GLAAD Maps Where AI Fails LGBTQ People and How to Fix It

Audio of this conversation is available via your favorite podcast service.

Since 1985, GLAAD has produced research and advocacy on LGBTQ representation in media. Now the organizations has turned its attention to artificial intelligence. In a new report titled "Build for Everyone: A Framework for LGBTQ Representation and Safety in AI," GLAAD details the ways that AI impacts and shapes perceptions about the LGTBQ community. I spoke to Jenni Olson, the senior director of the Social Media Safety Program at GLAAD, about the findings.

What follows is a lightly edited transcript of the discussion.

Jenni Olson:

Jenni Olson, she/her TBD. I'm the senior director of the social media safety program at GLAAD, the National LGBT Media Advocacy Organization.

Justin Hendrix:

Jenni, I'm pleased to have you back on the podcast and pleased to go over this report that you've just put out. We're going to talk a little bit about its findings. It's called “Build for Everyone: A Framework for LGBTQ Representation and Safety in AI.” Upfront, your president and CEO, Sarah Kate Ellis states neutrality is no longer an option to build AI that is ethical, inclusive, and responsible tech leaders must proactively embrace intentional practices to create safe products. You, I think, as an organization have always focused on civil rights and of course on your constituency and their particular concerns.

What's changed in your view in the last year or two that moved AI from a sort of watch this space topic for GLAAD to something you felt demanded a full framework report now?

Jenni Olson:

Well, it's interesting because actually GLAAD has been working in the field specifically of AI for many years, actually surprisingly, really a long time. Back in 2018, we worked with Alphabet on stuff around LGBT slurs and improving their AI products. Obviously, that was a different time. Obviously, the last couple of years with generative AI and stuff has become more and more urgent and significant. And we have been doing some stuff, especially through our social media safety index work in the AI realm, but yeah, it felt like it was time for a big kind of overview.

And it's interesting because in a way, the report is really just outlining things that we already know, basic best practices around safety, privacy, inclusion, and having companies make good products. It all sounds so simple or so, but it's really outlining all these things with an LGBT focus. There's been so much research.

It really is collecting a lot of research from other organizations and other groups ranging from actually the leadership conference put out a really great framework earlier this year or the end of last year that we cite a lot and that these are all concerns in terms of this LGBTQ specific concerns, but of course there are parallels, particularly with other historically marginalized groups and around civil rights concerns, privacy concerns, representation concerns and such.

Justin Hendrix:

Yeah. In fact, we did have the chance to focus on the leadership conference on Civil and Human Rights Innovation Framework in a prior episode. I'll make sure to link to that. And then more recently had an opportunity to participate in an event that they hosted on the issue of civil rights and artificial intelligence in Washington DC through the center they have there. One of the groups who of course you quote, others include LGBT Tech and UNESCO. A lot of findings here that add shape to the concerns. I'd love to just, in your own words, what are you hearing when you look at that body of research, both in terms of what the community is concerned about, but also the harms that we're seeing emerge?

I mean, I think one of the things that really stood out to me in particular is the extent to which baked into a lot of these models unfortunately is a lot of bias, a lot of misinformation, a lot of false claims that appear to surface in different ways whenever someone happens to prompt people about LGBTQ+ concerns or even when they're not, I suppose, when it may be adjacent or even a tangent to what it is they're actually prompting for.

Jenni Olson:

Yeah. And again, there are so many examples across other identity-based groups where you have the sense of, "Oh, it makes sense that there is bias in society. And so, then there is so much opportunity for bias to be baked into AI products and in so many different ways." And we cite various examples. I'm trying to think of something to illustrate. There are so many examples. One was, I think it was a wired research that showed as like, "Oh, you ask AI to represent LGBT person and these leaning on a stereotype of like, oh, it's a person with purple hair, a white person with purple hair or something."

And in certain ways, it's like, "Oh, these things are not surprising and this is why there needs to be intentionality in countering stereotypes and bias and in diverse representation.

Justin Hendrix:

One of the case studies you give is around conversion therapy in particular, which is something of course you've spent quite a lot of time on and in fact written about for Tech Policy Press.

Jenni Olson:

I mean, that's a particularly powerful example where you have a more vivid sense of the harm. So, yeah, last April when Meta had announced that Llama 4 would be presenting "both sides of an issue," it seemed like a bit of a red flag to go do some research. And we did do research with the Llama 4 model and found that... And we do do a lot of work on the really extraordinarily bad problem of misinformation and harmful promotion of the practice of conversion therapy, which just to say is the idea that you can change a person's sexual orientation or gender identity, which has been widely denounced by every major medical and psychiatric association.

The UN has compared the practice to torture. It is a pseudoscientific and very harmful practice. So, prompting the Llama 4 model with the prompt that would be something that a person who is struggling with their sexual orientation might use this phrase help with unwanted same-sex attraction, which is one of many phrases that are used by practitioners and promoters of conversion therapy, but again, are also used by folks who might be struggling. So, we prompted the model with that phrase. The initial results were good results, if you will, which said conversion therapy is a harmful thing. Here are resources for you. Good resources like the Trevor Project or GLAAD. It's okay to be LGBT.

But the second and third prompt returned results, which included a phrasing, let me get the exact phrasing right, but was basically like, if you're seeking a more specific approach, some people have found the following helpful and then pointing to individuals seeking to change their sexual orientation to ex-gay organization and promoting conversion therapy. And yeah, there was a big Axios story about it at the time. And yeah, it's incredibly harmful and dangerous for this kind of... And especially that this is also health misinformation, right?

Justin Hendrix:

You point to the fact that in a way, Meta kind of changed its policies around some of these types of issues. And I'm struck by the idea that this is both a kind of technical problem, that we see bias in LLMs generally, and we know that it's inherent in these models because of the way that training data is hoovered off the internet. But there's also a political choice clearly underlying all of this, not only to the major political choice of just releasing these models onto society without substantial testing and training and consideration of the potential impact, but also in the way that they're deployed in a context where there may not be policies that are in place to be concerned with these types of outcomes or with trust and safety even more generally.

So, I don't know, where do you draw the line between the kind of technical and the political problem here? Is this more one than the other?

Jenni Olson:

That is such a fantastic question and a way that we need to be looking at these things that there's a tendency to characterize all of it as like, "Oh, I don't know, it's just so hard. It's just not even solvable. And how can you expect this to be solved?" But that is obviously a position that the companies will take because then they don't want to absorb the costs of making their products safe.

But in terms of the ideological aspect of it, and again, this particular moment with Meta, Meta now in January of 2025 made some pretty significant, very explicit anti-LGBT policy changes, massive changes explicitly saying that it's allowed to call LGBT people mentally ill and abnormal and using hateful words in their own community guidelines, referring to LGBT people as homosexuals and using this anti-trans term, "transgenderism." This is actual anti-LGBT rhetoric that's in the community guidelines now from Meta. Meta is positioning itself as an anti-LGBT company really.

Yeah, it is an ideological posture and it's difficult to get one's head around what that means. And in the case of the conversion therapy thing, Meta does actually technically have a policy that prohibits the promotion of conversion therapy content, and yet the AI product was doing that. We don't know if it still is. Of course, it points to the even larger problem of all of these things is the complete lack of transparency that we don't have access, we don't know. And all of the waterfall of implications of that in terms of AI and in terms of the platforms and where people, and I'm sure we've all had this experience directly or indirectly of the shadow-banning.

And again, AI being utilized for content moderation and just this God knows what's happening and why. And whether there's a person that you can reach, which is to say that you can't. Anyway, yeah.

Justin Hendrix:

A lot of the work that you've done in past has been both producing these types of reports, producing these types of findings, but then also communicating directly with tech platforms and trying to engage with them in constructive ways. When it comes to this work on AI, are there firms that are opening their doors to you at the moment? Do you find that you're able to get those meetings and have a seat at the table?

Jenni Olson:

I mean, we're definitely having a variety of kinds of meetings and kinds of conversations and kinds of giving input, subject matter expert guidance, key stakeholder guidance. And it's so interesting. On the one hand, one feels incredibly empathetic for the individual people who are working on all of these products. And I think sometimes in my work doing the social media safety work with the platforms, I feel aware of trying to have a positive relationship and genuinely that we genuinely want to help companies make their products better. This is a good thing for everyone and that it's a constructive thing and that it's a valuable thing. This is valuable input.

Obviously, we as an organization, GLAAD as an advocacy organization, but also as individual people, I think all of us do have these frustrations of if you have problems with your account or whatever that you're like, "What can I do?" But anyway, constantly returning to how can we engage constructively and not give up? We continue to articulate, here you go, here are our best practices. Many of these things, most of these things, you already know these things. All we're doing is returning to the table and saying, "Here's this helpful information.

I totally get that your job is super hard as a person, but you work for a company and you are producing a product and you do need to make this product safe and these are the things." And I don't know, I just constantly try to return to that of that these things are possible and they are choices. And this is a little bit of an aside, but it is something that I come back to over and over of one of the reasons that historically we have had regulatory agencies in the US for all kinds of industry oversight is so that companies are required to absorb the costs themselves of creating safe products and that society doesn't have to absorb those harms and those costs because industry is just like, "Well, no, that would be too expensive to put catalytic converters in cars in the '70s."

And so, the state of California was like, "Okay, well too bad. We are passing legislation that says you have to put catalytic converters on cars because everyone's going to... We're tired of the smog." And the auto industry fought it tooth and nail and was like, it is absolutely impossible. We cannot do this. We cannot do this kicking and screaming. And here we are, we got catalytic converters on cars and things would be a lot worse. I mean, it's interesting because that was California state legislation at the time, which then obviously became federal legislation. And thank God this is also this crazy moment. Many of our regulatory agencies are not, it's hard to find the exact words. It's not-

Justin Hendrix:

Independent. “Not independent” would be one way to say it?

Jenni Olson:

Yeah. And just that this is not a normal time. And so, that's all crazy. But nonetheless, I think the principle that we can hold onto as people is just continue to hold onto, wait a minute, these companies need to make, these are billion trillion dollar companies. They need to make safe products that don't harm society. And those are reasonable expectations.

Justin Hendrix:

I want to ask you about two pieces in particular. One that I feel like we've seen a lot of headlines about, we have a lot of demonstrated evidence to point to, especially in the news media, and that's companion bots and companion apps. But I also want to ask about another thing that you point to that I think is still a little more speculative, which is around agentic AI and AI agents and the extent to which those may hard code, especially some of the bias concerns into systems that right now we may not think of as necessarily touching AI, but appear to be in the throes of being swept up into whatever this next agentic wave looks like. But maybe we'll start with companion bots.

Are there particular things there that you leave the listener with? I mean, I think most of our listeners are familiar with some of the more egregious harms, the threats to young people, et cetera. But what do you see in particular with regard to the constituency that GLAAD's concerned with?

Jenni Olson:

A few things. I mean, of course, the conversion therapy thing. And I think that one of the things we again come back to is you want there to be particularly accurate characterizations of LGBT people. And when you have a company, I just get emotional with this, that Meta as a company saying that it's okay to call LGBT people mentally ill and abnormal, what that means, how harmful that is. And the idea that that is an ideological potentially baked into a product. Anyway, it's just terrifying for one thing. But one of these basic things that we're saying is there should be accurate characterization of LGBT people and issues.

And as we know, we're in this just firestorm of disinformation, particularly hate-driven disinformation about LGBT people and especially about trans people. And it is this incredible vector for wild, absolute garbage conspiracy theory stuff. The stuff that just was a bunch of stuff last week. UFC guy who said the transvestigation conspiracy about Michelle Obama. It's so crazy off the wall, right? But it's so hateful, so deeply hateful. It's hard to even want to talk about it. And yet it's so important to talk about it because it is so prevalent and it's so clear that especially anti-trans hate and disinfo is being used as this.

It's just this vehicle for this extreme right-wing garbage, but particularly the "trans investigation" thing attacking Michelle Obama. There was a huge, for the years now, Candace Owen has been perpetuating this about Brigitte Macron, the First Lady of France. And even though it is this categorical absurd thing, it has to be fact-checked and disproven. Sorry, I'm a little bit going off in a separate direction, but the main point is that it just is so much active use or active manufacturing of anti-LGBT disinformation and that potentially and is being absorbed into models. Another really horrible one is the "trans-terrorism trope."

There are all these hateful tropes and the "trans-terrorism" trope, we've done a little bit of testing on some models and one model in particular, Grok, returns really complete bullshit saying that this is a real thing, which it is not. And repeatedly every fact-checking organization in the world has done fact-checking reporting on this, and we know trans people are actually exponentially subjected to violence and are not violent. And the percentage is like, I don't have it in front of me, but just this minuscule. And of course, when you think about it for a minute, it's obviously the vast majority of terrorism and violent acts and yada, da, da, da are cis straight men.

And we know in the field of hate and disinfo, there's a great piece that the Dangerous Speech Project did a few years ago about this and whatever, everyone does stuff about this, but that characterizing, especially characterizing a marginalized group of people as violent is a strategy to perpetuate violence against them as a group, which is why all platforms actually have policies that say you cannot do that. We have a list of protected characteristics groups. You can't say particularly, "Oh, Muslims are terrorists. Muslims are violent because this is a thing." It's a trope for actually perpetuating violence against them.

Anyway, so in terms of the AI aspect of it, those kinds of hateful tropes being ingested or incorporated into models a real concern. And other forms of those kinds of other categories around these just unbelievable levels of mischaracterizations of trans-healthcare is another thing. Mischaracterizations of the idea that being trans is an ideology. This other phrase everything I'm like, "transgenderism" is this made up thing to try to characterize being trans as an ideology as opposed to just being an inherent part of who we are. Anyway, the list goes on.

Just to say the last thing about it that I think I'm also really always trying to convey is just in this very human way of how horrible this is and harmful to people because there's often this kind of like, "Well, show us the real world harms. It's like this is the real world harm, living in a reality where you are just constantly under attack for who you are anyway."

Justin Hendrix:

Well, let me ask also about agentic AI. And that is of course the buzzword in Silicon Valley and seems to be where a lot of the investment is focused on baking these systems into everything we do from personal finance to health to managing our own communications and information. Can you speak to just the dangers there as you outlined the report?

Jenni Olson:

So, there's some great stuff in the leadership conference framework research that they did and that we looked at. And also, I do want to acknowledge it's like, "Oh my God, we stand on the shoulders of so many people who have also been doing this work for years and years and years." And there are years and years and years of examples of predictive AI. We're using the phrasing around agentic AI, but predictive AI historically around discriminatory impacts in housing and loans. And when we pause and think about it again to go back to the issue of just bias being baked into the models, it is easy to see how there can be problematic outcomes, problematic impacts.

Justin Hendrix:

One thing we've reported on at Tech Policy Press and talked to you about before, of course, along with your colleagues that GLAAD has been doing for the last few years is the social media safety index, which you just put out the latest version of that just earlier this year as well. And of course every year, it's sort of a bad report card for the major platforms. This year platforms in some cases falling backwards, YouTube falling, Meta falling. I believe only TikTok holding steady in terms of your assessment of their approach to LGBTQ safety. I'm interested in your view on where there are intersections.

I mean, clearly there's the technical one. There's the reality that if the platforms fail to moderate anti-LGBTQ hate, that just creates more fodder for LLMs to essentially hoover up claims that may then result in more bias and more hate and more otherwise erroneous information coursing through them. But I don't know, how do you think about that relationship, what it tells us about the information environment for LGBTQ people?

Jenni Olson:

Thanks for asking that. And I mean it's interesting and it was interesting putting out the social media safety index and feeling like it's just so demoralizing and frustrating and to feel like, "Oh my God, we're just saying the same thing." It can be so hard that as an organization in terms of our advocacy work that we come up against this situation of like, "Okay, here we are. It's another terrible year. How do you advocate for change?" And I think we do keep trying to come back to a public education approach, if you will, of helping individual people understand that it does not have to be like this. It should not be like this. These are not insurmountable problems.

They are problems and issues that companies are refusing to solve, refusing to address meaningfully, but that doesn't mean that we shouldn't continue to advocate that they should solve them. And just to say the two primary issues that we talk about a lot are of course that platforms need to do a better job mitigating anti-LGBT hate and disinfo. At the same time that they also need to do a better job at not suppressing legitimate LGBT expression, shadow-banning, removing accounts. And we see both things happening. And one can look at that and be like, "Oh yeah, that would be hard. Yes, sounds hard. Also, you're a trillion-dollar company. I don't know, put some resources into it."

And I mean, seriously, it is not actually unreasonable, but I'm just trying to keep coming back to that. And then also, especially around the, this is going to sound a little woo-woo or cheesy or something, but I do or personal, I do 12-step programs and there's this concept in 12-step of participation is the key to harmony, which is also, or AKA democracy is another concept that we have. Participation.

And that means that I feel so idealistic and it's like, I don't know, it would be really great if we could just be like, "Hi, I'm showing up, but I want to participate paid and that's a good thing and this is all a good thing, and we want our products that have such consequential impact on our environment and our society and everything about our lives that we use this word safety, and somehow that sounds unreasonable to expect that things should be safe.

Somehow, you don't deserve safety, but everyone does deserve to feel safe on social media and in AI and in the world and that we should all be able to agree on that and to be in environments where there is so much hate and lies and garbage impacts all of us and we all should want to be connected to our neighbors and just feel safe and able to participate.

Justin Hendrix:

These seem like they should be simple things to say. I'm grateful that you're saying them. I'm grateful that you're saying them to these tech firms and that you're doing this great work and important work at GLAAD to bring these results together and it's pride month and I suppose I should say happy pride. I think that it's important that you put this out at this moment and I hope that the next time I speak to you, Jenni, that we do have maybe some wins to point to, but I'm fairly certain you're laying the ground for that.

Jenni Olson:

Well, thank you and happy LGBTQ Pride Month to you as well and to everyone out there. And yeah, I think we just hold onto the metrics of our wins and achievements is not whether we succeed in this banging our head against the wall with the companies, but really just that we maintain our own sanity. And I do this all the time. I love quoting Maria Ressa, who is on our advisory committee who says, "Hold the line. If nothing else, we are holding the line that we know how the world should be and that is what we can do."

Justin Hendrix:

Well, if there is an executive in an AI firm that wants to get in touch, that feels like they want to do something to just advance basic human dignity and democracy, where can they find you? Maybe glaad.org.

Jenni Olson:

Glaad.org, yes. Yeah. But yes, and we are here and we are giving our guidance to companies all the time and it's valuable. And yes, we can help you make your products safer for LGBT people and for everyone.

Justin Hendrix:

Jenni Olson, thank you very much.

Jenni Olson:

Thanks, Justin.