Most LLMs are released worldwide, and country-specific iterations are not available that require a global approach from regulators. It is also not clear in which technology category LLMs fall into from a regulatory perspective. However, based on the differences between LLM and previous deep learning techniques, new regulatory categories may be needed to address the unique challenges and risks of LLM.
A regulatory body should only design regulations for an LLM if the developer of the LLM claims that its LLM can be used for medical purposes. or if the LLM is developed, adapted, modified or directed specifically for medical purposes. While LLMs in widespread use today don’t fit into either category, medical replacements for LLMs that are specifically trained based on medical data and databases probably do.
One prominent example is Med-PaLM, published by researchers at DeepMind and Google. In that study, the authors proposed a framework for human evaluation of model answers along multiple axes, including factuality, accuracy, likelihood of harm, and bias. Additionally, using a combination of prompting strategies, their model achieved an accuracy of 67.6% on the US Medical Licensure Examination questions, outperforming previous state-of-the-art models by more than 17%. Human evaluations revealed significant gaps in the responses provided by LLM, so a directed-prompt adjustment was introduced and the resulting model Med-PaLM showed promising performance, although Clinicians were still inferior. Since then, GPT-4 has allowed him to achieve over 85% accuracy on the same test.14.
With the release of GPT-4, which can analyze not only text, but also images, we expect to see it grow into a model for analyzing uploaded documents, research papers, handwritten notes, audio, and video in the near future. (Table 2).
This emphasizes the notion that just regulating the current LLM model is not enough, as new iterations with these advanced features can be expected to be implemented at a similar rate to previous iterations. . Without considering these future additions, a regulation focused solely on language models could miss important updates by the time they are widely accessible.
Businesses that own approved devices and decide to implement LLM in their services face additional challenges. So, if the AI-based medical technology LLM recently introduced is already approved for medical use, how will the FDA regulate it? Table 3 summarizes the regulatory challenges. is.
There are proposals for regulation of LLMs, but they come from outside the medical establishment. In their research paper, Hacker et al. proposes a new terminology to capture his chain of AI value by distinguishing between developers, implementers, expert and non-expert users, and recipients of his LLM output. The authors also proposed his four strategies for ensuring that these models are credible and deployed for the benefit of society as a whole.In particular, regulation should focus on specific high-risk applications rather than pre-trained models per se, and should focus on (i) transparency mandates, (ii) risk management, (iii) non-discrimination. and (iv) Content Moderation Rules.15.
Noting that existing audit procedures fail to address the governance challenges posed by LLMs, Mökander et al. propose three contributions to fill the gaps. 1) establish the need to develop new audit procedures that capture the risks posed by LLM; 2) Outline a blueprint for auditing his LLM in a viable and effective manner leveraging IT governance and systems engineering best practices. 3) discussing the limitations of the prospect of auditing LLMs;16.
Such potential solutions could serve as benchmarks for new regulations in healthcare. In any case, regulators and legislators need to act quickly to keep abreast of LLM’s unprecedented evolution and progress.
In a sign of increasing pressure on regulators, in March 2023, a group of prominent computer scientists and tech industry executives such as Elon Musk and Steve Wozniak issued a “to all AI labs: We called for an immediate suspension of training for at least six months on AI systems “more powerful than GPT-4.”17. Their letter states, “Over the last few months, the AI Lab has aimed to develop and deploy ever more powerful digital minds that no one, not even their creators, can understand, predict, or reliably control. We are caught up in an uncontrollable competition.” This suspension should be public, verifiable, and include all key stakeholders. If such a moratorium cannot be enacted immediately, national governments should intervene and implement moratorium measures. “
Prominent AI experts such as Andrew Ng opposed the idea, calling for instead seeking a balance between the enormous value AI creates and the real risks. We agree that the moratorium cannot actually be enforced without government intervention. “Suspending emerging technologies that governments don’t understand is anti-competitive, sets terrible precedents, and is a terrible innovation policy.”18.
To add to our concerns, it is worth mentioning that Italy became the first Western country to temporarily block ChatGPT in April 2023, citing privacy concerns and lack of proper regulation. To do.19.
