US Government Accountability Office GAO

For the purposes of this report, we use “model” to refer to the result of an algorithm “trained” on a set of data. Training is the iterative process of feeding data (called training data) through an optimization process to improve model performance. Of the generative models, foundation and frontier models are among the most capable. Foundation models are models trained on a broad array of data that can be adapted to a wide range of tasks – for example, summarizing text, writing code, composing music, or creating images. Foundation models may be fine-tuned or augmented to tailor to customer needs, or they may be integrated into multiple AI systems across a variety of areas. Frontier models are the most advanced foundation models, with respect to new or powerful capabilities, and may pose increased risks related to fact checking or contextual awareness of outputs and responses than foundation models.

Many generative AI models use natural language as input, meaning any text could be used as a prompt. But this ability also means the quality of the prompt has a significant effect on the response. Prompts serve as the starting point for a model’s generation. Models tend to respond like-for-like to input. For example, a prompt that contains harmful content is more likely to produce harmful output. Another difference is that it is usually less clear why a generative model produced a certain output. This “black box” effect is notable with the large neural networks underlying generative AI. As a result, reproducibility and accountability are lower, degrading the user’s ability to evaluate and understand the model’s inner workings.

larger generative AI models may have abilities that are not predicted by extrapolating from those of smaller models. Such “emergent” abilities may be unintentional or unexpected and may not be apparent until a model is fully developed or deployed.

In business, generative AI has the potential to assist with travel arrangements, analyze accounting data, and help manage investments. For example, generative AI may help answer traveler inquiries by generating information about destinations, attractions, events, travel services, local customs, and visa requirements. In the accounting and auditing sectors, generative AI may automatically create coherent, informative, and well-structured financial reports based on historical data, including balance sheets, income statements, and tax documents. This process may significantly reduce the operational risks of manual errors. In the finance sector, generative AI may play a role in digital advisory services and provide portfolio management services without significant human involvement. However, generative AI may also contribute to worker displacement, especially in jobs with routine tasks, such as administration or basic analysis.