The Power and Danger of AI for Businesses

Sharp UK's Director of Transformation and Security, Matt Riley, explains the differences and dangers of Generative AI and Semantic Search.

Artificial Intelligence, commonly known as AI, is intelligence demonstrated by machines, as opposed to the intelligence of humans and other animals. It involves ingesting information, learning, and providing outcomes based on that learning. A terrible example of this occurred back in 2016 when Tay, the Microsoft AI chatbot based within Twitter, had to be taken down after 24 hours. Twitter users tweeted Tay with offensive and demeaning remarks so Tay assimilated these traits and responded accordingly.

Currently, there are two types of AI in use: Generative AI and Semantic Search.

Generative AI

Through the media, many people will be familiar with ChatGPT and Bard which are two forms of Generative AI. You ask questions and it will generate responses for you based on a ‘chat’ with the AI. What it generates is frankly brilliant but it should always be taken with a pinch of salt. It is returning responses based on its learning and responds based on the probability that an answer is correct. It is prone to bias, relying on the training that it received being correct. It will make conclusions on more data than humans can understand and therefore is difficult to check (known as the black box problem). It can also ‘hallucinate’ this is where it doesn’t have the data so takes an educated/probability-based approach to guess the answer.

An example of this was during Meta's demonstration of Galactica, an LLM (large language model) specifically tailored for science researchers and students. When prompted to generate a paper on the subject of creating avatars, the model referenced a fabricated paper on the same topic, purportedly authored by a real expert working in a related field.  

Semantic Search

Semantic Search is a data searching technique that aims to understand and interpret the data it is searching to return the most accurate result. A simple example of this is holiday vs. vacation. If you worked in a multinational company and wanted to understand annual leave requests across the group, you would run into a problem. Part of the business uses ‘holiday’ and another uses ‘vacation’. A Semantic Search AI tool is designed to understand the context of these two words so would return a result based on the understanding that they both mean the same thing.

Some positives of Semantic Search are the natural language understanding, the ability to handle complex queries, and the ability to decipher user intent. However, Semantic Search algorithms can be complex and difficult to interpret, leading to concerns about transparency and biases in search results.

So what is the danger?

Outside of the potential bias, debates over consumer privacy and unclear legal regulation, unreliable training, the black box problem, and hallucinations, the main challenge for businesses is where their team members use AI. AI learns from its inputs; if you feed it confidential data it will learn it. If you entered a secret recipe from a well-known brand into it, then it would learn it. Trade secrets, intellectual property, confidential information, financial information or anything sensitive should never be input into an AI or chatbot, otherwise, it potentially becomes permanently public.

10 Interesting Facts About AI