While you can introduce new concepts to AI when it is chatted with, the mechanism for how this happens is quite different from how humans learn. Machine learning (ML) based AI systems can learn and grow with the data (even conversations). But there is no real-time learning from every single interaction. Rather, the models need huge sets of structured data and a model built to process that data over time.
As an example: The GPT-3 model from OpenAI itself can also be customized — it can be trained on new datasets, which means it could, for example, learn to be more in touch with the preferences of a specific type of user and focus more on certain subjects. To prove the theory, in 2021, IBM Watson showcased that if AI were trained on inputs particular to the domain, it could enhance its responses. But a user does not "teach" AI strictly by talking to it — conversations merely provide feedback that developers could use to retrain the AI, enhancing its performance and accuracy. This is usually a supervised process as human experts check the data before retraining can begin.
AI-generated customer service data indicates that 68% of businesses utilizing this method saw a proven improvement in customer satisfaction, following a retraining of the AI with data generated by users — like common questions or feedback. This data is available for AI to pinpoint patterns and trends to refine its response. An example would be that during a conversation between a user and an AI, a user might use a new term (even a new word); the AI will not learn that term right away, but it can be added to the training set for subsequent iterations through the knowledge base.
Supervised learning methods are commonly used when dealing with new information in ML models exploited in conversational AI systems such as chatbots. Specifically, this means AI developers monitor conversations to determine where the model is performing poorly, and supplement with new labeled data to enable more accurate predictions. This is how the Google AI corrects its search algorithms by increasing its outcomes through user-interaction and user-response. So, let us say you talk to a virtual assistant about a certain subject, while it will not comprehend your choice automatically, The more data and information it gets during the process, it learns automatically to align with the virtual assistant and describe and manage more similar conversations.
So in that sense no, you can not necessarily directly teach an AI by talking with it, but you can definitely provide knowledge that will end up helping train it. And this information is valuable data from which the AI can analyze, learn to improve accuracy and understanding in the future. talk to ai is a great experience that allows an understanding of the way AI evolves from time to time.