Building Trustable AI
Although nearly 80% of businesses report that they are either using or planning to use AI (explodingtopics.com/blog/companies-using-ai), consumers and businesses alike remain skeptical of the technology. A May 2024 article in the Harvard Business Review referred to “the AI trust gap” and cited disinformation, security, and the black box problem as the top three concerns (hbr.org/2024/05/ais-trust-problem). But creating trust is possible by providing visibility into the information sources, governing the quality of the underlying content, explaining the development process, and offering methods of validating the outcomes.
Building Confidence for GenAI
Some of the deepest suspicions about AI arise from generative AI (GenAI), which creates newly created content in response to prompts after being trained on a large language model (LLM). However, in order to produce responses that are accurate and meaningful to a particular organization, curated information about products and services needs to be incorporated. This can be achieved in various ways, including retrieval-augmented generation and providing additional training, or tweaking, for the model. Verification methods can be incorporated, including human oversight and automated inquiries of the data to look for specific facts known to the organization, to ensure accuracy in responses.
When successfully implemented, GenAI can leverage corporate knowledge in an efficient and reliable way. One leading tax analyst firm developed a content creation platform using Progress Semaphore to generate complex tax and legislation reports for its customers. Its manual process using subject matter experts was expensive due to the complex nature of the data environment and the need to create a unique report for each customer. The Progress Semaphore platform assisted users in finding and assembling information through an interactive content creation process. Based on saving each employee 1.5 hours a week, estimated annual savings were $30 million.
Confident in the knowledgebase being used to create the reports, the company then wanted to use GenAI for the same purpose. In this case, the GenAI application was doing the research on its own, writing summaries, and creating the reports. The production of the finished report was reduced to just minutes. “Considering the time-savings, this translated into an even higher level of savings,” said Philip Miller, senior product marketing manager for AI at Progress, where he oversees the messaging and strategy for data and AI-related initiatives.
Progress Semaphore is a software solution that helps people find meaning and insight by connecting and contextualizing data. It does so through advanced data classification, extracting facts and information from unstructured documents. Using taxonomy and ontology to classify information and relationships, workersthen create a knowledge graph in a user interface-driven environment. As a result of this processing, the resulting information is clean and relates appropriately to the organization. The search function is carried out by the Progress-owned MarkLogic search engine. The knowledgebase can be used to generate responses to prompts or to support chatbots and other agentic AI actions.
The system is designed to validate the information, checking to see if the AI has followed the rules, which are set up using natural language in combination with the LLM. The combination of symbolic AI as reflected in rules, along with neural network analysis, provides the strength of both deterministic and non-deterministic analysis and is referred to as neurosymbolic AI. “This combination was dismissed for many years because the sophisticated neural networks were not yet available to support it,” Miller commented, “but now the two can be combined in AI platforms.”