Why You Shouldn't Use LLMs to Screen Job Applicants' Resumes (And the Ethical Way to Filter Candidates)

The Seductive Myth of the Unbiased AI Recruiter

We have all been there. You post a job opening for a mid-level developer or a marketing manager, and within forty-eight hours, you are staring down a digital mountain of three hundred CVs. Your eyes glaze over. The temptation to spin up a quick script using OpenAI or Claude to parse, rate, and filter these candidates is incredibly strong.

On paper, it sounds like a triumph of modern engineering. You write a prompt, feed in the job description, loop through the PDFs, and ask the model to output a ranked list. Proponents argue it is faster, cheaper, and even "unbiased"—after all, a machine does not care about a candidate’s gender, ethnicity, or whether they went to the same university as the hiring manager.

Except, it does. Or rather, it cares about the patterns it has inherited from the vast, messy, historically biased corpus of the internet. Handing your hiring pipeline over to an LLM is not just an ethical shortcut; it is a remarkably effective way to build a mediocre, homogenous team while systematically locking out exceptional talent.

Why LLMs Flunk the CV Screening Test

To understand why LLMs should never be given the keys to your recruitment gate, we need to look at how these models actually process human experience. It is not just a matter of ticking a compliance box; it is about understanding the fundamental mismatch between statistical pattern matching and human potential.

1. The Homogenisation Trap

LLMs are, by their very nature, engines of predictability. They are trained to predict the next most likely token. When you ask an LLM to evaluate a resume against a job description, it looks for the most typical representation of that role.

This means the model will naturally favour candidates who have taken the most conventional, well-trodden paths. The self-taught engineer who spent three years building open-source tools while working in retail? The career-changer who brings invaluable cross-disciplinary skills from hospitality? The candidate with a non-traditional educational background? The LLM will discard them because their profile does not align with the standard statistical template of the role. You end up hiring the same carbon-copy candidates, strangling your team's diversity of thought at the very first step.

2. Hallucinations and Invented Career Trajectories

If you have spent any time in our Claude troubleshooting hub or OpenAI guides, you know that LLMs hallucinate. Usually, this is a minor annoyance. When screening candidates, however, a hallucination can be career-ending.

An LLM parsing a poorly formatted PDF might misinterpret dates, invent gaps in employment, or completely misrepresent a candidate’s actual responsibilities. Because LLMs present their findings with absolute, unearned confidence, a hiring manager looking at an AI-generated summary will rarely double-check if the candidate actually lacks the required years of experience or if the model simply failed to parse a two-column resume layout.

3. Over-Indexing on "Vibe" and Jargon

Candidates are already using AI to write their resumes. When you use AI to screen those resumes, you are essentially letting two algorithms play a game of keyword bingo in a dark room. LLMs are incredibly susceptible to style over substance. A candidate who has stuffed their CV with high-value industry buzzwords and polished prose—courtesy of a sophisticated prompt generator—will score incredibly high. Meanwhile, a highly competent candidate who describes their actual, practical achievements in plain, unpretentious language will be filtered out. You are not hiring the best performer; you are hiring the person who knows how to game the model.

The Ethical Way to Streamline High-Volume Hiring

If we cannot let the LLM make the cut, how do we handle the sheer volume of applicants without burning out? The key is shifting the AI's role from evaluator to administrator.

1. Use LLMs for Pure Extraction, Never Evaluation

It is ethically sound to use LLMs to clean up and structure data, provided you do not let them pass judgment. You can use an LLM to extract a list of programming languages mentioned in a CV, or to pull out the candidate's notice period into a clean spreadsheet.

`json { "action": "extraction", "fields": ["notice_period", "current_location", "languages_mentioned"] } `

Never ask the model: "Is this candidate a good fit for the role?" Instead, ask: "Does this resume mention experience with React?" Keep the questions binary, objective, and easily verifiable by a human. If you run into formatting issues during extraction, check our guide on structuring outputs in the glossary.

2. Implement Blind Skills-Based Assessments

If you want to remove human bias and save time, do not rely on a resume screen at all. Instead, invite applicants to complete a brief, practical work sample that mimics the actual job. Ensure this test is anonymised. Evaluating someone's actual ability to write a short code snippet, draft a customer reply, or organise a project plan is infinitely more predictive of job performance than an LLM's automated vibes-based assessment of a two-page PDF.

3. Establish Structured, Human-Led Rubrics

When humans do review the resumes—and yes, you still need to do this—use a rigid, structured rubric. Score every candidate on the same three or four criteria. This keeps your human biases in check far better than an LLM ever could, because you are actively forcing yourself to justify your scores based on evidence, rather than outsourcing your critical thinking to a black-box model.