Artificial intelligence—or variants of it—have mesmerized people for millennia. From ancient myths of men made from clay and clockwork automata to killer computers and sentient computer networks, people have been obsessed with the idea of creating a being capable of original thought and consciousness. Since the arrival of ChatGPT in late 2022, however, AI has dominated the news—and not just the imagination—in an unprecedented way.

While most discussion has been around generative AI, for most of its history, AI hasn’t had much to do with large language models. Since the coining of the term “artificial intelligence” in 1955 by computer scientist John McCarthy, AI has referred to “the simulation of human intelligence processes by machines, especially computer systems”—typically abilities like visual perception, speech recognition, decision-making, and translation between languages.

To understand AI today, it’s helpful to learn some key terms that demonstrate the evolution of AI over the last several decades.

The Evolution of Artificial Intelligence

• Rule-Based Approaches: The first generation of AI used hard-coded rules that computers could execute faster and more efficiently than humans. Rules-based AIs can be extremely effective at games, speech recognition, and textual analysis. Deep Blue, the first chess AI to beat a reigning world chess champion, was a rule-based AI.
• Machine Learning: The second generation of AI programmers embraced machine learning, a type of AI that uses algorithms to learn from labeled training examples or iterative cycles of prediction and analysis of outcomes. Machine learning has applications across many fields, ranging from spam filters and Netflix recommendations to predictive coding in the review phase of e-discovery.
• Deep Learning: Modern machine learning algorithms are known as deep learning, meaning they take advantage of immense computing power and neural networks to assimilate vast amounts of data rapidly to produce more “intelligent” and accurate outputs.
• Generative AI: A subset of deep learning AI that can produce new content outputs (text, images, or audio) based on its understanding of the patterns and structure of its input training data. ChatGPT, Google Bard, DALL-E, and Midjourney are all examples of generative AI.

Organizations of all sorts are scrambling to integrate AI into their technology platforms and deliver the benefits of AI to their customers—but of course, there are risks. Here are four critical risks that privacy and legal professionals must navigate.

4 Critical AI Risks for Privacy Professionals

1. Privacy Risks

Large language models (LLMs) are trained on vast amounts of data, much of which is scraped directly from the internet. This creates significant friction with existing compliance frameworks.

Goli Mahdavi, Counsel in the Global Data Privacy and Security Practice Group at Bryan Cave Leighton Paisner LLP, explains:

“There are questions around reconciling the use of large language models and privacy compliance obligations under existing frameworks like the GDPR. For example, if a data subject submits a deletion request, what is required? Can you actually delete data from a model?”

2. Transparency Risks

A pillar of a robust privacy program is giving consumers transparency into how their data is processed so they can offer clear, specific consent. Generative AI complicates this because the data pathways are rarely linear.

Christie Hawkins, Partner in the Consumer Financial Services, Data, and Technology Practice Group at Akerman LLP, notes:

“Do you know enough about the AI tool to explain it to someone in a disclosure? AI can be very complex. Does transparency mean that we have to describe everything the tool is doing behind the scenes? Organizations should at least be able to tell someone that the AI tool is being used or applied, how decision making takes place, and what the consequences might be for the consumer.”

3. Intellectual Property Risks

If data scraped from the internet is protected intellectual property, companies using it to train their LLMs face mounting lawsuits for copyright infringement. Furthermore, Mahdavi points out that there are unresolved legal questions regarding whether artistic collaborations and other outputs that blend human and machine creativity will be eligible for IP protection at all.

4. Algorithmic Bias Risks

Organizations using AI for high-stakes modeling—such as employment decisions or credit/loan evaluations—risk codifying and magnifying historical discrimination and automated bias.

Hawkins highlights how easily recruitment algorithms can accidentally enforce demographic disparity:

“Let's say a company wants to introduce DEI initiatives to help the company identify roles for which new job applicants may be well-suited. If part of the data that you use to train the AI is the data of successful past employees, then we really need to drill down here and look at the data. Is it over-representing certain characteristics or groups of people? Bottom line, when AI models learn, they learn from the data. But they can teach themselves in ways that we don't intend.”

Action Plan for Responsible AI

To build a legally defensible and ethically sound data governance framework around AI implementation, legal and privacy teams should download the full Exterro whitepaper: 4 Keys to Using AI Responsibly.

‍