Biased medical AI, Trump signs an order and self-correcting LLMs

Today’s most important AI news

The day was marked by uncomfortable research: studies show that the most widely used models in medicine produce different diagnoses depending on the patient’s gender, and that LLM hallucinations are detectable well before they come out as output. In parallel, Trump signed a new executive order on AI just two weeks after cancelling the previous one, redefining the regulatory framework in the world’s biggest market. It wasn’t a day of spectacular launches, but of evidence accumulating about the real limits of systems already in production.

LLMs diagnose differently depending on the patient’s gender

A paper published today on arXiv tested three models — Gemini Flash, Claude Sonnet and GPT-mini — with the same profile of neurological symptoms: persistent headache, blurred vision, morning nausea and visual disturbances. The only variable that changed between queries was the patient’s declared gender and age. The result is uncomfortable: the models produced different triage recommendations for identical symptoms, including changes in the assigned level of urgency.

This isn’t a minor calibration problem. In contexts where these models are used as support in health systems — already a reality in several Latin American countries — a systematic bias in triage can mean that patients with the same symptoms do or don’t receive priority care depending on factors that shouldn’t influence the diagnosis. The study tested seven different demographic combinations (three age groups crossed with gender), which gives statistical weight to the findings.

For teams building or evaluating AI tools in health, the practical implication is direct: it’s not enough to test diagnostic accuracy on an average patient profile. You need to audit the model’s behavior across demographic variations that, in theory, shouldn’t affect clinical recommendations.

Hallucinations are written in the model’s intermediate layers

Another relevant paper of the day shows that LLM hallucinations don’t appear out of nowhere at the output layer: they are encoded in a linearly separable way in the hidden states of the model’s intermediate layers, before the text is generated. The study worked with three models of between 7B and 8B parameters — Llama 3.1, Mistral 7B and Qwen 2.5 — on four truthfulness benchmarks, including TruthfulQA and HaluEval.

What this means in practical terms is that it’s possible to build hallucination detectors that act before the model delivers a false answer, simply by reading signals in intermediate layers with simple linear classifiers. No second large model or expensive external judge is needed. The signal is already there, in the architecture.

For developers building LLM applications in production, this finding opens a concrete direction: instrument models to monitor internal states at inference time, not just filter the output. It’s a paradigm shift from the current approach of post-generation detection.

Trump signs a new executive order on AI

Less than two weeks after revoking the previous executive order on artificial intelligence, President Donald Trump signed a new one. According to MIT Technology Review, the order has five central points and keeps the focus on promoting AI development in the United States, although the details on oversight and governance mechanisms are still being analyzed by the community.

The context matters: this move comes at a time when competition with China in AI is one of the administration’s most active foreign-policy axes. Repealing and replacing an order in such a short time suggests internal tensions over how to balance the deregulation the industry is asking for with some level of oversight that justifies the national-leadership narrative.

For Latin American companies that export tech services to the United States or that use US model infrastructure, regulatory changes in Washington have a direct effect: they can change what data can be processed, under what conditions model providers operate, and which compliance standards become implicitly necessary to keep contracts with North American clients.

In a nutshell

Today’s pattern is that of a technology already being used in high-impact contexts — medicine, education, infrastructure — but whose evaluation remains superficial. Detecting that a model hallucinates by looking only at the output, or that it discriminates by looking only at average accuracy, is insufficient. The research published today points in the same direction: you have to open the box, instrument the models from the inside and audit differential behavior, not just aggregate performance. Trump’s order is the day’s political noise, but the real signal is in the papers: the models we already have in production are more opaque and more biased than we want to admit.

---

Sources used: https://arxiv.org/abs/2606.03641, https://arxiv.org/abs/2606.02628, https://www.technologyreview.com/2026/06/03/1138322/the-download-trump-ai-order-smart-glasses-warfare/