – without anyone intending it.
Textio, a US-based provider that specializes in exactly this kind of text analysis, has documented the link even more directly: in an analysis of nearly 79,000 engineering job ads, postings that ended up resulting in male hires contained almost twice as many masculine-toned phrases as feminine ones – and the reverse held for postings that resulted in female hires. Machine-intelligence roles came out as the most masculine-worded by a wide margin, compared to data science, backend, frontend, or full-stack. The language in a job ad doesn't just say something about a company's tone – it actually predicts who ends up applying and getting hired.
This is exactly where the constructive side of NLP comes in. Tools that scan ad copy automatically don't work off a rigid word blacklist; they work with statistical patterns drawn from real application and hiring data – so they don't just flag "the bad word," they learn from the correlation between wording and the applicant pool it actually produces. In practice, that means: before a job ad goes live, run it through a language check that flags tone, clichés, and gender coding – a simple, cheap lever with a direct effect on the diversity of applications, long before anyone even looks at a resume.
The Flip Side: When the Fix Becomes the Problem
As clear as the benefit is for job ads, things get murky once the same family of technology meets spoken language in the interview itself. This isn't about written words anymore – it's about automatic speech recognition, tone analysis, and in some cases even facial-expression evaluation during the interview.
A case from the US shows where that can lead: a deaf, Indigenous employee at Intuit applied for an internal promotion and was required to complete an automated video interview on the HireVue platform. Her request for human-generated captions as accommodation was denied, and her application was subsequently rejected.
The US civil-rights organization ACLU filed a complaint afterward, alleging that the automatic speech-recognition system systematically underperforms for non-white speakers as well as deaf and hard-of-hearing people – with the effect that qualified candidates get screened out for technical reasons, not professional ones.
It's worth pausing here for some context, because bias in automated screening has already been front and center on this blog once before – in the post on false negatives and the rescue-loop audit playbook for ATS filters. This time, I'm deliberately taking a different angle: instead of asking who a score-based filter wrongly rejects, this post looks specifically at language and speech as a source of bias – the level where NLP operates directly, with a clear focus on the new legal landscape around exactly this technology. The two perspectives complement each other: one looks at the filter as a whole, the other at the linguistic raw material that feeds it and that it judges.
Regulators have responded to exactly this risk, and unusually clearly. The EU AI Act classifies automated applicant management as a high-risk application starting August 2, 2026, including CV screening, automated ranking, and performance evaluation. Certain practices aren't just regulated but banned outright: emotion recognition based on facial expressions in video interviews is explicitly one of them. That puts exactly the combination of NLP and sentiment analysis at the center of the HireVue case squarely in the "not allowed" category – not "use with caution."
What the EU AI Act Means for DACH in Practice
High-risk systems face seven core obligations starting August 2026: a conformity assessment before deployment, documented risk management across the entire system lifecycle, demonstrably representative and unbiased training data, technical documentation including activity logging, transparent disclosure to candidates about whether and where AI is being used, human oversight with a documented ability to override an AI recommendation, and robustness plus cybersecurity.
Violations can get expensive: high-risk breaches carry fines of up to €15 million or 3 percent of global annual revenue, while banned practices can reach up to €35 million or 7 percent of global annual revenue.
For Germany and Austria, as EU member states, this applies directly and immediately. Switzerland is a different story, even though plenty of articles gloss over this: Switzerland isn't an EU member, so the AI Act doesn't automatically bind Swiss companies. That said, anyone recruiting EU candidates, using tools from EU vendors, or operating as part of an international group will feel the requirements in practice anyway – through supply chains, parent companies, or simply because the vendor in question builds its software to be AI-Act-compliant regardless.
Companies in Switzerland that already align with the seven obligations now are covered either way, no matter how the Digital Omnibus debate over a possible delay to the deadline plays out.
GDPR remains relevant in parallel, equally across all three countries. Article 22 gives applicants the right not to be subject solely to an automated decision with legal or similarly significant effects – which means an automatic rejection with zero human review is already on shaky ground, independent of the AI Act. Anyone using NLP-powered pre-screening should have a human in the loop for data-protection reasons alone, not just because it becomes explicitly mandatory in 2026.
What TA Teams Can Take Away
In practice, this comes down to three levels.
First, the easy, immediately actionable step: run every new job ad through a language check before it's published – especially for leadership roles, where the StepStone numbers show the bias is strongest.
Second, take a critical look at anything in the hiring process that evaluates tone, pace, or speech patterns: is it merely transcribing and analyzing what was said, or does sentiment, facial expression, or speaking speed also feed into an evaluation? The latter falls largely into banned-practice territory starting August 2026 and is worth questioning right now.
Third, the organizational homework: build an inventory of every AI tool used in recruiting before the deadline gets closer – whoever already knows today which tool does what will have a much easier time demonstrating the seven obligations than someone who only starts looking in summer 2026.
In the end, a familiar pattern emerges in recruiting: the same technology can expose and reduce bias, or reproduce it in a new form – depending on where in the process it's applied and how much human oversight remains. NLP applied to job ads is already a pragmatic win today. NLP applied to speaking behavior in interviews is, as things stand, more of a warning sign than a tool – which is exactly why it pays to take a second look before a vendor sells you the next interview round with the word "AI-powered."
Sources
Transparency Notice on the Use of AI
Artificial intelligence tools were used in a supporting role in the preparation of this article. They are primarily used for research and information structuring, linguistic and grammatical review, translation, and, in some cases, the creation or editing of illustrations and visual content.
The concept, substantive statements, assessments, and conclusions are those of the author and reflect his personal views, professional experience, and convictions. All content is reviewed editorially before publication and remains under the author’s responsibility.
Articles on this website are neither created nor published automatically by AI. Artificial intelligence is used as a supporting tool only; full editorial and substantive responsibility remains with the author.
Comments