Artificial intelligence and transgender people

Artificial intelligence is not neutral. It learns from data produced by human beings, reflects the social structures in which it is developed, and amplifies the patterns it finds in training datasets. For transgender people, this means facing a technology that can be both a tool of liberation and a mechanism of oppression — often at the same time, on the same platform, within the same algorithm.

A 2025 study presented at the ACM Conference on Fairness, Accountability, and Transparency (FAccT) measured for the first time attitudes toward AI among marginalized populations in the United States: nonbinary people scored the lowest in positivity toward AI (3.84 out of 7), followed by transgender people (4.12), compared to a mean of 5.12 among cisgender people [1]. This is not generic distrust. It is the cumulative effect of concrete experiences of algorithmic discrimination.

This wariness has solid foundations. But the story of AI and transgender people is not only a story of harm.

The risks: when the algorithm misgenders

Facial recognition and binary classification

Automated Gender Recognition (AGR) technology is perhaps the clearest example of how AI can be structurally incompatible with trans identities. AGR systems derive gender from physical features — jawline shape, cheekbones, the presence or absence of makeup — and reduce it to a binary classification: male or female. A 2019 investigation by Quartz tested facial recognition services from Amazon, Microsoft, and IBM on photos of trans and nonbinary people, finding significantly higher error rates compared to cisgender people [5].

The problem is not a technical flaw fixable with more diverse datasets. As Access Now emphasized in a campaign conducted with All Out and supported by researcher Os Keyes, the very goal of these systems — automatically classifying people’s gender from their face — is incompatible with fundamental rights [2]. You cannot “fix” a technology that denies self-determination by definition. Adding a third category is not enough: the very principle of deducing gender from physical appearance is wrong.

And the consequences are not abstract. AGR systems are used in public surveillance, airport security, and identity verification for financial services. A trans person who is misgendered by an automated system at an airport can be subjected to additional screenings, questioning, and delays. A trans person trying to open a bank account with a biometric verification system can be blocked because their face does not “match” the gender on their documents.

The binary in interfaces and data

The problem extends beyond facial recognition. Most computer systems — from online registration forms to electronic medical records, from social media profiles to government databases — are built on a binary gender architecture. The “sex” field offers two options: M or F. Sometimes neither.

For nonbinary, genderqueer, or genderfluid people, this architecture is a form of systematic erasure. It is not possible to exist in a database that has no field for you. And when AI systems are trained on these data, they learn and reinforce the binary: they generate outputs that do not contemplate the possibility of a gender outside male and female.

A 2025 study on AI and personalized medicine, conducted through focus groups with trans people, highlighted how healthcare AI systems tend to reinforce binary gender stereotypes and marginalize trans identities, with concrete implications for the quality of care received [8].

Content moderation: the invisible censorship

If you create online content and you are a transgender person, you have probably experienced the automatic removal of a post, a video, or a story. Not because it violated platform rules, but because an algorithm classified your content as “sexually explicit” or “adult.”

An academic study published by ACM in 2021 documented this phenomenon with concrete data: transgender people are among the three groups of social media users who experience disproportionate content removal (along with political conservatives and Black people) [4]. Trans content is removed as “adult content” despite complying with platform guidelines. Educational videos about gender identity, transition stories, even simple selfies are classified as “explicit” by moderation algorithms — while identical content posted by cisgender people passes without issue.

Research conducted at UCLA identified systematic patterns: the terms “transgender” and “nonbinary” attract more automatic flags than other terms [13]. These are not random errors but a structural bias in the training data of moderation models. The GLAAD Social Media Safety Index of 2024 recommended that platforms not rely excessively on AI for moderation, suggesting that automated systems should be used to flag content for human reviewers rather than automatically removing it [3].

This algorithmic censorship has real effects: it reduces the visibility of trans people online, limits access to information and resources, and contributes to a sense of digital isolation that compounds physical isolation.

Deepfakes and automated disinformation

Generative AI has introduced a new category of risk. Deepfakes — artificially generated videos, images, and audio — can be used to create non-consensual content that disproportionately targets already vulnerable people. Estimates indicate that approximately 90% of sexually explicit deepfakes target women, and transgender people are particularly exposed to this form of digital violence.

But the risk is not limited to pornographic deepfakes. Generative AI can produce disinformation at scale: fabricated articles attributing statements never made, invented scientific studies, manufactured testimonials. In a context where misinformation about transgender people is already a widespread phenomenon — as documented in the article on common myths about trans people — automation makes the production of false content faster, cheaper, and harder to identify.

A GLAAD investigation also highlighted how generative AI systems tend to represent LGBTQ+ people in stereotypical ways — young, white, with purple hair — reproducing and amplifying cliches rather than reflecting the real diversity of the community [3].

Hiring algorithms and workplace discrimination

AI systems used in recruitment present another front of risk. A 2024 study from the University of Washington demonstrated that large language models (LLMs) favor names associated with white people in 85% of cases and names associated with women in only 11% of cases [10]. Many resume screening systems do not recognize gender identities outside the male/female binary and can automatically downrank applications from people whose profiles do not conform to historical hiring patterns.

For transgender people, this means facing an additional layer of discrimination: not only the human bias of the recruiter, but also the bias encoded in the algorithm that filters resumes before a human being ever sees them. A trans person who has legally changed their name may be penalized by a system that detects discontinuities in their resume. A nonbinary person whose gender does not fit the software’s predefined categories may simply be excluded from the process.

The opportunities: when technology opens doors

Telemedicine and access to care

If the risks of AI are real, so are the opportunities. AI-assisted telemedicine is changing access to gender-affirming care, especially for those living in geographic areas where specialized services are absent or in contexts where social stigma makes it difficult to physically visit a clinic.

A study published in Nature Scientific Reports in 2025 documented how telemedicine expanded access to Gender Expression Care from 24 zip codes concentrated in a single urban area to 158 zip codes across a wide geographic region [11]. The data also show a 56% reduction in missed appointments when visits take place via telemedicine [11] — a significant finding, considering that logistical barriers (distance, transportation, costs) are among the main obstacles to accessing care for transgender people.

Platforms like QueerDoc offer access to gender-affirming hormone therapies (testosterone, estradiol, progesterone) through telemedicine consultations, eliminating the need to travel to specialized centers that may be hundreds of miles away. For transgender people in rural areas or regions with restrictive legislation, telemedicine can literally be the only viable option.

AI-assisted voice training

Voice is a central element of gender expression, and voice training — vocal exercises to feminize or masculinize the voice — is a long and often expensive process when done with a specialized speech-language pathologist. AI is making this process more accessible.

Apps like Genderfluent use neural networks to provide real-time feedback on the perceived gender of one’s voice, allowing users to track their progress during exercises [12]. Voice Whiz offers real-time pitch visualization and gender perception analysis through on-device machine learning. TruVox, developed by the University of Cincinnati, is an open-source app that combines vocal exercises with speech component visualizations [12].

These applications do not replace the professional support of a speech-language pathologist, but they lower the barrier to entry: they are free or low-cost, usable independently, and available anytime. For those who cannot afford regular sessions with a specialist or who live in areas where these services do not exist, they represent a concrete first step.

Informational resources and support

AI-based chatbots can provide basic information about gender identity, transition pathways, legal rights, and local resources. They do not replace support from mental health professionals or specialized organizations, but they can serve as a first point of contact for someone exploring their identity who does not know where to turn.

In contexts where access to accurate information is limited — due to language, geographic, or social barriers — an AI system that can provide answers based on scientific sources, available 24/7 and in one’s own language, can make a concrete difference. The translation capabilities of language models also make resources accessible that would otherwise be available only in English.

Research and data analysis

AI accelerates research on the health of transgender people: analysis of large clinical datasets, identification of patterns in treatment outcomes, development of personalized protocols. Natural language processing tools can analyze scientific literature to identify research gaps or synthesize evidence from hundreds of studies.

This is particularly relevant in a field where data have historically been scarce. Research on trans health has suffered for decades from small samples, short follow-ups, and selection bias. AI does not solve these structural problems, but it can help extract maximum value from available data and identify priorities for future research.

The root problem: who designs, who decides

At the root of all these problems is a question of representation. Who designs AI systems? Who decides which gender categories to include in a database? Who labels the training data? Who defines what counts as “explicit content”?

The answer, in most cases, is: cisgender people, in academic and corporate settings where transgender people are underrepresented. A 2024 study published in the journal AI and Ethics analyzed how AI “hype” impacts the LGBTQ+ community, highlighting a fundamental disconnect between those who develop the technology and those who bear its consequences [7].

The catalogue of best practices for developing inclusive AGR solutions, published in 2024 by ACM SIGAPP, proposes concrete guidelines: eliminating mandatory binary classification, allowing gender self-identification, including trans and nonbinary people in development teams and testing processes [9]. But these are recommendations, not requirements. And in the absence of regulation, adoption remains voluntary.

The 2024 Forbidden Colours report on the impact of AI on LGBTIQ+ people proposed a European regulatory framework that accounts for the specific needs of marginalized communities, emphasizing that the European AI Act — while representing a step forward in regulation — does not explicitly address risks to LGBTQ+ people [14].

What can be done

There is no single solution, but there are clear principles.

First: ban inherently harmful technologies. Automated gender recognition based on physical features cannot be “improved” to become inclusive. Its operating principle — deducing gender from appearance — is incompatible with self-determination [2]. It must be banned, not corrected.

Second: include transgender people in the design process. Not as research subjects, but as designers, developers, and testers. The 2025 study on AI and personalized medicine showed that focus groups with trans people produce insights and solutions that cisgender development teams simply do not see [8].

Third: regulate algorithmic moderation. Platforms must be transparent about the criteria of their moderation algorithms, submit them to independent audits, and ensure effective appeal mechanisms when content is wrongly removed [3].

Fourth: invest in the opportunities. Telemedicine, voice training apps, and AI-based informational resources have enormous potential to improve the lives of transgender people. But this potential is realized only if these technologies are developed with transgender people, not for them.

Fifth: protect sensitive data. Gender identity, transition status, and health data related to gender-affirming care are particularly sensitive information — as explored in the article on privacy and gender identity. AI systems that process these data must be subject to enhanced protection standards.

A future to build

Artificial intelligence is neither an ally nor an enemy of transgender people. It is a tool that amplifies what it finds: if it finds biased data, it produces biased results; if it finds representative data, it produces more equitable results. The difference is made by human choices — who collects the data, who designs the models, who decides how they are used.

Transgender people are right to be skeptical. The FAccT 2025 study data reflect concrete experiences of algorithmic discrimination, not abstract prejudice [1]. But skepticism must not translate into disengagement: the opportunities offered by AI — access to care, tools for expression, informational resources — are too important to be left exclusively in the hands of those who have never had to face a system that does not recognize their identity.

The future of AI for transgender people depends on one thing: participation. Participating in design, regulation, research, and critique. Not as an exception to be managed, but as an indispensable perspective for building technologies that work for everyone.