If you are using AI models to generate texts or as a customer chatbot, you should be careful that they are prone to errors. They cannot be completely trusted. Even famous and sophisticated models like GPT-4 make errors, let alone other simpler ones. In 2022 Air Canada Chatbot promised a bereavement policy out of its creativity. This plan however was totally hallucinated by the model and the company did not aim to do it. Yet in February 2024 a Canadian court confirmed that the airline should pay the bills.
That is because the models are using probability to connect the nodes and fine tune the model’ however truth in many cases is not probabilistic. The drawback is that the same feathers that make errors in the AI models are exactly the same that makes them useful. In generative models the next words are chosen based on the probability distribution function. Mathematically it means that for every word there exists a non-zero probability that it becomes chosen for each position in a sentence.
There are ways to address the issue, but all the methods to alleviate the problem come with a cost. The models that are “fine tuned” are more sharp and powerful but at the same time they make mistakes more frequently. In the process of fine tuning, the model encodes the statistical relationship between word and phrases. A model that is not fine tuned may answer many questions with “I don’t know” no mistakes, but not impressive as well!
There are other methods to limit the error probability, for example the AI can be restricted to choose the first two or three words with the higher probability for each vacancy. This would decrease the mistake probability but also kills creativity as well.
According to scientific research, GTP-4 hallucinates in 3 % of cases, Claude 2 in 8.5 % of cases and Gemini Pro 4.8 %.
Source: Economist
