When reaching out to customer service of any major company, it is very likely that the first interaction happens via a chat interface with a chatbot on the other side. It is usually easy to tell that one is interacting with a bot due to its unnatural and pre-written responses. However, recently, artificial intelligence (AI) has made significant advancements towards creating human-like text from user input, to the point where it is difficult to say if one is interacting with a human or a machine [1]. These technologies could change communication and constitute a wide range of applications in business, industry and healthcare. Yet, as with many new technologies, there are concerns about the ethical application of AI and its potential misuse [2-4]. In this article, we cast light on how generative AI works, and explore its positive and negative impacts.

Generative AIs are systems that can produce unique output, such as text or images, on-demand based on user input. To achieve this, two components are crucial: the architecture of the AI and its training data. An AI’s architecture defines its capabilities and is specific to its application. For example, text-generating language models, such as the currently most capable ChatGPT, use probabilities to predict the words of a sentence one after the other. How these probabilities are assigned, i.e. which word is chosen next, depends on the context of the sentence and on text that the AI has seen previously–the so-called training data. This data is used during the development of an AI system to ‘teach’ it what type of content it should produce. In the case of text generating AI, a model trained on poetry will likely produce poems, while a model trained on statute books will produce legal texts. ChatGPT was trained on more than 570 terabytes of text from the internet, which allowed it to generate text across a wide range of topics and styles [5,6].

One potential application of language models is in customer service. Companies can, for example, use them to generate automated responses to customer inquiries, which can improve response times and reduce the workload on customer service agents. This technology has already been implemented by several companies [7]. Another potential application of language models is in the medical field. Researchers are exploring their use to analyze electronic health records and assist with diagnoses. For example, researchers successfully used language models to analyze electronic health records and identify patients at risk of sepsis, a potentially life-threatening condition [8,9].

While the likely benefits of AI and language models seem clear, also negative impacts must be considered. One major concern is that biases such as racism and sexism could enter AI systems via the training data [10-12]. For example, in 2016 Microsoft launched a chatbot named Tay on Twitter, which was designed to learn from interacting with users through conversations. However, within 24 hours of its launch, Tay had become a prolific spewer of racist and sexist content, which led to its rapid shutdown [13]. To avoid such scenarios, AI developers are now being more careful with the selection of training data and both the inputs to AIs and its outputs are filtered for inappropriate content [14].

Another issue of current AI systems is the generation of false information. For example, upon the release of chatGPT, multiple journalists reported that the AI made up scientific studies and presented false information in a very legitimately sounding manner. This happens because the main task of a language model is to generate natural sounding and contextual sentences, not to supply accurate information [15]. Developers are working to improve this issue, but so far the most practical solution seems to be to advise users to fact check the AI output [16].

Looking to the future, it seems clear that AI will play an increasingly important role in our daily lives. However, since this technology is still relatively new, it is likely that its general application will have unintended consequences and it is difficult to predict how it will impact our communication. One thing is certain, however– the time of identifiable chatbots is coming to an end.