2MM: AI Roundup: Food and Drug Administration reviews liver injury prediction tool, Joint Commission launches healthcare artificial intelligence certification, governance playbooks aim to standardize adoption, and pediatric hospitals bring generative tools to the frontline

Food and Drug Administration reviews artificial intelligence tool for liver injury prediction

The Food and Drug Administration (FDA) is reviewing an artificial intelligence based drug development tool designed to predict drug induced liver injury, one of the more stubborn safety problems in clinical development. The agency’s Center for Drug Evaluation and Research accepted a letter of intent for the tool, according to Reuters reporting published on June 3, 2026. Drug induced liver injury remains a major reason drug candidates fail, and it is especially difficult because traditional preclinical methods do not always predict human risk well. That is the part that matters most for physicians. A therapy can look promising early, only to run into liver safety problems once human exposure expands. The FDA said the artificial intelligence tool could help improve early safety assessment, reduce reliance on animal testing, and support better decisions before human trials begin. This does not mean artificial intelligence is suddenly replacing toxicologists, pharmacologists, or clinical safety teams. It means regulators are looking at whether models trained on complex datasets can help flag risk earlier than conventional methods alone. The timing also fits a broader push to make drug development less slow, less expensive, and less dependent on late stage surprises. For clinicians, the downstream hope is straightforward. Better early safety prediction could mean fewer failed trials, fewer patient exposures to drugs with hidden toxicity, and more efficient movement of safer candidates into human testing. The caution is that liver injury prediction is biologically complex, and model performance will need to be transparent, reproducible, and clinically meaningful. A tool that works in a curated dataset may not perform the same way across therapeutic areas, patient populations, dosing patterns, and comorbidities. Still, this is a useful example of how artificial intelligence is moving into the infrastructure of drug development, not just the clinic. The FDA’s review will be worth watching because it may help define how predictive models are evaluated before they influence high stakes safety decisions.

Joint Commission launches responsible artificial intelligence certification for healthcare organizations

The Joint Commission has launched a voluntary certification program focused specifically on responsible artificial intelligence use in healthcare organizations. Announced on June 1, 2026, the program is designed to recognize hospitals and health systems that are adopting artificial intelligence with sufficient governance, monitoring, education, and accountability structures in place. The Joint Commission announcement describes the certification as the first of its kind built exclusively for healthcare organizations. That is important because artificial intelligence adoption has moved faster than many hospitals’ internal oversight systems. Clinical documentation tools, imaging support, triage models, revenue cycle products, and patient facing chatbots are all being introduced into systems that may not yet have a unified way to evaluate risk. For physicians, this could become a practical signal that an organization is at least trying to put structure around the technology. Certification will not make artificial intelligence safe by itself, and it will not answer every question about bias, liability, or clinical accuracy. But it may push health systems to formalize basic safeguards that should already exist. Those include inventorying tools, defining who owns model oversight, educating staff, monitoring performance after deployment, and creating pathways to report problems. Reporting from Healthcare IT News emphasized that the program is meant to assess whether organizations have governance and monitoring processes in place, rather than simply whether they own a particular technology. That distinction matters. The real risk in healthcare artificial intelligence is often not one bad model, but a collection of tools introduced without enough transparency or follow up. A voluntary certification could also create pressure on vendors, since hospitals may start asking for clearer documentation of model performance, training data, limitations, and update cycles. In the long run, this may be less flashy than a new diagnostic algorithm, but more important. Healthcare artificial intelligence will only be as trustworthy as the systems that govern it.

Pediatric hospitals bring generative artificial intelligence to the clinical frontline

Pediatric hospitals are beginning to bring generative artificial intelligence into the daily work of clinical care, but the most realistic use cases are not the dramatic ones. At CHOC Children’s, leaders are thinking about artificial intelligence as a way to reduce administrative burden, organize data, and help clinical teams work through complex information more efficiently. A May 22, 2026 report from Healthcare IT News described the technology as a tool that could ease clinician burnout, speed care, and reshape how hospitals think about data. That framing feels right for pediatrics. Children’s care is filled with context that does not always fit neatly into one note or one lab result. Growth curves, developmental history, school concerns, parent observations, medication changes, subspecialty input, and longitudinal follow up all matter. A useful generative artificial intelligence tool could help summarize that information, draft documentation, prepare discharge instructions, or organize referral questions before the physician enters the room. That may sound less exciting than autonomous diagnosis, but it is probably safer and more useful in the near term. Pediatric clinicians need tools that help them see the full picture faster without pretending the model understands the child better than the care team does. The risks are also different in pediatrics. Children are not simply smaller adults, and developmental norms vary widely by age. Pediatric datasets may be smaller and less representative than adult datasets, which raises concerns about bias and generalizability. A polished artificial intelligence summary can still miss a crucial family detail, a subtle developmental concern, or a safety issue buried in the chart. For physicians, the right question is not whether generative artificial intelligence belongs in pediatrics at all. It is where the technology can reduce friction while keeping clinical judgment firmly in human hands. If deployed carefully, these tools may give pediatric teams more time with families and less time wrestling with the electronic health record. That would be a meaningful win, even if the technology never makes a single diagnosis on its own.