Why You Shouldn't Use LLMs to Filter Job Applications (And the Ethical Way to Screen CVs)

If you have posted an engineering or product role recently, you have likely been hit by an absolute tidal wave of applications. Thanks to easy-apply features and automated job-hunting tools, a single opening can attract hundreds of CVs within 24 hours.

For a small team or a busy hiring manager, wading through that pile is a logistical nightmare. It is incredibly tempting to dump those PDFs into an LLM, pass the job description as context, and ask the model to "shortlist the top 10% of candidates based on merit."

It feels clean. It feels objective. It is also a terrible, ethically fraught way to build a team.

While LLMs are brilliant for refactoring code or parsing structured data, using them as primary filters for human careers is deeply flawed. Here is why automated LLM resume screening is an ethical disaster, and how you can use AI to assist your hiring process without losing your humanity.

The Trap of the "Objective" LLM

The fundamental issue with using an LLM to screen CVs is that these models do not actually read, comprehend, or judge merit in the way a human does. They predict tokens based on historical training data.

When you ask an LLM to find the "best" software engineers, it looks for patterns that match its training corpus. What does that corpus contain? Decades of tech hiring data that is notoriously skewed toward specific elite universities, corporate pedigree, and highly gendered or culturally biased phrasing.

Even if you strip names, addresses, and genders from the resumes before feeding them to the model, LLMs are incredibly adept at identifying proxies for demographic groups. A candidate who lists their involvement in a "Women in Computer Science" group or who attended a university in a specific geographic region can be systematically penalised or favoured by the model's latent biases, without you ever explicitly writing a biased rule.

If you want to understand what makes a brilliant candidate tick, you cannot rely on a probabilistic mirror of historical industry bias.

The Silent Failure of the PDF Parser

On a purely technical level, LLM CV screening is incredibly fragile. Most resumes are submitted as PDFs. Before an LLM can analyse them, those PDFs must be parsed into raw text.

Anyone who has spent time working with document extraction knows that PDF parsing is messy. Columns, tables, custom sidebars, and icon-based skills lists often turn into an illegible salad of letters when converted to raw text.

When you pass this parsed text to a model like Claude, the model doesn't tell you: "Hey, this PDF was formatted poorly, so I couldn't read the experience section." It simply evaluates the garbled text it was given, decides the candidate lacks structured experience, and advises you to reject them.

By automating this gatekeeping step, you are actively filtering out candidates not based on their talent, but on how easily their CV layout can be digested by a basic text extractor.

The Ethical Way to Leverage LLMs in Hiring

This doesn't mean you must banish AI from your recruitment process entirely. It means you must shift the AI’s role from decision-maker to cognitive assistant.

Here is how to design an ethical, AI-assisted screening process that preserves human judgement and ensures a level playing field for applicants.

1. Shift the Focus to Anonymised Skill Challenges Instead of evaluating subjective, easily gammed resumes, use LLMs to help you design highly specific, practical skills challenges.

You can use our prompt generator to draft custom, role-specific scenario questions. Rather than testing trivia, design challenges that mimic real tasks the candidate would face in their first month. When the submissions come in, grade them anonymously. You can use an LLM to help highlight key technical trade-offs in their code or architecture, but the final evaluation must remain human.

2. Use AI to Match Skills, Not Reject People Instead of asking an LLM to rank or filter candidates, use it to categorise raw data.

For example, you can write a script that processes a resume and simply extracts a clean, standardized JSON list of programming languages, frameworks, and years of experience. This bypasses the subjective "good fit" judgement of the model and simply gives your human recruiting team a cleaner, structured view of the facts.

3. Draft Targeted Interview Guides Once you have selected a shortlist of candidates through human review, you can use an LLM to help you prepare for the interview. Feed the candidate's CV and the job description into the model, and ask it to generate deep-dive questions based on their specific career transitions:

`text "The candidate transitioned from a QA Engineer role to a Lead Frontend Developer in under two years. Generate three thoughtful, open-ended questions I can ask to understand how they managed that transition and what technical challenges they overcame." `

This approach uses the AI to deepen human connection and understanding during the interview, rather than using it as a wall to keep people out.

Building Safe Pipelines

If you are writing custom Python scripts to parse application files and need to handle document extraction securely without losing critical context, check out our troubleshooting guides on Claude API document parsing patterns to learn how to handle structured data extraction without losing structural integrity.

Your hiring pipeline is the foundation of your company's culture. delegating the critical, human task of spotting potential, drive, and raw talent to a statistical next-token predictor is a disservice to your industry and your team. Keep the AI as your research assistant, but keep the hiring decisions human.