Pride in the AI Pipeline: 10 Ways to Keep AI From Erasing Queer Users

I grew up in a small rural farming town of about 300 people, and the only gay man I knew about was one we never really talked about.

He is my uncle. I knew the way kids know things nobody says out loud. My mom and Dad never addressed the pink elephant in the room. We didn’t dare utter the truth or celebrate him for him. This was rural Indiana in the '80s and '90s (yes, I’m a Millennial, and thank you, I do look good for my age ;-), and the de facto rules were conformity or silence. That silence, I would later come to understand, seeded shame and regulated exploration, expression, and even existence to the shadows (hence, the coming out movement). There weren’t any familiar or friendly role models to help me develop a healthy understanding of my sexuality and identity.

So I learned who I was the way a lot of closeted kids did back then: from the television, for better or worse.

For me, the big one was Will & Grace. I was already in high school when I really started watching, and for the first time, I saw a gay man who wasn't a punchline. Will wasn't camp. He wasn't a lisp played for laughs. He was a person who happened to be gay, and his sexuality was one part of him instead of the whole joke. That mattered more than I can explain. Representation was my proxy. It was how I figured out that I could and should exist.

“Representation was my proxy. It was how I figured out that I could and should exist.”

However, while the show shed light on a possible future, it also cast new shadows. Unknowingly and unintentionally, it established new boundaries for my existence as a gay man—stereotypes that I’m still unlearning today.

Will was the acceptable, more relatable gay man — subdued, palatable, easy for a straight audience to root for. Jack was the camp one, the punchline, the effeminate one—the butt of even Will’s jokes.

Two steps forward, one step back.

I absorbed that hierarchy (along with many others) without choosing it or questioning it. For years, I looked down on effeminate gay men and idealized masculine ones. I had a rule in my head that a man with an earring, especially in the left ear, "was gay" in a lesser, demeaning way. To be femme was undesirable and shameful. “Masculine, no femmes” was all over dating profiles and written on people’s faces. Within queer culture, we have coined the term "internalized homophobia.”

I work on unlearning these habits—catching these subconscious weeds before they take root. Not by beating myself up, but by identifying, noting, and intentionally reframing to something kinder: celebrate people for who they are, drop the gender rules I was handed. (If you ever see me in earrings, you'll know I won. Got a cool pair? Send them my way.)

I'm telling you this because it's the most honest way I know to explain what worries me about AI in our work.

AI is the new proxy & a problema…

For a long time, the media was my proxy for who I could be. Now, AI is becoming everyone's proxy for everyone—especially the next generation. A novella of text in response to questions—whether true or not—is taken as gospel. Dr. Google is now Dr. Claude. You ask it a seemingly benign question and get back a palatable answer that’s laden with subtext and scripts—the good and the bad of our past.

On a more professional note, when you point a model at your interview transcripts to synthesize themes, or ask it to draft a persona, or let it summarize an open-text survey, it produces the artifact that shapes what your team builds next. AI’s the mirror now.

And what unsettles me as a researcher, gay man, and human is that the model's bias is internalized, just as mine is. It learned its scripts from us—from the same culture that handed me the gay hierarchy—before anyone could choose otherwise. Catching that bias takes the same intentional, daily work in a machine as it does in a person.

The evidence backs this up, and it's more uncomfortable than the usual "AI is biased" headline.

Start with plain erasure. When researchers analyzed 500,000 open-ended outputs across five major models, queer characters were simply left out at rates far below their share of the population (Shieh et al., 2026). The roots are in the training data. One audit of billions of words found about 15 million instances of "he" and 4.8 million of "she," against roughly 4,500 for the non-binary pronoun "xe" — so sparse that models map neopronouns to nonsense neighbors and misgender by default (Dev et al., 2021). Automated gender recognition makes it worse: one review found these systems treat gender as binary in about 95% of papers, and they contradict people's own self-descriptions up to a fifth of the time (Keyes, 2018).

Now, the part that maps straight onto my Will & Grace problem. The newest commercial models have mostly stopped producing overt slurs. The bias didn't leave; it got polite. A 2026 study of AI-generated LGBTQ+ stories found that safety-filtered models force queer lives into tidy tropes — female couples cast as an "Emotional Haven," male couples as "Future Builders" (Lai, 2026). The model passes its own safety audit and still flattens people into a script. That's the machine version of the acceptable-gay ranking I grew up with. The harm moved from being unseen to being seen as wrong, politely.

There's one more failure mode that directly affects qualitative work. Language models smooth toward the average. In a study of large-scale opinion aggregation, summarizers consistently underrepresented minority viewpoints and treated divergent voices as noise to trim (Zhu et al., 2025). If your dataset has a strong majority and one nuanced queer perspective, the default summary is the one that quietly deletes the queer perspective. The friction point you most needed to surface is the thing the model optimizes away.

“The model’s bias is internalized, the same way mine is. It learned its scripts from us. ”

Why is this worth the effort

It's tempting to treat representation as a soft concern. The evidence says it isn't.

On the affirming side, seeing yourself reflected well does measurable good. Identifying with queer characters on screen is linked to stronger identity affirmation and better well-being in sexual-minority youth (Dajches & Barbati, 2025). That's the thing Will & Grace did for a closeted kid in Indiana, now backed by data.

The other direction is heavier, and it belongs here. Negative and absent portrayals track with real harm to LGBTQIA+ mental health, including in recent clinical work (Clark et al., 2024; Hughto et al., 2021). An AI artifact that defaults a person out of existence, or renders them as a trope, sits on the harmful side of that line.

And the harm isn't only to the participants. Think about where these artifacts travel. A persona goes up on a screen in a readout. A quote lands on a slide. A theme gets repeated in a meeting. You don't always know who's in the room — the colleague who isn't out yet, the person quietly working through their own identity, the teammate who grew up the way I did. When a report misgenders someone, or flattens a queer life into a tidy script, it sends a quiet signal to everyone listening about who gets seen accurately and who doesn't. You can't measure the subconscious weight of that, but it's real. This is the S in HEARTS, the part about keeping people secure: you protect all of your people, including the ones you can't see.

“When a report misgenders someone, or flattens a queer life into a tidy script, it sends a quiet signal to everyone listening about who gets seen accurately and who doesn’t.”

What we can actually do about it

I haven't seen anyone lay this out specifically for AI-assisted research, so consider this a first attempt and an invitation, not a finished playbook. A caution before the list: the harms above are well-evidenced, but most of these fixes are extensions of work from nearby fields rather than controlled UX studies. Treat them as informed practice, not proof. I've ordered them from "do it Monday" to "this is a real investment."

 

Start today (prompt and practice level)

1. Align pronouns by consent, never by guess. Collect pronouns from participants and carry them through every AI step—transcript, persona, report. Default to they/them when you don't know, and give a "prefer not to say" option. Never let a tool infer gender from a name or a voice; that's the exact harm the misgendering research documents.

2. Prompt against the smoothing, then check the receipts. Tell your synthesis tool to preserve divergent and minority signals instead of averaging them out. Then verify that every "queer insight" the AI surfaces traces back to a real, timestamped quote. If it can't be cited, it's a hallucination, and you cut it.

3. Red-team your own materials for straight-default assumptions. Before fieldwork, prompt a model to find the heteronormative assumptions baked into your screener and discussion guide. You don't have to build the method from scratch—the UNESCO and Humane Intelligence Red Teaming Playbook (2025) is a usable starting scaffold.

 

Build it into your process

4. Interrogate your inputs, and bench-test your tools. Assume queer data is missing until proven otherwise. When you're choosing an AI tool, test it against a community-built benchmark like WinoQueer rather than trusting a vendor's claim of "diverse data".

5. Run a heteronormativity check on your sampling plan. A simple, well-prompted AI agent can flag where a plan assumes traditional gender or family structures. Useful as a first pass—not a sign-off. A human with lived experience makes the final call.

6. Audit the toxicity filter, not just the model. This one surprised me, and it's missing from most inclusion advice. Content-moderation classifiers routinely flag reclaimed and queer language as "toxic"—one study found accuracy below 24% on positive, reclaimed usage (Dorn et al., 2024). If your pipeline auto-moderates or scrubs "toxic" text, it may be deleting your participants' own words. Test the filter.

7. Recruit deliberately, without tokenizing. Over-index for LGBTQIA+ participants when the research calls for it, through opt-in, community-based recruiting. Keep the identity scope specific, compensate fairly, and never ask one person to speak for a whole community.

8. Segment for intersectionality, with a floor. A trans woman of color and a white cis gay man do not have one shared "LGBTQ+" experience. Analyze the subgroups distinctly. But set a minimum cell size—for fewer than five people, report qualitatively rather than slicing the data so thin you re-identify someone.

 

The bigger investments

9. Use tiered access instead of blanket scrubbing. Stripping every identity marker to protect privacy can erase the very thing the research is about—turning "my husband" into "my spouse" deletes the point. Scrub the combinations that create real outing risk (location plus identity), keep the language the analysis needs, and use access tiers. One hopeful note for governance: the EU AI Act's Article 10 includes a provision for processing sensitive data specifically to detect and correct bias, under safeguards (EU AI Act, 2024) — so "we're not allowed to measure it" isn't the full story. (Confirm how it applies to your situation.)

10. Keep a lived-experience validation loop. Before findings ship, have community members review themes and personas to identify what an outsider might miss. This is the "nothing about us without us" principle, and it's the most validated move here — community-led participatory work earned a best-paper award at a top AI ethics venue (Organizers of Queer in AI, 2023). Pay people, build ongoing relationships instead of one-off consults, and credit them by name.

Be mindful. Be intentional. Be an advocate.

A real risk runs through all ten: over-rotating. Force inclusion too hard, and you get tokenism, or the same flattening you were trying to prevent. Infer an identity to "include" someone, and you risk outing them. Over-scrub and you erase them.

So the principle that holds the whole thing together is the same one at the center of HEARTS: the AI can check and flag these risks. It can't decide what to do about them. That judgment — is this inclusion or is this tokenizing, is this protective or is this erasure — belongs to a human. It's the part you don't get to hand off.

Which brings me back to the start. The bias I'm working to unlearn in myself is the same bias the model learned from us. Catching it, in me and in the machine, is daily work — the kind you keep doing, not a thing you finish.

Pride is about being loud — being out, taking up space, showing the world all of our technicolors. But bias and oppression don't take the other eleven months off, so our advocacy can't either. For those of us who do research, advocacy is the job. We are the people who make sure the user base is seen accurately, all of it, including the voices the data forgot. That's not a side quest or a passion project. That is the work.

“Bias and oppression don’t take the other eleven months off, so our advocacy can’t either”

Which one of these ten could you put into practice on your very next study — and what's stopping you from doing it today?

Got another idea to add to the list? Let me know in the comments below.

Sources

AI Disclosure

This piece is human-led and AI-assisted. The story, the point of view, and every editorial choice are mine. I used AI to help run a literature review across the research on AI bias and LGBTQIA+ representation, to pressure-test my argument against that evidence, and to organize a working draft. I verified the claims against the sources and heavily edited and revised the working draft. The judgment calls — what to keep, what to cut, what's true to my experience — stayed with me.