The Dark Side of Generative AI:
A Taxonomy of Negative Possibilities

By: Mark Cummings, Ph.D., Zoya Slavina, Katarzyna Wac, William Yeack, CSE

Much has been written in the last year about the negative side effects of generative AI. This article provides a good general understanding, including a general background and taxonomy, of the negative side effects, as a way to organize thinking about them.

The development of generative AI (GenAI) started in approximately 2017 and its power is rising exponentially. This exponential growth is likely to produce an expanding palette of benefits as well as negative side effects.

GenAI technology is based on Large Language Models (LLMs). LLMs use symbols. Symbols can include alphanumerics, phonemes, images, brush strokes, musical notes, etc. LLMs consider all symbols as a part of a language model. LLM-based systems operate on the basis of probability. An oversimplified explanation of LLMs is that they start with a symbol (letter, word, ideograph, image, etc.) and determine what symbol is most likely to follow. Based on this process they can create new content. This ability to create new content is a fundamental differentiator from previous forms of AI resulting in a quantum leap in capability. To achieve this, these models are trained on massive data sets. For example, data sets crafted using internet/web crawlers. That is, a capture of all the information that is publicly available on the Internet. Some models are also trained on private information. In conjunction with this data, the models use a huge number of parameters—tens of billions or more. How these, often non-transparent and black-box, large probabilistic systems produce their results with potential negative side effects is still not well understood. Yet, the resulting benefits are so attractive that they are being implemented widely, despite the potential for negative consequences.

The negative side effects of GenAI that have been documented to date are organized into the taxonomy presented in Figure 1.

Figure 1. Taxonomy of GenAI Negative Side Effects

The first two major categories under GenAI are negative side effects and benefits. There are likely to be more negative side effects that have not yet been documented, and more are likely to emerge. However, it is probable that many of them can fit into this taxonomy.

To differentiate between positive and negative effects, financial gain/loss and quality of life (QoL) factors are used.

The first major categories of negative side effects are Cybersecurity and Hallucinations. Cybersecurity involves the unauthorized use of communications and computing resources. Hallucinations is a term that is emerging in common usage for