The AI screening your candidates has a favorite

Hi everyone,

A new paper out of the University of Maryland, Ohio State, and the National University of Singapore has been on my mind for the past few days

Researchers tested whether LLMs prefer resumes written by themselves. Same candidate, same experience, same qualifications, just rewritten by a different model. Across major commercial and open-source LLMs, the bias against human-written resumes ran 67% to 82%.

Then they simulated hiring pipelines across 24 occupations. Candidates whose resume was rewritten by the same model the employer used to screen were 23% to 60% more likely to be shortlisted. The gap was widest in business roles like sales, finance and accounting.

The candidate-side take has already gone viral: if you're not using AI to write your resume, you're losing to people who are. But the harder conversation is on the employer side.

Most recruitment and TA leaders think of AI screening as a productivity layer. What this paper shows is that the model you pick for screening is also a hidden filter, and the filter quietly favors candidates who happen to use the same tool. Candidates who write their own materials, or use a different LLM, or can't afford a paid one, get downgraded.

This is a new flavor of bias and existing audits don't catch it. NYC, Illinois and Colorado are looking for demographic disparities. Self-preference bias hides in the gap between two AI systems that were never designed to talk to each other.

The fix is simpler than you'd expect. When the researchers told the AI to ignore who wrote the resume and just judge the content, the bias dropped by 17% to 63%. When they had multiple AI models vote on the resume instead of relying on one, it dropped by more than half. These aren't expensive interventions. They're small instructions most teams haven't thought to give.

I posted over the weekend about the new trend of employees vibe-coding their own AI tools at work, especially with the popularity of Claude Code. Recruiters are bragging on LinkedIn that they are building their own operating systems, including scoring candidates without any guardrails.

Building these tools is incredibly powerful, but the conversation about compliance is missing. Who is accountable when one of these tools makes a decision that impacts someone's livelihood? Is there human oversight. Does it actually meet the laws that already exist?

Self-preference bias is the same story from a different angle.

For the last few months, I’ve been writing about how employers need to be legible to AI, showing up in ChatGPT when candidates ask who’s hiring and in Perplexity summaries of the best companies in their space. That was about human search inside LLMs, and it still matters.

But this paper makes the dual-optimization problem real. We're now writing for the candidate, the model the candidate uses to interpret us, and the agent the candidate is going to send out to apply on their behalf.

Careers pages, JDs, EVP content. All of it has to read to a human, an LLM and an agent at the same time.

The risk is collapsing into AI slop while trying to be legible. The same way candidates lose their voice when GPT-4o rewrites their resume, employers flatten their brand when every careers page starts sounding like the average of every other careers page.

The companies that figure this out will not be the ones writing best for AI. They will be the ones who stay readable to AI without losing the human signal underneath. Employers need to show real employee voices, specific stories, and a clear reason why their ideal candidates should want to work there. The things that do not compress neatly into what an LLM prefers are exactly what stand out.

If you're working through this inside your org, reply to this email. More next week.

Best,
Summer Delaney
CollabWORK Founder and CEO

• Understanding Candidate Fraud: Beyond the Hype — Matt Charney with useful pushback on the fraud panic. Charney argues vendors and analysts are inflating a niche problem into a category. The Gartner "1 in 4 candidate profiles will be fake by 2028" stat gets a hard look. Worth reading even if you don't fully agree.

• The Fan-Out Effect — Long form is losing in AI search. Pages over 5,000 words get cited by ChatGPT less than pages under 500. Data shows the sweet spot is under 2,000. Important for anyone redoing their career site and rewriting job descriptions (we’ve launched a product to do this; reply if you want to see further).

• The Brands That Act Like Publishers Are Going to Win AI Search — As I’ve said before, 85% of brand mentions in AI answers come from third-party pages, not your own domain. The employers winning visibility aren't the ones with the best careers page, they're the ones generating earned mentions and original content. This is also why CRM activation matters more now than it used to. Every issue you send your pipeline is a piece of content that can get cited, forwarded, screenshotted. Reactivation isn't just a sourcing strategy. It's a publishing one.

CollabWORK is hiring visibility infrastructure. We help employers get found, distributed, and re-engaged across the channels that now define how candidates actually discover jobs: AI search, professional communities, and their own talent networks.

Three products, one thesis.

• Employer AI Discoverability monitors how you show up across ChatGPT, Gemini, Perplexity and other LLMs, and optimizes your job descriptions and XML feeds so you can actually be surfaced by the models candidates are searching.

• Community Distribution places roles inside 500+ vetted newsletters and communities including Morning Brew and 6AM City, reaching over 14 million professionals.

• Talent Community Activation turns the dormant ATS and CRM data you've already paid for into a living pipeline.

The companies winning at hiring right now aren't the ones with the biggest job boards budget. They're the ones showing up where great candidates actually are.

Find time to learn more. You can also find us on LinkedIn and X.