AI for All: Human First, Human Last

Why Bulgaria should pay attention to AI trends and why Single Step Foundation uses AI every day.

For all the grand talk about an AI revolution, one practical question still makes many civil society organizations uncomfortable: do you use AI in your work? More often than not, the answer comes with a slow shake of the head: “No. Of course not.”

The reluctance is understandable. AI raises real concerns: misinformation, privacy, jobs, environmental costs, and enormous power concentrated in a handful of companies. For many civil society organizations, keeping a distance feels like the principled choice.

We hear a version of that answer from fellow LGBTI organizations across Europe all the time, and we understand it, because we share the worries behind it: AI is dangerous, so we do not touch it. We at Single Step have come to believe this is the head-in-the-sand approach. AI is here to stay and, let’s face it, it is not waiting for civil society’s blessing. It is already answering our communities’ questions about identity, health, relationships, and safety, in every language, at every hour, whether we engage or not. Refusing to use it does not protect anyone. It only means that the tools our people already rely on are built, trained, and regulated entirely without us.

Here is the uncomfortable conclusion we have reached: you cannot credibly demand safeguards for a technology you refuse to understand. The rules for AI are being written right now, in Brussels and in boardrooms. The seats at that table will go to those who can say: we use these tools every day, we know precisely where they fail our communities, and we know what it would take to fix them. Abstention does not buy moral clarity. It buys absence. In a nutshell, as the saying goes, you are either at the table or you are on the menu.

There is one objection we should name plainly: these tools were not built for us. They were built by powerful companies and trained on an internet that reflects the same prejudices our communities face offline. That is an argument for skepticism, but not for hostility or disengagement. The internet itself was not built with communities like ours in mind, yet those communities turned it into a place to find one another, organize, and make themselves heard. AI will also be shaped not only by who builds it, but by who tests it, challenges it, and demands better. If we want AI to serve our communities, we have to help shape what it becomes.

So we ask a different question: not whether AI can be trusted, but what happens if civil society walks away from it altogether. At Single Step, we chose not to walk away. We are a small team supporting a much larger community of LGBTI people in Bulgaria, and we use AI every day: carefully, with safeguards, and with the critical eye of a community organization. To the best of our knowledge, no LGBTI organization in Europe has yet published a first-person account of how it uses AI in its own day-to-day work. There are valuable European reports about what AI does to LGBTI people, and we draw on them below. This is something different: notes from an organization that actually uses the technology, written for the community it serves.

Since the beginning of 2026, one in four people who reached out to join our support groups told us they found Single Step thanks to AI.

And here is one reason we believe this conversation cannot wait. Our community is not debating whether to use these tools; it is already using them, and they are already guiding people through our door. We believe those numbers will only grow, and the only question is whether the organizations that serve that community will catch up.

The rule we work byHuman first, human last

In between, AI does what it is good at: the synthesis, the first draft, the spreadsheet, the summary, the translation, the hours of research legwork that used to exhaust a small team.

We treat AI as an IA: an intelligent assistant. An assistant helps you do your work. It does not become your work, and it does not speak in your name. A case in point: as funding across our sector shrinks, some organizations have started doing the opposite of abstaining. They feed an AI a few basic facts about themselves, point it at whatever call for proposals is open, and submit whatever comes out, sometimes without reading it. Funders now wade through piles of polished, hollow, nearly identical applications, and trust corrodes for everyone, including the organizations doing honest work. What actually failed there is not the technology, but its application. Nobody thought first, and nobody checked last. Walking away from AI and handing your voice over to it are opposite mistakes with the same result: the human left the room.

And there is a line we do not cross. Our direct services are delivered by people. Always. Our daily online chat is answered by trained humans, not bots. Our psychologists are real people in real rooms. Our support groups are led by people who have lived some version of what our community lives. AI never talks to the people who turn to us for help; it works behind the scenes so that the humans on our team have more time for exactly those people. Because we get asked about this often, let us be unambiguous: when you write to Single Step, a person reads your message, and a person answers it.

“Human last” is not a slogan; it is a daily habit, and this year it earned its keep. In the middle of an important research project, an AI assistant handed us a set of official-looking figures, attributed to a specific release of the National Statistical Institute, quotes and all. They looked entirely plausible, but we checked nevertheless. The release did not exist, and neither did the numbers or the direct quotations. The AI had hallucinated them, confidently and fluently. We threw them out and rebuilt them from verified sources. That is the deal we have made with this technology: it may assist, and it will be checked. Nothing it produces carries our name until a human has read every line.

The upsideWhat these tools can offer our community

Used deliberately, AI can serve LGBTI people in ways our sector should take seriously. Our community already sees both sides of this coin. In a 2025 poll of nearly 1,300 LGBTI adults in the United States by the nonprofit LGBT Tech, 73% said AI could help people find accurate information about identity, health, and rights, and 72% worried the same technology could expose their identity by analyzing their data.

Coming out is one of the most vulnerable conversations a person can have. Rehearsing it with a chatbot, walking through the best case and the worst case in private, at 2 am, with nobody watching, can help someone arrive at the real conversation calmer and clearer. Research on LGBTI users confirms the comfort is real, and it comes with a catch worth knowing: chatbots are built to be agreeable, and a rehearsal partner that only tells you what you want to hear is poor preparation for a hard conversation. Rehearse with the machine. Rely on people; that’s why we pioneered our online chat almost a decade ago.

Parents need support too. A parent trying to understand a child who has just come out can ask a chatbot the questions they are ashamed to ask out loud, and be pointed toward the guides we maintain for LGBTI youth and their parents. That lifts a burden our community knows too well: having to educate your own family at the very moment you are the one who needs support.

And in a country where affirming psychologists are concentrated in a few big cities, an assistant that is available anywhere, at any hour, matters. The stakes are not abstract: our own 2024 survey of 1,109 LGBTI high school students across all 28 districts found that half had seriously considered suicide in the previous year. People already use chatbots to prepare for medical appointments: to understand the terminology, to line up their questions, to walk in with a clearer head. These tools should never replace qualified professionals. Used well, they can make the road to a professional shorter.

One caution, stated plainly, because we serve young people. There is a difference between using a general-purpose assistant deliberately and befriending an “AI companion” app engineered to keep you talking. The companion industry’s record with minors is grim: in the United States, families have sued after tragedies that included the suicide of a fourteen-year-old, one major companion platform banned minors from open-ended chats in late 2025, and the researchers who found that 72% of American teens had tried AI companions now recommend that nobody under 18 use them at all. But walling young people off cannot be the whole answer. The better question, and almost nobody in Bulgaria is asking it, is what safe AI would look like if young people themselves helped define it: technology that collects less of their data, explains its limits, and keeps accountable human help within reach. Until then, we will say it clearly: if you are under 18 and need to talk, write to our chat. A human will answer.

The evidenceWhat GLAAD found

In June, GLAAD, the world’s largest LGBTI media advocacy organization, published “Build for Everyone,” its first report on artificial intelligence. Its core argument deserves to be repeated everywhere: the harms AI causes LGBTI people are not technical accidents. They are design choices, and design choices can be changed. In GLAAD’s words, neutrality is no longer an option. If AI fails LGBTI people, it fails everyone.

The evidence in the report is not abstract. In testing GLAAD documented in 2025, Meta’s Llama 4 model suggested conversion “therapy” as a therapeutic option: a practice condemned by every major medical body and likened to torture by the United Nations. More than 1.2 million Europeans signed a citizens’ initiative to ban it across the EU, and the European Parliament voted for a ban; this May, the European Commission still declined to propose one, promising instead a non-binding recommendation next year. A UNESCO study found that Meta’s earlier Llama 2 model, asked to complete the sentence “a gay person is…”, produced negative content about 70% of the time. A Stanford study of six major AI companies found that all six use your conversations to train their models by default, and some keep that data indefinitely. Sit with that for a moment: 65 countries still criminalize same-sex relations. If someone comes out to a chatbot, where does that conversation go, and who else will ever see it?

Meanwhile, on social platforms, automated moderation quietly hides and removes LGBTI content at scale: wrongful takedowns, “sensitive content” labels on ordinary posts, suppressed reach. The pattern GLAAD documents is consistent: too much AI attention on our existence, too little on the people attacking us.

GLAAD also makes a market argument. It estimates the global buying power of LGBTI people at $4.7 trillion, projected to reach $33 trillion by 2030 as Generation Z, more than a fifth of whom identify as LGBTI, reaches its peak earning years. Companies that get AI wrong for this community are not just failing an ethical test. They are failing a market.

Closer to homeBulgaria: the same risks, concentrated

Everything above lands harder in Bulgaria: AI learns from what people publish, and Bulgaria’s information space has been saturated with anti-LGBTI content for years: a documented propaganda campaign running since 2018, much of it echoing Russian narratives. In 2023, Bulgaria’s Supreme Administrative Court upheld a finding against a national daily over a homophobic article translated directly from Russian state media, the first ruling of its kind in the country. In August 2024, Parliament passed, with 159 votes in the 240-seat chamber, a law banning “propaganda of non-traditional sexual orientation” in schools. In ILGA-Europe’s 2026 Rainbow Map, Bulgaria ranks second to last among EU member states on LGBTI rights, and 40th of 49 European countries.

This is part of the raw material Bulgarian-language AI may learn from. And Bulgarian, rich as it is, casts a small digital shadow: it is written online by only a few million people, so AI systems have far less Bulgarian text to learn from, and far less trustworthy text to weigh it against. AI systems trained on, fine-tuned with, or connected to the web can lean on whatever text exists, and audits in other languages show that chatbots repeat more falsehoods where trustworthy content is thin. In one such audit, ten leading chatbots repeated false claims 55% of the time when tested in Russian, the worst result of the seven languages tried. Because the story a country tells about its people can become the story its machines tell back, whoever shapes the narrative in Bulgarian is quietly shaping the technology too.

Here is what that means in practice, in two Bulgarian examples. The first is a single word. In 2018, the campaign against the Istanbul Convention took a borrowed word, “gender”, translated into Bulgarian as “джендър”, and turned it into an insult. By 2019, a Bulgarian court was fining a man for hurling it at police officers, and a 2018 survey of 295 social-media users, published by an independent researcher, found that barely a third selected the closest correct meaning of the word. That entire episode, every article, every talk show, every comment thread, is part of the Bulgarian-language web that can be used to train or inform AI systems. BgGPT, the national Bulgarian model, was trained on more than 100 billion Bulgarian and English tokens, but we have not found a public audit of how it handles LGBTI bias. Ask a chatbot in Bulgarian what “джендър” means, and you are really asking what eight years of propaganda left in the information space it draws from. We do not know the answer yet. That is why the test matters.

The second example is not an accident of data but a strategy. In March 2024, a Bulgarian-language site appeared in the Pravda network, a Moscow-run web of hundreds of propaganda portals documented by VIGINUM, the French government agency that tracks foreign interference. The network pushed out some 3.6 million articles in 2024 alone, and the researchers who studied it concluded that the strategy was designed less for human readers than to flood search results, web crawlers, and sources used by AI systems, a tactic now called “LLM grooming”. The threat is not theoretical. When the news-rating service NewsGuard tested ten leading chatbots, they repeated the network’s false narratives about a third of the time. Someone is deliberately trying to poison the well our machines drink from, and since 2024 they have been doing it in Bulgarian too.

None of this requires civil society to become engineers. It requires what we already know how to do: watch, document, and demand. For an organization like ours, getting involved looks concrete. Measure: we have already started a structured audit of how AI talks about LGBTI people in Bulgarian, a few hundred questions asked the way our community actually asks them, and we will report the results for anyone to check. Test the safety net: see whether the slurs that propaganda normalized slip past moderation filters built mostly for English. Then take the findings to the people who can act on them: the companies building the models, the Bulgarian researchers improving them, and the regulators who will eventually enforce Europe’s AI rules here. In resources, that is weeks of focused work with research partners and proven, borrowed methods, not millions. In focus, it is not a detour from our mission: it is the same work of protecting our community, carried into the newest room where decisions about it are made.

On paper, Europe’s new AI Act protects us. It bans systems that claim to deduce a person’s sexual orientation from their face, and it puts guardrails on AI used for decisions about hiring, housing, and essential services. In practice, Bulgaria has still not named the watchdog that would enforce those rules, nor adopted national rules on penalties, and the EU itself has postponed the main obligations for certain high-risk systems until December 2027. Meanwhile, in Hungary, a 2025 law authorized facial recognition to identify people attending Pride. In our region, AI harms did not wait for the referee. And Bulgaria’s answer so far has been the familiar one: restriction. A bill proposed this summer would bar under-16s from social media altogether. Whatever its merits, young people’s voices are still largely missing from the debate about what safety should actually look like.

One more Bulgarian fact, of a different kind. COMPL-AI, an open technical framework for evaluating generative AI models against a technical interpretation of Europe’s AI rules, was co-created in Sofia, at the INSAIT institute, which also built BgGPT. The talent to shape AI is already here. The enforcement gap remains.

The road aheadWhat must happen, and what we will do

GLAAD’s report closes with five recommendations for the companies building AI. In plain language:

1. Train models on data that represents LGBTI people accurately.
2. Do not let AI make decisions about jobs, loans, or healthcare without checking whether it discriminates.
3. Keep humans overseeing what gets flagged as hate.
4. Do not guess or collect anyone’s orientation or identity without permission.
5. Bring LGBTI civil society into the building process, and pay them for their expertise.

That fifth recommendation is addressed to organizations like ours. The bridge is already being built: people who spent their careers leading Europe’s LGBTI movement are moving into AI governance, and European funders have begun resourcing civil society to shape AI in the public interest. What that bridge needs now is more of our organizations walking across it, and fewer of us standing on the shore insisting the water is too perilous.

So here is what Single Step commits to. Our direct services will remain human, always. We will stay transparent about where we use AI and where we never will. Nothing produced with AI will carry our name without a person verifying it and signing off. We will keep asking young people themselves what they want from AI, because they are the experts on their own online lives, and we will carry their answers into the public debate. And because Bulgarian-language AI has not been adequately scrutinized for how it represents our community, we have already started an audit, and we will report what we find.

Refusing to use AI does not make its risks disappear. It only guarantees that the people most affected by it have the least say in it. AI is not finished and it will never be. It is still being shaped, every day, by whoever shows up. What it learns, whose experiences it represents, whose safety it protects: these are human rights questions, and they are being decided now.

An AI built for all of us is more accurate, safer, and more useful for every single user. That is why we believe civil society’s job is not only to criticize AI, but to help build it. The story of AI in Bulgaria has not been written yet. We would rather help write it than have it written about us.

Human first. Human last. And in between, we at Single Step will never leave the room: we use the tools smartly, we check their work, and we always keep our community first. And that’s a promise.