Artificial Intelligence (AI) has been spreading across industries and growing in prominence and acceptance and the emergence of ChatGPT on the scene this year has launched the conversation into the stratosphere. ChatGPT has put a very public face on one of the most significant areas of progress, natural language processing (NLP). The intuitive interaction that happens in a conversational format allows the user to ask a simple question and, fueled by powerful processing algorithms, the NLP generates the answer. However, there are two major questions we have to ask ourselves: Can we trust the answers that AI engines generate, and how do we know that we can?
The validity of an AI-generated response depends on the type of question, the data being queried, the data that the AI has been trained on, and the complexity of the answer. Knowing when and how to use AI is a skill all its own, and it’s a critical one to develop if you want to use AI to leap ahead of the pack.
Deciphering the hype
The Gartner Hype Cycle identifies the phases in the life of a new technology. Right now, AI is sitting squarely atop the “Peak of Inflated Expectations” with tools like ChatGPT and Google Bard being touted as solutions for a massive variety of problems.
Those AI tools are a great solution…for some problems.
AI engines that are designed to operate on large data sets can process vast amounts of information in a short amount of time. This allows them to provide solid answers to straightforward questions, such as weather forecasts or sports scores, with a high degree of accuracy.
But a tool like ChatGPT doesn’t automatically give you the source of their information. You can—and should—ask yourself where those responses are coming from. When we do ask ChatGPT where it gets answers we get a fairly transparent explanation:
“As an AI language model, I have been trained on a vast amount of text data, including books, articles, and websites. My training data comes from various sources, such as news outlets, scientific journals, encyclopedias, and online forums. Additionally, I have access to real-time information through APIs, which allow me to retrieve data from various sources, such as news feeds, weather services, and stock market data. However, it's important to note that I do not have personal experiences or emotions, and my responses are based solely on the information contained within my training data and the inputs I receive from users.”
Some of those sources, such as scientific journals and encyclopedias, are great. It would take us hours to search through that data on our own, making ChatGPT a powerful option for finding answers in those areas. But that information can be mixed in with other sources, like online forums, which makes the validity of the final result unpredictable.
The bottom line is that how, where, and when we use AI tools matters.
Creating transparent, trustworthy AI
When it comes to more complex questions that require an understanding of context, nuance, and human emotions, AI engines will struggle. We might ask a question like "What is the best way to motivate employees?" An AI engine could generate an answer based on data from studies and surveys, but that won’t tell us the whole story.
AI-generated answers also have significant room for bias and error because they are only as unbiased and accurate as the data they are trained on. (Remember how ChatGPT is accessing online forums? Yikes.) If the training data contains bias or is simply wrong, then the AI engine will generate wrong answers. We’ve already proven that this is a problem, as we’ve watched AI systems perpetuate systemic discrimination in hiring situations or lending operations because they were trained on biased data.
To address these shortcomings, data scientists are working to develop more robust and transparent AI systems. “Explainable AI" (XAI) algorithms provide a justification for the answer generated by the AI engine, allowing users to understand how the answer was reached and whether any biases were present.
Another approach is to use "human-in-the-loop" (HITL) systems, which combine the strengths of AI with human expertise. In HITL systems, the AI engine generates an answer, which is then reviewed and validated by a human expert. This approach can help ensure the accuracy of AI-generated answers while also reducing the potential for bias.
Developers are working to create all kinds of AI solutions. But are they solutions that companies need?
What users really want from Generative AI
Most companies are still trying to decide what they want and need in AI tools, but their first priority should be to select secure, transparent tools that bring real value to their organization. Sure, features like natural language query (NLQ) can feel like magic…but if the magic is just an illusion, or doesn’t help your business strategy, then what’s the point?
Combining technology helps us get more insight out of our dataset. For example, Knowledge Graphs are an incredible way to visually represent data and the communities and connections present, but exploring them has typically required someone who could write sophisticated queries. But if you were to combine a Knowledge Graph with a Large Language Model (LLM) then you would have the ability to ask a question and explore that data visually, no code required.
Companies should be looking to empower data analysts to do far more than they can with BI tools that were never designed to explore complex data. Knowledge Graphs and Natural Language Queries are just two of the many possibilities delivered in the Virtualitics AI Platform. Our focus is on Intelligent Exploration: using AI to find insight and guide analysts as they explore their data. With the right advanced analytics tools, organizations can put data science capabilities in the hands of their analysts so they are empowered to discover opportunities, do deep analysis, and get stakeholder buy-in.
Are you ready to level up your data analysts and put AI to work? Schedule a 1:1 demonstration with our team.