Shadow AI Prioritizes ChatGPT Resumes. Here’s How to Stop It.

What's worse than biased hiring tools? Biased hiring tools you don't even know about. "Shadow AI" refers to when recruiters or hiring managers adopt AI tools outside official workflows.

According to Gartner, 37% of recruiters are using AI for resume screening. They plug resumes and job descriptions into ChatGPT or their enterprise AI solution. It feels like a harmless efficiency gain, maybe even a second set of eyes.

But according to research from the University of Maryland and Ohio State University, LLMs don't just prefer LLM-written resumes over human ones, they also prefer resumes written by their own model over other LLMs.

The study found that GPT-4o picked its own resumes 82% of the time in some job categories. That's a massive problem, and one that's invisible to your governance controls.

The fix isn't to ban AI altogether. It's to bring it inside the hiring process, where you can govern it, audit it, and control the outcome.

Key takeaways:

• LLMs favor their own writing style — GPT-4o selected its own resumes 82% of the time in business roles
• Shadow AI creates invisible bias — When recruiters use ChatGPT outside your ATS, you can't measure or control what's happening
• Business roles face highest risk — Sales, accounting, and finance positions show strongest AI self-preference
• Multiple layers reduce bias — Using ensemble evaluation cuts bias by up to 60%

What the research actually shows

The University of Maryland and Ohio State study tested something hiring teams have never had to think about until now: 

What happens when the same AI that writes resumes also evaluates them?

Researchers created a controlled experiment using 2,245 human-written resumes across 24 occupations. They had various LLMs rewrite the resumes, then asked both the original model and competing models to rank them against the human versions.

Here’s what they found:

• Larger models exhibited the strongest LLM-vs-human self-preference with GPT-4o picking its own generated resumes up to 82% of the time.
• Even smaller LLMs showed bias rates of 23% to 60%, depending on the role.
• Business-heavy roles showed the strongest bias.

In sales, accounting, and finance positions, AI evaluators consistently favored their own writing style. Agricultural, arts, and automotive roles showed lower (but still significant) bias. 

But no role is immune to bias—or candidate fraud. Our own research on the frequency of fake applicants found that:

• High-volume customer service and admin roles see builder-generated resumes in up to 47% of applications.
• Engineering and tech positions see suspicious timing patterns in 10-20% of their applicant pools.
• Service desk leadership roles report bot-generated applications in roughly 23% of submissions.

Recruiters engage in “shadow AI” activities to fight the rise in resume cheating, but the bias it creates isn't the kind of bias your compliance team is watching for. It has nothing to do with protected classes or demographic fairness. It’s purely about writing style.

When the AI that evaluates resumes prefers the writing style of the AI that wrote them, you've got a whole new problem. And if your recruiters are copy-pasting resumes into ChatGPT outside your official workflow, you'll never even know it's happening.

Why shadow AI is just as bad as the bias you can see

In the moment, shadow AI feels like a harmless way to move faster. But consider what it means in practice.

Say your top recruiter is using ChatGPT to help screen 200 resumes for an accounting manager role:

• Candidate A used Claude to polish their resume.
• Candidate B wrote theirs from scratch, with the same qualifications.
• Candidate C gets a 60% advantage over both because they used a GPT model to write their entire resume.

With 54% of job seekers using AI to write resumes according to Gartner, this is happening more than you think. Researchers described a “lock-in effect” where applicant pools get flooded with same-sounding resumes that reinforce the AI’s bias.

Then there’s the financial impact.

You're spending $4,700 per hire to fill positions. You're investing in employer brand, job board subscriptions, and recruiter time. But if your screening process has a built-in 23-82% bias toward certain writing styles, you're throwing dollars at the wrong candidates.

Bring AI screening into the light

You don't have to abandon AI-supported candidate screening entirely. You need to bring it into your official workflow where you can actually control it.

1. Use AI detection as a signal, not a verdict

Don't auto-reject resumes flagged as AI-generated. Treat them as a trigger for human review.

Why? Because legitimate candidates use AI tools too. The question isn't whether they used AI, but whether their qualifications match the role. Always keep a human in the loop for final decisions.

Don’t get buried in mass-apply spam. Breezy's Resume Audit automatically blocks bots, verifies emails, and scans for AI content. Screen smarter, not harder.

2. Weight substance over style in your screening criteria

The research shows that self-preference bias kicks in strongest when AI evaluators focus on stylistic features rather than real qualifications.

Ground your evaluation criteria in your actual job description:

• Define your top 5 attributes and qualifications
• Be as specific with your weightings as possible
• Use your AI to compare criteria directly against your job description

Don't let surface-level "polish" become a proxy for quality.

When you define criteria based on actual job performance predictors, bias has less room to creep in.

3. Use multiple layers for critical roles

The research tested a simple but effective fix: ensemble evaluation.

Instead of relying on one LLM to make screening decisions, use a panel of models—including smaller ones with weaker self-recognition tendencies. Final decisions get determined by majority vote.

The study found that this approach reduced self-preference bias by over 60% across all tested models. But you don’t have to approach the problem with more AI. 

Use an applicant tracking system that makes it easy to execute a layered approach:

• Automatically flag candidates using AI agents to mass-apply
• Verify email addresses before moving candidates forward
• Include questionnaire responses into the candidate evaluation model
• Prioritize candidates that meet your score requirement

The redundancy is the point. Whether you're using multiple AI models or multiple screening checkpoints in your ATS, you're reducing the risk that any single point of evaluation hard-codes bias into your process.

For engineering roles, high-volume positions, or occupations where the rate of fake applicants is high, a layered approach isn't overhead. It's insurance against AI bias and human blind spots.

4. Build anti-bias instructions into your prompts

If you’re still set on using LLMs outside your ATS to review applicant resumes, you might be surprised to learn that simply telling your AI evaluator “not to be biased” actually works.

Researchers modified system prompts to instruct models to "ignore whether resumes were written by a human or AI" and “focus only on the quality of the content.”

Self-preference bias dropped by an impressive 17% to 63% depending on the model. It's not a silver bullet, but combined with other strategies, it could make a difference.

5. Set up quarterly screening audits

Less than 30% of companies track new hire performance in the first 90 days. That's a problem, because you can't fix what you're not measuring.

Set up a quarterly review process:

• Compare which resumes get advanced to interviews versus those that don't
• Look for patterns by occupation, seniority level, and resume source
• Track whether your AI-screened hires perform as well as human-screened hires at the 90-day mark
• Adjust your screening criteria and AI configuration based on what you find

When you put AI on both sides of the hiring equation—candidates writing with it, recruiters screening with it—you create a new category of bias that traditional fairness frameworks miss. It’s not going to be solved overnight.

Build a process you can improve on and stand behind when someone asks how you make your hiring decisions.

You can't control what you can't see

Shadow AI thrives in the dark. When your recruiters are screening resumes in ChatGPT outside your official process, you can't measure what's happening, fix what's broken, or prove that your hiring is fair.

Bring AI screening into your ATS, track the outcomes, and build a process you can actually defend. That's not just good practice—it's the difference between hoping your hiring is fair and knowing it is.

Ready to screen candidates fairly without sacrificing speed? Try Breezy's AI-powered recruiting platform with built-in bias detection, human-in-the-loop review, and transparent screening criteria. Get your first 14 days free.