6 minute read
AI and You
Exploring the Boundaries, Benefits, and Dangers of Generative AI
Artificial intelligence is not going to steal your job. Artificial intelligence is not going to destroy the human race. But this powerful new technology is here, and you will be using it.
Chances are that you have been using it already. If you’ve dictated a text message into Siri, asked Alexa anything, or used predictive text in an email, then you have used AI in one form or another. It’s been integrated into our daily lives in subtle ways already.
When ChatGPT launched this past spring, AI flexed its muscle with a chatbot that produced seemingly thoughtful responses to inquiries of all things imaginable. Generative AI gave its best presentation of mimicking human interaction.
What ChatGPT showed is AI’s ability to combine increased computing power with the availability of massive data sets to make quick analysis of patterns and create unique content. ChatGPT shows the future of AI is in a data-driven world, where the system can compute enormous amounts of data, and generate responses to that data.
“It’s not necessarily new, but ChatGPT shows us the new capacities unlocked by the scaling of the parameters,” says Kai Shu, Gladwin Development Chair Assistant Professor of Computer Science at Illinois Institute of Technology.
“There’s been a lot of research over the past decade in the scaling of parameters and massive training data. And now we’re seeing the phenomena.”
Shlomo Argamon, professor of computer science at Illinois Tech, has spent many years conducting AI research, and says generative AI, such as ChatGPT, performs the same fundamental task as text prediction, but draws upon a much larger amount of surrounding context due to the huge data sets that have been used to train highly complex and sophisticated statistical models. This results in its ability to construct long, intelligible responses to inquiries, and mimic human-like behavior.
“This is a tremendous leap forward,” Argamon says. “Until ChatGPT, systems couldn’t converse and seem nearly human-like.”
Although generative AI systems can produce responses that seem like human interaction, they do not incorporate anything that we would think of as understanding, reasoning, and knowledge. In fact, these systems have no concept of reality. The responses are more akin to hallucinations generated to look plausible based on analysis of the massive data sets used for training the system. As a result, these systems can generate responses that contain made up “facts,” which can be inaccurate.
“For example, if a student asks it to write a term paper complete with references, it will do that,” Argamon says. “However, none of the references will actually exist. They look realistic but are just made up.”
Most of the facts contained within the paper will be accurate, as the majority of the data the AI system is analyzing to generate the content will be correct, but not because the system “knows” it’s correct. It also won’t be drawing on a limited number of references that it can list. So when it is asked to generate references, it will do just that. It will create references based on the patterns it finds within the data it is analyzing.
AI systems also can be exploited to deliberately generate misinformation by malicious users. Deep fake technology can use speech patterns to generate video showing people saying things that they never said. Generative AI technology can be used to develop news stories that aren’t factual to intentionally mislead the public.
“A lot of future research will go into exploring how large language models can generate misinformation in order to combat it,” Shu says. “There is still some action needed to be able to detect AI-generated misinformation.”
Malicious users also can raise the ugly head of bias through AI in a number of ways. Biased training data can be incorporated to lead systems into generating biased responses. Systems also can be programmed to extract private information from protected data, such as an individual’s race, ethnicity, age, and gender, or even their health information.
There are three stages in the life cycle of training data, and Shu says mitigations can be implemented in each stage to reduce bias.
During pre-process, data is collected. Bias within the data can be determined during this stage, and scrubbed before it is used as training data. During the in-process stage, data is actively used to train the AI model. A predictive layer can be added to detect and mitigate biased information. The model is fully trained by the post-process, but bias can still be mitigated at this stage. By examining the system’s responses, it can be determined what biases it is producing and steps to mitigate them can be taken before the system is deployed for real-world use.
“There are a lot of challenges along the way from research to deployment,” Shu says. “If you are going to deploy AI in the real world, it must be done in a responsible and trustworthy way, or people aren’t going to be able to rely on it.”
Other mitigation includes making algorithms transparent so the user can understand how AI systems draw the conclusions it reaches. New policies also should be developed, specifically in the areas of data privacy and security, Argamon and Shu argue.
Although the results that AI is producing is impressive, there is little knowledge as to how, exactly, these systems operate.
“AI is more of an art than a science at this point,” Argamon says. “The capabilities of these systems are beyond what we could have believed, but how they work is still poorly understood.”
However, the capabilities make AI a valuable tool that has been widely used in fields, such as health care, engineering, software development, and many more. New jobs have emerged with the deployment of new AI systems.
“You’ll see that prompt engineer is a new job,” Argamon says. “It might not always be a standalone job. It most likely will be absorbed as part of another job. But at this point, it’s a new job.”
Shu and Argamon agree that AI, when integrated into the workplace and under human supervision, has the capabilities to fulfill mundane tasks and increase worker productivity. Tasks such as scheduling, writing briefs, or filling out spreadsheets could be completed more efficiently with AI, freeing time for workers to work on more important duties.
“It can take over some grunt work,” Shu says. “There are potentially a lot of benefits to enhance productivity and efficiency.”
Illinois Tech faculty have long been leaders in AI research in a variety of areas, including enhancing cybersecurity, understanding health issues, and brain imaging. Argamon and Shu are jointly working on an AI project to determine authorship of specific documents that is being funded by a $1.6 million grant from a $11.3 million contract from the Human Interpretable Attribution of Text Using Underlying Structure program of the Intelligence Advanced Research Projects Activity. The pair are working to develop “AUTHOR: Attribution, and Undermining the Attribution, of Text While Providing Human-Oriented Rationales.”
AUTHOR promises to capture the writing style of specific uncredited authors through natural language processing and machine learning techniques to create stylistic fingerprints. On the flip side, AUTHOR will be used to develop authorship privacy, anonymizing identities of authors, especially when their lives are in danger.
AI research has also crossed fields at Illinois Tech into areas such as engineering, law, and finance.