Stefan Harrer, PhD’s Post

Chetna Gakhar S.

Regulatory frameworks may lag practice by more than one era - some content currently being reviewed is still grappling with general purpose AI systems and generative AI, let alone agentic AI. I like the progressive idea in the article about making AI systems undergo a structured 'training' process, but unfortunately that is often impossible in practice when parts of the system are not open source. Alternative ideas we are experimenting with are performance, safety and security quality testing ('AI assurance'), strict restriction of egress and tools ('containment') and having fine toothed definitions on the spectrum of human in the loop when assessing agency levels of agents. Glad to see some innovative approaches on the table, but there is a lot of work to do to overcome outdated ways of thinking. As always the most important model might be our mental model.

Thanks for this post. A possible way forward is to open up the structural model of regulation to explore more how we might focus on the input side of innovation rather than just the output, so if regulation can be used to demand more patient centric and societal approaches that might curtail some of the risk, as well as ensuring better understanding of what is realistic so instances where things go wrong don't lead to panic and fixes can be found.

Great post and definitely agree that AI adoption in healthcare and life science has far outpaced regulation. Majority of AI solutions are not independently evaluated or monitored after implementing, the most important time. We at PsychedAboutAI have built evaluation frameworks for this purpose

Thanks for sharing, Stefan Harrer, PhD ! It is urgent‼️ Without this kind of solutions we will all suffer from the shock that relevant care we were used to cannot be given anymore and we will miss the opportunity for real personalized care and prevention. We should not hold ourselves back #transform #health #equality

Thanks for the insightful summary. Not a healthcare person. But I wonder if it’s correct to frame this as health regulations playing catchup, or it should be framed as AI agents not meeting the right standards yet for healthcare, a domain where it is as high stakes as it gets?

One of the most powerful AI agents shaping the healthcare ecosystem is the one integrated into the search engines we all use - and this context faces similar regulatory challenges where the framework is still built for a pre-internet era. Let alone one where Dr Google has evolved into 'Chat GPt'. The relationship between patients and these answer engines is deepening due to the conversational nature of the exchange and the empathy embedded in the language used by AI. Where restrictions prohibit the companies responsible for the quality use of their medicines from offering easily digestible answers in front of firewalls, the default sources are third party websites that largely contain unverified content that is frequently locally irrelevant and often out of date. It's time to permit the open publication on company websites of 'patient responsive content' that answers real-world queries with accurate and accountable advice. This will displace the other sources and offer AI something to grab onto and deliver with confidence into the 'zero click search' experience many patients (and potentially HCPs) engage in now.

Agentic systems are moving much faster than regulation. Sandboxes and adaptive oversight may not just be temporary fixes, they might need to become standard, because static approvals can’t keep up with systems that change and adapt. The real question is how regulation itself can evolve alongside the technology.

Stephen Gilbert

Pieter Jeekel Levie Hofstee