University of Illinois System

Awareness

What is Generative AI?

Generative Artificial Intelligence (AI) can be defined as a technology that leverages deep learning models to generate human-like content in response to complex and varied prompts (pp. 10-11, Lim et al. 2023). Generative AI “learns” from large volumes of information — typically, significant portions of the open web — to answer questions in sentences, paragraphs, and pages of seemingly original text.

Artificial Intelligence (AI) is a field, which combines computer science and robust datasets to enable problem solving.

Generative AI can be defined as a technology that leverages deep learning models to generate human-like content in response to complex and varied prompts.

Chat Generative Pretrained Transformer (ChatGPT) is a generative AI tool that allows an individual to have human-like conversations with a computer program. ChatGPT can answer questions and create content, such as emails, essays, presentations, etc.

ChatGPT is a product created by OpenAI; however, Microsoft (Bing), Google (Bard), and others have competing offerings.

Multiple versions of product are available = different features/capabilites and complexity.

Generative AI is not limited to the creation of text. Can generate programming languages, images, video, music, and a human voice.

What is a Large Language Model?

A large language model is a computer program designed to understand and generate human-like text. It is trained on a massive amount of written information, like books, articles, and websites, so it learns patterns and rules of language. It is called "large" because it requires a lot of computing power and memory to run due to its complexity. But at its core, a large language model is all about using its training to understand and generate human-like text.

Generative AI applications that use large language models can produce text responses. However, the technology is not limited to the creation of text, it can be used to generate programing languages, images, video, music, and a human voice.

What are potential Risks?

Risk: Large Language Model — Unreliable Outputs

Some generative AI applications can produce unreliable outputs. Over time, this risk may be reduced but information provided by generative AI is not always accurate, requiring an individual to verify the output. Some of the common output risks faced by users are:

  • Factual Inaccuracies: Partially true outputs that are wrong on important details.
  • Hallucinations: Completely fabricated outputs. No actual ‘understanding’ of content; it simply predicts text.
  • Outdated Information: ChatGPT’s “knowledge” cutoff is September 2021.
  • Biased Information: Training data bias can result in biased outputs. Large language models can make their own language models that reinforce bias.
  • Copyright Violations: Outputs may resemble copyright-protected work.

Risk: Generative AI — Intellectual Property, Data Privacy, and Cybersecurity

Generative AI can introduce risk associated with intellectual property (IP), data privacy, and cybersecurity. This risk may vary amongst generative AI applications and over time may reduce, but it does require the individual using the technology to understand these potential risks.

Intellectual Property

  • Information entered into ChatGPT can become part of its training set.
  • Any proprietary, sensitive, or confidential information entered as prompts could be used in outputs for other users.
  • The ownership of output from ChatGPT does not belong to the user. At best, it is public domain, but it may contain other people’s intellectual property.

Data Privacy

  • OpenAI may share ChatGPT user information, including interactions with the platform and links clicked, with third parties without prior notice.
  • Third parties may include vendors or service providers, affiliates, and other users.
  • Users may enter personal or sensitive information about themselves or others (e.g., protected by HIPAA, FERPA), which ChatGPT may then store and use in future responses.
  • However, it is possible to request OpenAI to delete your data.

Cybersecurity

  • Personal or sensitive information stored by OpenAI could be accessed by hackers.
  • Hackers can also use “prompt injection,” or use prompts that can manipulate ChatGPT to give away information it shouldn’t.
  • ChatGPT can also be tricked into writing malware or ransomware codes.

Risk: Regulatory and Legal

The regulatory and legal landscape will likely not be able to keep pace with the rapid development and deployment of generative AI. New and existing laws will impact how generative AI is used. Examples of potential risks include:

  • U.S. and international law may restrict the use of tracking and monitoring technologies. For example, companies using AI chatbots supported by third-party vendors are currently facing class-action lawsuits for alleged violations of federal and state wiretapping laws. International privacy laws may also be implicated.
  • To the extent AI is used to generate content to be used by the University, there may be copyright considerations regarding the source of the content.

What are potential opportunities?

There are many potential opportunities associated with generative AI, below are four high-level examples:

  • Integration into Existing Applications: Generative AI has and will continue to be incorporated into existing software applications to improve their functionality. Generative AI can be used in content generation platforms to generate written articles and reports based on user inputs or specific parameters saving time and effort for content creators. It can be integrated into image editing software to automate tasks such as background removal, content-aware filling, and intelligent upscaling, making editing workflows more efficient.
  • Proprietary GPT Engines: Organizations will likely launch their own generative transformer using their repositories of information to create a ‘walled garden’ version of GPT. These can be used for internal use only — leveraging their unique domain or industry knowledge — or making their knowledge available for license as an adjunct to other generative transformers.
  • Monetization: Organizations may institute models whereby generative transformer platforms will be charged for training data. High-value training data repositories will likely have monetization frameworks in place to value their data, and GPT access will likely introduce a pricing model based on the quality of data.
  • Automation: Generative AI interfaces that enable autonomous task completion. Individuals provide a list of tasks to be completed and the system takes care of the rest. Unlike current generative AI, which requires detailed prompts, this technology generates its own prompts to complete the given goals, resulting in a more streamlined experience.

What's Coming?

No one knows for sure what is to come and that’s why it is important for individuals to understand and engage in conversations about the risks and opportunities of generative AI. In general, there is substantial uncertainty around how:

  • The technology will evolve
  • Lawmakers will regulate
  • Industry will or will not self-regulate
  • Society will understand and adopt this technology
  • Competition will drive new and different capabilities
  • Third parties will build off the open-source nature