What do hallucinations, articles on rectangular land surveys, and pirate ships in coffee cups have in common? They all can be generated by artificial intelligence (AI). UMRA members learned that and much more about generative AI from our April 23 Forum guest speaker, University of Minnesota School of Law Professor Daniel Schwarcz. 

Generative AI is a form of machine learning. It is different from the computer programs we are familiar with that involve setting rules for computers to follow. Instead, with generative AI, humans feed extremely large amounts of data into a virtual neural network and the program teaches itself in ways that are not totally understood. 

The goal, when creating text, is to generate a series of words that make sense. A training algorithm provides feedback to the virtual neural network and, eventually, a coherent sentence and paragraphs are created. Human feedback enhances the learning. 

The end result is a language prediction model that remains static but is used to generate responses to questions. Currently, the Generative Pre-trained Transformer or GPT-3.5 and GPT-4 versions are in use. GPT-4 was launched in March 2023 and is 10 times more advanced. A new, greatly improved version is likely to be released soon, Schwarcz said.

AI is best at creating summaries of existing material, Schwarcz said. He demonstrated this by asking the application for information on rectangular land surveys in the Minnesota territory, as suggested by UMRA member Rod Squires. Squires, a scholar of land surveys, was less than impressed with the depth of the resulting summary but agreed it was accurate. 

Hallucinations

Schwarcz said that when AI can’t find information it creates incorrect predictions called “hallucinations.”

Schwarcz has conducted research on student use of AI. In his work, students who were at the low end of the grade distribution benefited from use of AI, while those at the upper end were hurt. He thinks that is because AI replaces independent thinking and requires the student to take time to edit and remove hallucinations. Similar research involving writing classes showed AI did not improve the quality of the work product, but students were 20-30 percent faster. 

Other research presented various standardized exams (AP, GRE, LSAT, and bar exams) to AI, and compared the scores with the results from actual people. Two years ago, AI generally scored better than 20-30 percent of the actual test takers. When the research was repeated one year ago, AI often scored better than 90 percent of the people. This shows the speed at which AI is improving, Schwarcz said. 

Generative AI can also create non-text output and, for a bit of humor, Schwarcz showed several animated images created by AI based on his prompts. The video of pirate ships fighting in a coffee cup is one that will stick in my mind. 

So, is generative AI a good thing or a bad thing?  

Schwarcz did not provide a definitive answer, but he did raise numerous questions about AI’s future.

• Will its exponential growth rate in learning continue? 
• Do we have enough data to teach AI? 
• How does copyright law protect people whose data (writing) is being fed to AI models? 
• What will AI do to educational and cultural attitudes about expertise and hard work?
• Will employers use AI to support or replace humans? 
• How do we keep humans in the loop when AI makes critical decisions? 
• How do we regulate AI to prevent harm? 
   

The answers will determine whether we rue the day AI was let out into the world or praise the technology for its incalculable benefits. 

—Julie Sweitzer, UMRA president-elect and Program Committee chair