Risk Assessment as executed by EFSA (and used here as a more general model for regulatory science) is a structured decision support process where selected scientists are asked to provide scientific recommendations to inform managerial decisions. Upon receipt of the mandate it starts start with a problem formulation phase that includes: the clarification and acceptance of the mandate that takes place in dialogue with the requestor; the translation of the general question into a scientifically answerable assessment question; the definition of the related conceptual model and selection of the overall approach for the assessment. The conceptual model illustrates all the sub-questions derived from breaking down the assessment question, along with a description being purely descriptive to mathematical with all the sub-questions expressed as parameters/variables. EFSA recognized the value of complementing problem formulation (i.e. description of the what) with an upfront definition of methods for conducting the assessment (i.e. description of the how the assessment will be conducted), which include the methods for answering each sub-question and for integrating evidence across sub-questions, including uncertainty analysis which can arise from limitations in the evidence (i.e. heterogeneity, degree of relevance, degree of internal validity and/or precision) and in the methods used throughout the assessment.
This plan phase is followed by the do phase which is the actual assessment, where the pre-defined and documented plan is implemented and conclusions are drawn in light of the identified uncertainties. The different approaches to answer sub-questions are: (1) Collecting, appraising, synthesizing and integrating evidence coming from the scientific literature collected or directly submitted; (2) extracting, assessing and analyzing data from databases and sources other than literature; (3) Eliciting expert judgement, when evidence is scarce and/or of limited validity or (4) Carrying out primary research studies. Combination of different approaches can be adopted for the same sub-question or, for broad assessments containing many sub-questions, for the various sub-questions.
The above described process has reached its limits both in the ability to execute timely as in the physical ability of the available human brains to read, appraise and integrate the exponentially growing amount of evidence in a structured way.
Automation of the process, in which AI models play a role, hence practically moving the workload from the human experts to the machine is the only way forward. However AI-powered models are still far from demonstrating human-like causal reasoning, imagination, top-down reasoning, or artificial general intelligence that could be applied broadly and effectively on fundamentally different problems such as described above for the plan and do phase of the risk assessment process.
Nevertheless a human centric approach where the core accountability of the planning of the risk assessment is human while the execution (do) is more a human augmented approach becomes more and more reality allowing human and artificial intelligence to work in tandem. These models could be applied to narrow, well-defined area’s of the risk assessment where AI guardrails could be manifested in terms of human intervention or by incorporating rule-based algorithms that hard-code human judgment.
Examples will be given on area’s of the risk assessment process where this is approach is currently implemented or under full development such as on the use of those human centric AI approaches such as for the answering of the sub questions and specific parts of the plan phase. read more... read less