50 shades of AI in regulatory science

Introduction

In the rapidly advancement of artificial intelligence (AI), particularly Generative AI, its integration into the regulatory science practice has never been more challenging. The benefits of AI such as ChatGPT in various domains have been widely recognized including in regulatory application within drug discovery and development. Still, while it comes with unprecedented opportunities, we are also faced with formidable challenges. As we recognize AI’s nuanced roles, from assisting and augmenting human capabilities to adopting autocratic and autonomous functionalities, it becomes clear that a one-size-fits-all setting to apply AI is impractical and counterproductive without considering context-of-use.^(p1) Within this context, the delineation of AI’s roles into “50 shades of ‘A’” (i.e., assisting, augmenting, autocratic, and autonomous) not only offers a framework for understanding but also highlights the necessity for tailored regulatory measures that recognize each function’s distinct characteristics and implications.

Integrating AI into regulatory science marks a significant evolution in how regulatory bodies approach review, oversight, compliance, and innovation. AI’s role extends from enhancing operational efficiencies to redefining the boundaries of human–machine collaboration. To navigate the landscape of AI in this field, it is crucial to understand the foundational categories that define its application and the context within which each is being used. Additionally, as AI is being incorporated into every stage of drug development and chemical risk assessment, aligning on shades of AI terminology makes communication easier and collaborations more effective.

Within regulatory science, AI is part of a toolbox, and how we use this tool depends on the regulatory application and the shade of AI is related to it. Each shade drives how we treat AI in terms of statistical performance, reproducibility, and explainability; this matrix can help us address AI within the context of use in regulatory applications. Importantly, each shade represents a unique interaction between human intelligence and machine capabilities, from automating tedious tasks to enabling machines to learn and operate with minimal human intervention. Here we discuss these distinctions, presenting definitions alongside exemplified applications of these AI categories in the regulatory domain. Through this exploration, we aim to stimulate a dialogue among researchers, policymakers, and practitioners about AI’s transformative potential and challenges within regulatory frameworks.

Presented below is Table 1, which outlines the four pivotal definitions of AI and examples pertaining to regulatory application. We certainly understand that this is just a subset of AI shades which can also overlap. The confidence in AI within regulatory science hinges on the specific context-of-use and the models’ degree of explainability, transparency, and statistical boundaries, which varies based on the significance of the AIdriven decisions and the extent of human involvement in these decision-making processes. Essentially, the criticality of the outcome influenced by AI necessitates varying levels of understanding and clarity about how AI models arrive at their conclusions, underlining the importance of these factors in situations where AI plays a pivotal role alongside human experts (Figure 1).

TABLE 1.

Definitions, Examples and Matrix of “50 Shades of AI” in Regulatory Science

Category	Definition	Regulatory Application	Matrix
Assisting Intelligence	AI applications designed to automate routine tasks such as data gathering and analysis. This form of AI supports human activities by handling repetitive or time-consuming tasks, enhancing efficiency without making independent decisions.	Literature screening, document formatting compliance, or deduplication of regulatory documents are routine tasks that often require minimal false negatives to ensure all the data that are captured for the manual inspection to improve review process. AI significantly streamlines these routine tasks by reducing time-consuming and labor-intensive efforts but without losing critical data.(p2) This enhances overall efficiency and allows human experts to concentrate on the more complex and nuanced aspects of the regulatory review process, optimizing both productivity and resource allocation.	The enrichment process aimed at improving regulatory efficiency significantly benefits from human intervention, where statistical measures like recall are valued more than explainability or reproducibility. In this context, the concern for explainability becomes less critical since human experts make the final decisions, ensuring the application of AI in the process remains transparent and accountable.
Augmenting Intelligence	AI that extends human capabilities by providing insights or identifying patterns not readily apparent to humans. This type of AI works in tandem with human intelligence to enhance science-based decision-making processes.	AI predictive models can be used to signal the potential area of concerns based on the historical data, for example, in reviewing Investigational New Drug (IND) application. It does not alter the regulatory decision-making process but it does require more data and evidence as signaled by the AI models. Augmenting intelligence enhances human decision-making by offering predictive analytics with specific performance metrics to suggest possible outcomes. It improves evidence-based review process in assessing safety and efficacy of a regulated products.(p3),(p4)	In the context of augmenting human decision-making with AI, a model’s predictive performance drives its application in regulatory decision-making. A model with high sensitivity may trigger a different decision-making as compared to the ones with high specificity. While reproducibility and explainability are desirable, the emphasis is placed more on a model that can be tested and verified for specific performance. In such scenarios, it’s important to understand the data used for model development, the verification processes applied to assess the performance, and the availability of the algorithm itself. In these cases, the importance of explainability may diminish, suggesting a balance between the need for transparent, interpretable AI systems and the practicality of their performance verification to enhancing decision-making processes.
Autocratic Intelligence	AI that takes on a more independent role in decision process with minimum human guidance. This category of AI provides data and evidence that can be directly used in decision-making.	Examples such as AI models can help make decisions within specified domains and contexts with minimal human intervention. For example, it may decide on the progression of drugs through certain phases of trials based on predefined criteria.(p5),(p6) For the FDA, when ISTAND (Innovative Science and Technology Approaches for New Drugs) qualifies an AI model for context-of-use in drug development, regulatory decisions can rely solely on the model’s output, eliminating the need for further verification.	In this situation, clear definitions of explainability, reproducibility, and context-of-use within the regulatory framework are paramount. Rigorous stress testing under varied conditions to evaluate performance and reliability is key. Continuous monitoring of decisions and outcomes with mechanisms to alert humans if anomalies are detected are needed to ensure AI is adhering to ethical and regulatory guidelines.
Autonomous Intelligence	AI that is capable of learning and operating independently with no human intervention including making final decisions.	Autonomous Intelligence represents an even more advanced form of AI, capable of learning and adapting without human guidance, oversight and intervention. Its applications in regulatory science could be: (1) independent audits: periodic audits by external parties to ensure integrity, accuracy, and fairness; (2) continuous learning and adaptation: learning mechanisms from outcomes, including successes and failures, to improve decision-making; and (3) safety nets: systems in place to prevent or mitigate adverse outcomes without human intervention.	In this context, the primary focus is on responsible AI in regulatory application. Concerning an operation beyond human control, all the aspects of AI performance such as explainability, transparency, trustworthiness, and ethical considerations should be comprehensively verified and validated. Ensuring autonomous systems adhere to ethical standards and regulatory requirements would be paramount. Caution needs to be exercised since AI is the primary decision maker and it should not be implemented without full confidence.

FIGURE 1.

Explainability & Transparency Requirements based on decision impact and human engagement.

Understanding the context of use for each shade of AI is crucial, as it helps identify and address specific biases inherent in different AI systems. For instance, augmenting intelligence requires understanding of data and algorithmic biases, such as patient population prevalence in clinical settings. Similarly, autocratic intelligence necessitates awareness of echo chamber biases, where a narrow set of perspectives derived from the limited input data might influence decision-making. Whereas autonomous intelligence remains a concept yet to be fully realized within regulatory frameworks, and its realization remains uncertain. However, there is potential for the development of partially autonomous intelligent systems, akin to partially autonomous vehicles. For instance, a vehicle may possess the capability to execute tasks like parallel parking independently, yet rely on human intervention for highway navigation. Moreover, varying degrees of partial autonomy could be envisaged; extending the previous vehicle example, the driver may relinquish control even on highways, but continuous monitoring by the driver could still be mandated. While the ideal of fully autonomous vehicles entails zero tolerance for accidents, the current reality still witnesses occasional incidents. However, in such cases, thorough investigation allows for the identification of causative factors and implementation of corrective measures to preclude recurrence. Yet, within regulatory frameworks, the occurrence of such incidents is untenable, necessitating a preference for partially autonomous intelligence, where oversight and corrective action can be more readily ensured. As AI becomes increasingly integrated into regulatory science, the need for comprehensive governance frameworks becomes paramount. These frameworks must be adaptable to the varying degrees of human-AI interaction inherent in the different “shades” of AI. As each category embodies a unique aspect of AI’s utility in regulatory science, from automating mundane tasks to enabling machines to learn and adapt independently. By understanding these definitions and their implications, regulatory bodies, and stakeholders can better navigate the complexities of integrating AI into their workflows, ensuring that technology serves to enhance, rather than complicate, the regulatory landscape. These definitions, examples and matrices offer a gateway to further exploration of each category and the ability to delve into AI’s practical applications, benefits, and ethical considerations in this crucial field.

In conclusion, navigating the shades of AI in regulatory science requires a comprehensive and forward-thinking approach. By embracing the spectrum of AI’s roles and potential, we can develop a regulatory framework that is both responsive and responsible.

Data availability

No data was used for the research described in the article.