b'Illinois Supreme Court supports the use of GAI, it expects lawyers to exercise judgment, maintain control over their work, and uphold their ethical obligations at every step. The Core Framework What is Generative AI? On June 12, 2017, researchers from Google and the University of Toronto released a groundbreaking paper titledAttentionIsAllYouNeed ,introducinganewtypeoftechnologycalledatransformer . 12This breakthrough is widely regarded as the watershed moment for modern artificial intelligence, and it forms the technological foundation for nearly all AI systems that lawyers are likely to encounter in their daily workflow. 13 Is All You NeedSoon afterAttention was published, researchers from OpenAI demonstrated that the new transformer technology could be configured to understand and mimic natural human language. 14They did this using machine learning techniques now commonly referred to as " model training " andfine-tuning.15Later in this Guide, we will discuss how model training might threaten client confidentiality and what lawyers should do to mitigate this risk. These core research efforts formed the basis of what is now commonly known as Generative Artificial Intelligence. The family of GAI tools includes two variations that frequently arise in the practice of law: Large Language Models , orLLMs , which can quickly generate, interpret, and summarize human language in response to prompts;16andVision Models , which can generate realistic images and video from basic descriptions or examples, and are often the subject of legal discourse regarding deepfake fraud and evidentiary manipulation. 17 The term "model" is a very important concept for understanding GAI, because the data processing behind most AI-enabled applications (including tools marketed to the legal industry) is typically performed by one of a few central companies that specialize in building, training, and licensing GAI models. 18These companies 12 Advances in Neural Information Processing Systems 30 Ashish Vaswani et al.,Attention Is All You Need , in 6000 (I. Guyon et al. eds., 2017) (paper presented at the 31 stIntl Conf. on Neural Info. Processing Sys. (NIPS 17)). 13John Berryman & Albert Ziegler,Prompt Engineering for LLMs(OReilly Media, Inc. 2024). 14Alec Radford et al., (OpenAI 2018). 15Numa Dhamani & MImagpgrioev Einngg lLearn, guage Understanding by Gene (rMataivnen Pinrge -PTurbailniciantgions 2024). 16Edward Raff et al.,Introduction to Gene (rMataivnen AinIg Publications 2025). 17ARDC does not curHreonwtl yL arregceo gLnainzgeu aa ugne iMveordsealls t Weromrk of art for image and video generation models. The term Vision Models is used in this Guide in a representative capacity to cover a range of technologies, including diffusion models, vision transformers, and other generative image architectures. For a useful discussion of this area,seeAmit Bahree, Ge Snuehraatsi vPea Ai, I in Action(Manning Publications 2023). (OReilly Media 2025) (explaining the broad landscape of 18 Designing Large Language Model Applicationsmodels and model providers in Chapter 5).Page | 5'