The Federal Trade Commission’s recent decision to open an investigation into artificial intelligence developer OpenAI and its large language models highlights many of the unique aspects of the technology, and how it fits into existing legal and regulatory frameworks.
The FTC’s investigation primarily stems from a complaint against OpenAI filed by the Center for Artificial Intelligence and Digital Policy on March 30, 2023. The issues highlighted in the complaint are wide-ranging, but several recurring themes touch upon the fundamental challenges that are not only unique to OpenAI but also, at least in part, intrinsic to the underlying technology of LLMs as a whole.
These recurring themes include bias, harmful or offensive content, and the lack of transparency in how the LLM works.
AI’s Reflected Bias
The CAIDP has voiced its concern about bias in OpenAI’s language model, GPT-4. This critique mirrors a broader, long-standing concern about bias in AI, specifically LLMs. These sophisticated models are trained on vast datasets, including trillions of data points, that are invariably influenced by human bias. This bias is encapsulated in the data that the AI models learn from, leading to an inescapable reflection of these biases in the AI output.
The daunting task of eliminating bias from these massive datasets is almost impossible. Any dataset, no matter how carefully curated, is a reflection of our society and its inherent biases. Moreover, perceptions of what constitutes bias can vary widely, leading to disagreements about whether a specific dataset contains bias.
Therefore, as long as human society continues to grapple with biases, LLMs trained on data and information largely created or heavily influenced by humans are likely to mirror them.
Catching Harmful or Offensive Content
Another substantial issue raised by the CAIDP complaint is the generation of harmful or offensive content by GPT-4. At the risk of oversimplifying, LLMs operate on a probabilistic basis, choosing the most accurate grouping of words based on the input and contextualization. The model, however, lacks the ability to consistently discern whether a particular set of words is harmful or offensive.
A similar issue is the commonly reported “hallucinations,” where the LLM generates information not present or implied in the input data. This is an inherent consequence of the LLM attempting to generate a coherent and contextually appropriate response, and it can occasionally lead to outputs that are inaccurate, misleading, or potentially harmful.
However, “accuracy” and “harmful” are not concepts for which LLMs are built to account, and the “black box” problem addressed below adds a further challenge to addressing this.
OpenAI has taken certain steps to mitigate this issue by refusing to engage with prompts designed to elicit harmful or offensive responses. However, this filtering likely operates at a separate layer of the technology stack than the LLM itself, and guaranteeing that potentially harmful or offensive output is never generated is a tall order.
The addition of extensive guardrails could undermine the inherent benefits of the technology. Further, the availability of open-source LLMs will likely drive more people away from the large, centralized providers if guardrails make those products become significantly less useful.
Moreover, similar to the bias issue, the definitions of what constitutes harmful or offensive content are largely subjective and vary across cultures, adding another layer of complexity to this issue.
Transparency and the Black Box Problem
A significant concern raised in the CAIDP complaint is the perceived lack of transparency and explainability in GPT-4. This is a well-known problem in the field of AI, often referred to as the “black box” problem.
AI systems, especially those based on deep learning like GPT-4, are complex and intricate, often to the extent that their inner workings remain obscure even to their creators. In the case of LLMs like GPT-4, the model takes in a text prompt as input and generates a relevant text output. However, the process by which the model determines the best response, based on countless parameters and the nuances of language learned during training, is not easily explainable or understandable.
This lack of transparency can lead to uncertainty and mistrust, as users may not understand why the AI system made a certain decision or prediction.
OpenAI’s recent use of GPT-4 to analyze the workings of its earlier versions and glean insights is a clear demonstration of this problem. It shows that even the creators of these models need to employ the very same models to unravel the complexities of their own technology.
This inherent opacity of AI systems is a pressing issue in AI ethics and regulation, as it raises questions about accountability, trust, and control in AI-driven decisions. It’s a problem that the AI community continues to grapple with, and it underscores the need for ongoing research and conversation on transparency and explainability.
A Widespread Challenge
While the CAIDP complaint is specifically targeted at OpenAI, the issues brought to light are not unique to OpenAI. Other LLMs, like Google’s Bard, could potentially face similar challenges. CAIDP likely drafted their complaint about OpenAI because it currently dominates the AI landscape.
However, the themes raised in the complaint are of utmost relevance to the entire AI community. The issues of bias, potential for harmful content, and lack of transparency are problems that stem from the very essence of how LLMs are developed and operate.
These thought-provoking issues have far-reaching implications for the future of AI, and regulators and the global AI community will need to grapple with these concerns as AI technology evolves and permeates various aspects of our lives.
These issues also underscore the need for collective responsibility and dialogue to attempt to find solutions that balance the immense potential of AI with the need for ethical and fair practices.
This article does not necessarily reflect the opinion of Bloomberg Industry Group, Inc., the publisher of Bloomberg Law and Bloomberg Tax, or its owners.
Austin Mills is a partner in Morris, Manning & Martin’s corporate technology practice and serves as chair of the blockchain and cryptocurrency group.
Write for Us: Author Guidelines