Embedding models are the way AI learns to recognize patterns in language and information by converting text into mathematical representations of meaning; understanding this helps you see why AI can sometimes find relevant resources even when you don't use exact keywords, and why it can occasionally miss things that seem obviously related to humans. This literacy helps you work with AI's strengths rather than expecting it to think exactly like you do.
An embedding model is a specialized AI system that represents words, phrases, or concepts as numerical vectors—essentially mapping meaning as points in multi-dimensional space. The fascinating and problematic part: words that appear in similar contexts get mapped near each other. This is incredibly useful ("king" and "queen" are near each other, as are "doctor" and "nurse"). But it also means stereotypes and biases encoded in training data become encoded in the model's geometry.
For LGBTQ+ people working with AI, understanding embeddings matters because many practical AI systems rely on them: resume screening systems, recommendation algorithms, search ranking, content moderation, and semantic similarity tools. When a resume contains "trans" or "non-binary," the embedding model has learned associations between these terms and other concepts from training data. Those associations might reflect broader social patterns you'd want to avoid.
The fundamental algorithm: a model processes text and learns to predict missing words based on surrounding context. Through this process, it learns that if "gay," "lesbian," and "queer" appear in similar contexts across training data, they should be positioned near each other in embedding space. This contextual similarity is useful.
The problem: if the training data contains stereotypical text about LGBTQ+ people (associating trans women with sex work, linking gay men with promiscuity, etc.), these associations become embedded in the model's geometry. The model isn't making judgments; it's learning statistical patterns. But those patterns reflect historical discrimination in the training data.
Researchers detect embedding bias through targeted probing. They construct word pairs designed to reveal stereotypes: if you ask the model "What's the similarity between (trans, successful)?" and compare it to "(cisgender, successful)," do the scores differ? Systematic differences across many such pairs reveal embedded bias.
More sophisticated approaches use analogies: "If 'man' is to 'programmer' as 'woman' is to X, what's X?" If the model returns 'nurse' or 'secretary,' it has embedded gender stereotyping from training data. LGBTQ+ analogies reveal similar patterns: "If 'straight person' is to 'monogamous' as 'gay person' is to X."
Resume screening and hiring: If a hiring tool uses embeddings to filter candidates, and the embedding model associates certain LGBTQ+ terms with stereotyped professions or characteristics, qualifications get filtered invisibly. A resume mentioning "LGBTQ+ mentoring" might get downranked if embeddings associate LGBTQ+ with lower-status characteristics.
Recommendation systems: Content recommendation relies heavily on embeddings. If embeddings associate trans content with sensitive/mature content more strongly than they associate cisgender content similarly, trans creators get over-filtered or shadowbanned.
Search functionality: If your application uses embedding-based search and embeddings encode stereotypes, search results will reflect those stereotypes invisibly—users searching "trans professional" might not find results for trans people in professional contexts if embeddings associate "trans" more strongly with other concepts.
Organizations working with embeddings can attempt debiasing: identifying problematic associations and manually adjusting them, using different training data, or training models specifically to minimize detected bias. These work partially but not perfectly. You cannot remove a bias that's truly fundamental to the training data without changing the training data itself.
The practical implication: if you're building AI systems that use embeddings, audit them for LGBTQ+-related bias by (1) testing with LGBTQ+ terminology and (2) checking whether LGBTQ+ inputs produce anomalous results compared to equivalent non-LGBTQ+ inputs. If you're using embeddings others built, you have less control but can sometimes request transparency about debiasing efforts.
Embeddings appear in more contexts than most people realize: plagiarism detection, content moderation, keyword extraction, similarity recommendations. Each of these has different implications for LGBTQ+ people. A content moderation system with biased embeddings might overfilter LGBTQ+ content. A plagiarism detection system might flag legitimate work differently based on who submitted it. These downstream effects are subtle but real.
Try this: If you use any embedding-based system (most LLMs contain embeddings), test for LGBTQ+-specific behavior. For recommendation systems, search for equivalent concepts and compare results: "LGBTQ+ community" versus "straight community." Are results comparable quality? For resume screening, test how mentions of LGBTQ+ involvement affect ranking. These informal audits won't be definitive but reveal whether the system seems to encode stereotypes.
Peri can explain this concept, give practical examples, help you decide whether it applies to your situation, or recommend a journey if appropriate.
Explore related journeys or tell Peri what you're working through.