Artificial intelligence arrived in healthcare with a big promise: faster diagnoses, automated triage, personalized care plans. By 2025, that promise has a real address. AI tools are already helping radiologists, predicting outbreaks, and monitoring mental health inside companies in near real time. But as these systems become infrastructure, an urgent question surfaces: *What if AI is simply replicating, with greater sophistication, the same exclusions we already experience offline?* This isn't speculation. It's what the evidence shows. --- **The problem isn't the algorithm. It's the mirror it uses.** Researchers have documented how hospital triage systems underestimate pain in Black patients, how cardiovascular risk models were calibrated almost exclusively on European populations, and how mental health tools reproduce cultural biases that render certain kinds of suffering invisible. The mechanism is simple and dangerous: when training data reflects a limited reality, the model learns that those absences are normal. It doesn't decide to exclude anyone. It simply never learned to include them. In corporate emotional health, this goes deeper. Platforms trained on samples of white, high-income workers from the Global North tend to classify as noise the very signals of distress that don't fit the expected pattern. Stress from workplace racism. Anxiety tied to housing insecurity. Exhaustion from carrying two or three jobs at once. That data exists. Few systems were built to read it. --- **Neutrality has a cost.** When an algorithm learns from biased data, it reproduces inequality under the appearance of objectivity. And apparent objectivity is one of the most effective ways to perpetuate exclusion, precisely because it requires no justification. In Brazil today, where mental health now appears in sustainability reports and ESG commitments carry real weight inside companies, the risk is exactly this: using AI tools as proof of care without questioning what they are actually learning, and about whom. --- **Real care requires knowing who's on the receiving end.** At AfroSaúde, we start from a premise that rarely appears in the technical specifications of digital health products: care only works when it considers who is on the receiving end. Their histories, their silences, the contexts that shape their bodies and minds. That's where Mentalaize came from. It's a tool for diagnosing and monitoring psychosocial risks at work. It crosses clinical data with behavioral and interaction indicators, generating individualized action plans. But the difference lies in the layer most products ignore: the system recognizes social contexts and incorporates the real diversity of the Brazilian workforce. Race, territory, socioeconomic condition, and occupational history are not distortions to be filtered out. They are legitimate variables. This isn't about preventing sick leave. It's about understanding that sick leave is simply the moment when suffering became visible enough to be counted. --- **What this requires in practice.** With AI regulation advancing in Brazil and the European Union, the digital health sector faces a choice: adapt reactively to new rules, or lead a genuine shift. Ensuring AI doesn't reproduce exclusion isn't a competitive advantage. It's a requirement. And it starts with concrete things: diverse teams involved in design phases, not just in communications. Data that represents the plurality of people who will be affected. Bias audits throughout a product's life, not only at launch. And genuine listening to the communities these systems claim to serve. Innovation isn't just about moving technology forward. It's about being clear on who moves forward with it, and who gets left behind when that question is never asked.