At the IFASS meeting, members explored the potential and limitations of AI in standard setting through two presentations. The first presentation introduced AI, including machine learning, deep learning, and generative AI, which can create new content. This AI subtype, large language models (LLMs), can generate human-like text but lacks critical thinking abilities and requires constant, costly training.

The second presentation showcased an LLM trained for disclosure analysis, particularly analyzing sustainability reports. The analysis of 11,000 reports highlighted the need for careful review of AI-produced results, as many generated statements needed more support and had inaccurate citations. The study concluded that fluency often compromises accuracy.

In the context of standard setters' work, the possibility of training an LLM was discussed due to the specific and low-volume literature. The potential administrative applications, such as drafting meeting minutes, were also highlighted. Although using LLMs for comment letter analysis seems appealing, it was noted that the results would still require review. However, this could facilitate searches for specific comments. Employing multiple LLMs trained for different perspectives, mediated by another model, was suggested to generate valuable arguments that could stimulate original thought.

Access the full agenda and topics discussed during the IFASS meetings.