Today I presented a paper titled “A rule-based risk assessment tool based on large language model” which was a pilot study done recently at Taltech. The presentation took place at session “AI & INNOVATION” at “XXXV ISPIM INNOVATION CONFERENCE” which is taking place in Tallinn this week, June 9-12, 2024.
Presentation overview
This study aimed to design and test a Virtual Risk Assistant (VRA) system that combines rule-based reasoning, large language models (LLMs), and business process management to automate ergonomic risk assessments in workplace settings. The findings showed that the VRA enabled risk assessments approximately seven times faster than manual methods, with risk probabilities assigned by AI comparable to human assessors (statistically the same). Though, the AI tended to assign higher hazard levels which were statistically not comparable to human assigned levels.
Q&A session overview
The questions raised during the discussion session highlighted the audience’s interest in the practical and legal aspects of integrating AI into business processes. The discussion showed the importance of maintaining human oversight while leveraging AI’s capabilities to enhance efficiency and decision-making in risk assessments.
Summary of questions from the audience and my replies
(transcribed by OpenAI Whisper, large v3 and summarized by ChatGPT)
Question: Why did you choose VRA? Why did you choose the combination of business process management, large language models, and rule-based reasoning?
The main discussant appreciated the setup of the system and the clarity of the factors involved in the risk assessments. She asked why we chose to combine business process management (BPM), large language models (LLMs), and rule-based reasoning.
Answer: I explained that the integration of BPM with AI, specifically LLMs and rule-based reasoning, was to mitigate the biases of LLMs. Rule-based reasoning ensures that both the input and output of the LLMs are checked and corrected if necessary. This combination can introduce LLM-based tools into business processes effectively.
Question: The main discussant was also curious if we had plans to compare the current setup with an alternative setup, perhaps through an A-B experiment.
Answer: I clarified that our A-B experiment involves comparing the manual risk assessment process (A) with the VRA tool (B). This comparison helps us determine if the tool performs as well as human evaluators and identifies specific areas where it may fall short.
Question: Is it relevant to measure the time taken by the machine versus a human to solve the task?
Answer: I emphasized that time measurement is crucial. Businesses need to understand the cost-benefit outcome when deciding to adopt such tools. Our study included assessments of the time taken for both manual and AI-assisted evaluations to provide a clear comparison.
Question: Does it make sense to look at the time the machine needs to solve the task and the time of the human being?
Answer: I confirmed that this was indeed relevant. When businesses consider adopting new tools, they evaluate the cost-benefit outcome. Therefore, it is critical to include the assessment of time spent on evaluating workplaces manually versus using the VRA.
Question: Would this tool be applicable in higher-risk environments, like manufacturing or oil rigs?
An audience member raised an important point about the applicability of the VRA in more dangerous environments and the legal and moral responsibilities associated with risk assessments.
Answer: I acknowledged the complexity of applying such a tool in high-risk environments. The EU AI Act’s implementation will likely require human oversight for such tools, ensuring that a specialist in occupational health and safety remains responsible. This tool can aid but not replace human judgment in high-stakes contexts.
Question: What happens from a legal responsibility point of view if the tool makes a mistake in high-risk environments?
(Follow-up by Bethan)
Answer: I addressed the legal and moral responsibility concerns, emphasizing that current legislation would not allow AI tools to operate without human oversight in high-risk areas. The AI tool is intended to assist, not replace, human evaluators, ensuring accountability remains with a qualified professional.
Question: Who controls and owns the AI, and who is responsible for it?
Answer: I explained that as long as AI is considered a tool, responsibility lies with its human owners and operators. Future advancements toward AGI (Artificial General Intelligence) may prompt legislators to consider granting legal entity status to AI, but this remains a complex ethical and legal issue.