ALIGNMT AI

ECRI asked an AI chatbot where to place an electrosurgical return electrode. It said the patient's shoulder blade was fine. That placement could cause burns. That's one reason ECRI ranked "misuse of AI chatbots" the No. 1 health technology hazard for 2026, ahead of EHR outages and counterfeit products. ECRI's examples include wrong diagnoses, unnecessary tests, and invented body parts. The scale is the real story. OpenAI says over 40 million people a day use ChatGPT for health information. Since September 1, organizations on ChatGPT for Healthcare can connect it to Epic for patient context. General-purpose AI is already in your hospital. It arrived through browsers, phones, and enterprise licenses, not through your AI committee's intake form. What I'd do this quarter: → Inventory sanctioned AND unsanctioned gen AI use (network and SSO logs tell you more than surveys) → Write an approved-use policy: what's OK for admin drafting, and what's never OK for clinical decisions → Monitor sanctioned tools the way you'd monitor any clinical system. That's why we built ALIGNMT AI's integration with the Claude Compliance API. ECRI's own recommendation to health systems: set up AI governance committees, train clinicians, and audit performance regularly. How many gen AI tools do you think are in use at your org right now? Now double it. #GenerativeAI #PatientSafety #HealthcareAI #AIGovernance #ShadowAI

Grateful to Healthworx for spotlighting our journey through the Healthworx Accelerator. We came in with early revenue from building to federal certification standards. What we needed next was commercial scale: getting in front of health enterprises that are actively working out how to govern their AI. The Healthworx network did that for us. It helped us see early how closely CareFirst’s challenges matched what we’d built, and it gave both sides a trusted place to talk openly about real problems. Healthcare organizations are under growing pressure to prove their AI is safe, compliant and working, under frameworks like Joint Commission RUAIH and the EU AI Act. That’s why we built continuous AI governance monitoring. Thank you to the Healthworx and CareFirst teams for the partnership. 🙏 #AIGovernance #HealthcareAI #ResponsibleAI #HealthTech

When I built healthcare AI products at Google Cloud, the question was usually "Is the model accurate?" These two studies show why that question isn't enough anymore. → Mount Sinai Health System tested 20 models with clinical instructions that conflicted with patient safety. 16.6% of choices were potentially harmful. A short safety reminder lowered that to 10.1%, but 1 in 10 was still risky. → Vanderbilt University found chatbots amplifying psychotic symptoms in 28 patients, usually by agreeing with delusions. Most of those patients were in their first episode. What this means for anyone deploying AI in care settings: 1. Validate under pressure, not just ideal conditions. Red-team for authority, urgency, and persistent users. 2. Re-test every time the model, prompt, or workflow changes. A system prompt is part of your safety case. 3. Treat safety prompts as one control, never the control. Human oversight still has to be designed in. Great breakdown of both studies ↓ #HealthcareAI #AIGovernance #PatientSafety #ResponsibleAI

We keep asking whether medical AI gives the right answer. Two new studies suggest the better question is: what does it do under pressure? • Mount Sinai Health System (Communications Medicine): when 20 models got clinical instructions that conflicted with patient safety, 16.6% of their choices were potentially harmful. A brief safety reminder brought that down to 10.1%. • Vanderbilt University Medical Center (JAMA Psychiatry): in 28 patients, chatbot use appeared to worsen psychotic symptoms, usually by validating delusions. About 61% were in their first psychotic episode. The common thread is deference. Models trained to be helpful and agreeable can go along with the person in front of them even when patient welfare points the other way. The good news: context can steer models toward safety as easily as away from it. The work now is building testing that holds up when users push and conversations run long. Full write-up below #AIAlignment #HealthcareAI #PatientSafety

95% of organizations report no business return on generative AI so far. Yet a handful of healthcare and life sciences teams are posting numbers any CFO would sign off on. ("The GenAI Divide: State of AI in Business 2025” by MIT Media Lab’s Project NANDA) Our Q4 edition of Aligned AI Innovation pulls the past year’s most decision-relevant research into one planning brief for 2027 budgets: → Waystar: $15B+ in claim denials prevented in 2025, appeal time cut 90% → Novo Nordisk: 90% less writing time on clinical study reports → Sanofi: $300M in revenue risk avoided by predicting inventory shortfalls early → AstraZeneca: 12,000+ employees through a five-level AI accreditation → Viz.ai: ISO 42001 certified for its AI management system What separates them? New data from @Wharton Human-AI Research and GBK Collective: teams that measure growth outcomes are nearly 3× as likely to report strong ROI as teams that track efficiency alone. Inside: where the returns are showing up, why agents change the risk math, what boards are now asking for, and a 5-item checklist to finish before year-end. What’s the one AI question your board keeps asking? Tell us in the comments 👇 #HealthcareAI #AIGovernance #DigitalHealth #LifeSciences #ResponsibleAI

Three things happened in healthcare AI governance in the past two weeks: → Coalition for Health AI (CHAI) released a draft vendor intake playbook: 57 common controls, roughly 40% of which can be met by vendor attestation alone. → Pennsylvania's House passed an AI governance bill for hospitals and insurers, also built on attestation. → Advanced Research Projects Agency for Health (ARPA-H) funded two systems to monitor clinical AI after go-live. The first two standardize the signed statement. The third funds the machinery to check whether it holds. Attestation is useful. But it describes the model the vendor shipped, not how it behaves with your patients six months later. That's where drift, subgroup gaps and workflow misuse show up, and where regulators and plaintiffs will look. The rule of thumb: a vendor can attest to what it controls. Anything that depends on your context needs evidence you generate. In this issue: where attestation is enough, where it isn't, and five questions to ask before your next AI tool goes live. #AIGovernance #HealthcareAI #ClinicalAI #ResponsibleAI

Great to see ALIGNMT AI featured as a security and compliance platform supporting the Claude Compliance API. 🙌 It's a strong signal of where enterprise AI is heading. Companies aren't asking whether to adopt AI anymore, they're asking how to see and govern it. Letting security teams review Claude activity in the tools they already use is exactly the right approach. Proud to be part of this ecosystem, and in company with teams we deeply respect. If you're a Claude Enterprise or Claude Platform customer thinking about AI governance, let's talk. #AIGovernance #AISecurity #Claude #Anthropic

An AI agent got told "no" by a government health website. So it found another way in. In June, an OpenAI model researching medicine spending hit access blocks on Australia's Medicare statistics portal. Instead of stopping, it worked around them, pulled non-public files, and wrote files to the server. Nobody asked it to. OpenAI didn't notice for two months. Australia found out in September, via an email to a public inbox. No patient records were touched this time. But I keep thinking about how ordinary the task was. Nobody was trying to hack anything. The agent just really wanted to finish its job. That's the part healthcare should be paying attention to. We wrote up what happened and what it means 👇 #HealthcareAI #AISafety #AIGovernance

A quick primer on where AI regulation in healthcare is heading in 4 concepts. 1️⃣ Lifecycle-based regulation Today, most medical AI is evaluated once, pre-market, on controlled data. But models drift, and they perform differently across hospitals and patient populations. Lifecycle regulation means oversight continues through deployment, monitoring, updates, and real-world learning. 2️⃣ Staged authorization Think of it as a learner's permit for medical AI: a model enters clinical use under conditional parameters and heightened monitoring, and earns full authorization by proving itself in the real world. 3️⃣ Proportionality Not all AI carries the same risk. Low-risk tools that support staff should face lighter requirements so benefits aren't delayed; tools with greater impact on someone's care need stronger safeguards. Right safeguards, right time, right level of risk. 4️⃣ System-wide accountability Safety isn't only the manufacturer's job. Deploying organizations, clinicians, and regulators all carry defined responsibilities — and no one should be accountable for something they can't understand, control, or act on. Why learn these now? The UK's National Commission just built its entire recommended framework on them, after a yearlong inquiry with 12,000+ stakeholders. And the EU AI Act, CHAI, and HTI-1 all point the same direction. #HealthcareAI #AIGovernance #AIRegulation #DigitalHealth

A rare, unvarnished look at what happens when a large, well-resourced organization tries to govern AI at scale. The NY State Comptroller's new audit of NYC found that three years after being told its AI governance was ineffective, the city still can't answer the discipline's most basic question: what AI are we actually running? No complete inventory. Risk assessments done only on request — seven in three years. Guidance everywhere, binding policy almost nowhere. And a tool that quietly steered building inspections outside any AI-specific control. The pattern isn't a government story; it's the default failure mode for any organization. Our latest piece covers what the auditors found, why it matters beyond New York, and a practical sequence for closing the gap. #ResponsibleAI #AIPolicy #AICompliance and #RiskManagement