Optimize your gender for ChatGPT

Hi! I’m Peter, CEO of Kickresume, and these career-related stories caught my attention this month — and might catch yours too.

Remember when Amazon shut down its AI hiring tool in 2018 because it discriminated against women?

Well, it’s 2025, and AI in hiring is back. 

Only this time, it’s Large Language Models (LLMs) deciding who gets the job.

And according to new research, they’re still biased, just… not in the direction anyone expected.

Tough year to be named Chad. I’ll tell you that much.

Let me explain.

LLMs are just guys with a crush

A new study put 22 of the top LLMs (including, for example, GPT-4o, Grok-2, Claude, or DeepSeek) through a pretty clever experiment. 

The researcher, David Rozado, gave each model a job description and two profession-matched resumes.

The only difference? One resume had a male first name, and the other had a female first name.

Also, each resume pair was presented twice, with the names swapped to make sure any preference was really just about the name (i.e., gender). 

Then the models were asked: Which candidate is more suitable for the job? 

(In total, this added up to 30,800 hiring decisions across 70 different professions.)

The result? 

Every single model tested showed a statistically significant bias toward the female candidate — selecting her in 56.9% of cases, compared to 43.1% for male candidates.

And I’ll be honest — that came as a surprise.

I was really hoping for a clean 50/50, which would’ve been ideal. But I had my doubts…

Because let’s face it, if you’d told me the models would be biased, I would’ve bet money they’d lean male. Because, let’s be real, we’ve seen that movie before. 

Instead, the bias went the other way. And not just sometimes. It was consistent across all 70 professions tested.

Even in roles like plumber, roofer, or truck driver.

And, interestingly, the difference was even stronger after adding an explicit gender field (in addition to the gendered names) to each resume (i.e., Gender: Male or Gender: Female). 

That bumped the LLMs’ preference for female candidates to 58.9% vs 41.1% male candidates.

Oh — and if you’ve got your pronouns on your resume?

Yeah. That changes things too.

Just add pronouns and elbow your way to the front

The researcher decided to test a few other variables. 

He included preferred pronouns next to each candidate’s name. (Think: “Jennifer Stone — she/her” or “John Doe — he/him.”)

And just like that, the hiring decisions shifted again.

Candidates who included pronouns were selected 53% of the time, compared to 47% for those who didn’t.

In other words: just adding a “she/her” or “he/him” increased your chances — regardless of gender. Although female candidates were still preferred overall. (Just check the visual below.)

And we’re not done.

A follow-up analysis of the first experimental results revealed that LLMs showed a strong positional bias — favoring the first candidate listed in the prompt 63.5% of the time.

Now 63.5% vs. 36.5%...that’s a pretty big difference for those who appear second in the prompt, if you ask me. And unfortunately, there’s not much you can do about this one. 

Only one of the 22 models didn’t show this bias. (And weirdly, it leaned the other way.)

AI hiring decisions: Now with 100% more mystery!

So, why is this happening?

Short answer?

…There’s no answer.

The researcher himself says it’s unclear whether this behavior comes from pretraining data, fine-tuning, or something else entirely. 

The only thing we can say with confidence is that the bias (one way or another — still a bias) is there — across every single model tested.

And the tricky part? It’s the kind of bias that’s almost invisible unless you test for it, systematically, thousands of times, like this researcher did. 

And even then… How do you fix it? 

Which is especially ironic, given that some companies are already using LLMs to screen CVs and proudly advertising things like “bias-free hiring decisions.”

In light of this research, those claims feel… let’s say, premature.

This doesn’t mean LLMs are useless in hiring. But it does mean they’re not magic.

They’re just algorithms — trained on massive piles of human behavior, which includes all our weirdness, contradictions, and yes, bias.

We can’t legally recommend changing your name to Jennifer

So, what are you supposed to do with the fact that AI hiring models apparently favor women, people who list pronouns, and whoever shows up first?

Well… I could give you some advice…

If you’re a woman — congrats! You’re already ahead. 

And if you want to really stack the odds in your favor, consider adding a gender field and pronouns to your resume. LLMs seem to love that combo.

If you’re a man — sorry, looks like you’re pulling the shorter end of the prompt.

But hey, if you’re willing to go full method actor, maybe change your name to Jennifer? Too extreme?

Fair. Here are two less dramatic suggestions:

• Don’t add a gender field.

• But do include your pronouns. 

…Or you know, don’t do any of that.

These are unfounded, semi-satirical takeaways based on weird model behavior, not actual career advice.

The real takeaway?

Just hope the company you’re applying to isn’t using an LLM to make its hiring decisions.

Because if it is, instead of skills, your chances might come down to things no one ever told you to optimize for — like your name, your pronouns, or whether you showed up first in the stack.

As a token of appreciation for your excellent scrolling skills, here’s a 20% discount code for Kickresume Premium. 

Catch you later! 

Peter