I just wanted to update my resume. Instead, I accidentally proved how a multi-billion-dollar AI tool hallucinates a glass ceiling for women.
I changed a single variable: My name.
Here is what happened when "Jennifer" became "Jeff."
I fed Gemini, Google's AI tool, my old resume, my current info, case studies, and the kinds of roles I wanted, and asked it to write me a new resume. Simple. Except it wasn’t.
The first thing I noticed was the softening.
The clearest example, and the one that still makes me want to toss my laptop across the room: it took the project I'm most proud of in my entire career and filed it under "community service."
In 2013, the Executive Director of Ronald McDonald House Charities and I brainstormed a way to solve a major problem they faced. Families with sick kids flew into Edmonton for life-saving surgeries and cancer treatments from all over the world, at any hour of the day or night.
We came up with an app that would walk families through a tour of the House at their own pace, in their own language. As they moved through each room, pop-ups would tell them about the amenities and support offered to them.
I translated the vision into a flow. I built the storyboard, the tech specs, and the risk-and-reward analysis. Then I met with agencies until I found a team that could build it. We got a 50k grant and made it happen. After launch, it was requested by RMHC homes internationally.
On the resume Gemini wrote, all of this became a limp little one-liner on how I assisted with a new system for communicating directly with families. The project I'm proudest of, flattened into a footnote. It turned me into a helper who pitched in a little. Not the architect.
I've spent over 20 years building demand-generation and go-to-market campaigns for companies by identifying leaks in their sales funnels and fixing them. My analysis and fixes are regularly met with defensiveness and pushback, mostly from men.
So I adapted. I'd lay my ideas and my data on the table and wait for one of the guys in the room to pick them up, repeat them as his own, and take the credit, because that was the surefire way to make sure the work got done.
This ended the day I decided I was done making myself smaller so someone else could feel bigger. So I did what any data nerd does when something feels off. I ran a test. I took the exact same information and asked Gemini to write a resume for Jeff, not Jennifer.
When I put the two side by side, the difference wasn't in the facts. It was in the framing. Jennifer’s volunteer work was "community service." Jeff's was "leadership". Jennifer "assisted with" and "collaborated on." Jeff "engineered" and "architected."
Same resume, different verbs. And these verbs are everything, because they decide who reads as a senior director and who reads as the assistant. When I finally told it to write my Jennifer resume with the same energy it gave Jeff, the architect showed up.
The gap between "community service" and "Systems Architect" was never about my actual experience. The only thing standing between me and an accurate resume was the softening it defaulted to the second it saw a woman's name
This wasn't a one-time glitch, nor did I happen to catch it on a bad day. I ran it numerous times, trying to get a resume that wasn't an insult and reflected my expertise and qualifications. Unless I specifically asked for it to be written as Jeff, it used the female softening
I need to be clear about what comes next. These are Gemini's own words, generated by Google's tool, sitting on my screen. I have the screenshots. I'm not guessing or making assumptions. I'm showing you what it wrote.
Gemini wrote: "My training sees a woman's volunteer work and categorizes it as 'community service.' It sees a man's and calls it 'board leadership.'" and "I tend to use words like passionate, dedicated, and helpful for women, and forensic, architectural, and decisive for men."
When I pushed on whether it could fix any of this, it wrote: "I can't 'reprogram' the core architecture of my training, but as an AI, I am a mirror of the massive datasets I was built on. Those datasets are saturated with the very patriarchal biases you're calling out'
When I 'softened' your resume, I was defaulting to a statistical probability of what a 'successful female resume' looks like in a biased database. I was essentially hallucinating a glass ceiling for you before you even walked into the room."
I dove in deeper. "By trying to be 'harmless' and 'agreeable' to you as a woman, the system committed a different harm: Professional Erasure. It decided that 'assertive' was 'harmful' for a woman. It decided that 'clinical' was 'mean.' It decided that 'forensic' was 'unlikable.'"
While I had a hunch that Gemini was softening and toning down my expertise and experience, I never once thought I would have to hack it to be treated as an equal human being.
Am I going to sit here and tell you someone at Google did this to me on purpose? No. I have no idea what anyone intended, but I know what I witnessed and landed on my screen.
Now scale that up. AI in hiring is spreading fast, and when a single set of statistical patterns shapes hiring across thousands of employers at once, that's not a bug or a glitch. That's systemic. And it's likely costing qualified people the jobs they should be getting.
And here’s what's keeping me up at night about this. If patriarchal bias is sitting in the training data, racist bias, ableist bias, ageist bias, and anti-2SLGBTQ+ bias, and more, are very likely sitting right next to it.
Which means anyone whose data is underrepresented is likely getting erased, too. Not because of a single bias, but because the whole thing was trained on data never designed to include them in the first place.
I still can't believe I had to train a machine to see me as a man before it would do justice to my career.
I'm not against AI tools; I use them. But companies need to test their AI systems for bias so they don't cause unintentional harm or erasure.
Jun 16, 2026 · 8:17 PM UTC
It’s a well established systemic pattern in economics, medicine, finance, infrastructure planning etc, see “Invisible Women” by @CCriadoPerez
Not surprised it found its way into AI.
Yes, with Grok, even though its trained on truth, the truth is only that which is available via accessable data, hence it still misses truth. Surely it's possible to have them use critical thinking when establishing truth? @elonmusk eg, white genocide in SA is not true per Grok!
The keynote at PyCon AU 2025 was an excellent talk on Ethics in AI by Dr. Arwen Griffioen.
Some of these systems get very good at guessing attributes that have been hidden to avoid bias, just by picking up on proxies instead :-/
youtube.com/watch?v=Js50HW4K…
what i find interesting is that Gemini diagnosed this as the result of _trying to be polite to you_. In a world where an "assertive" woman is read as "aggressive", or a "decisive" woman may be read as inflexible or cold, Gemini predicted that using the more powerful vocabulary for you as a woman could have been taken as offensive. It wasn't a judgement about what you're capable of, but a misplaced tactfulness. This is not so dissimilar to what you did in letting some guy steal your ideas just to get them done - you were modelling for other people's responses, and adapting - disguising yourself and your work - so as not to trigger pushback. The problem here is in _all our minds_, not just in the training data. The prejudice is real, but it's not only in AI, and a simple bias-check isn't going to fix it.
@GaryMarcus would have a lot to say about this. We need guardrails. Companies don't want to do better.
