**4. Learning techniques of virtual assistants and ethical considerations**

In self-service technology, virtual assistants are on the higher maturity curve and are expected to understand and interact with consumers as "humans" to provide information or take action. If we look under the hood of virtual assistants, then we uncover three basic technology building bocks.


In this section, focus will be on the conversational platform which has to be designed with ethical consideration. There are many types of technologies deployed for virtual assistants ranging from simple click based predefined options, to pattern matching, natural language understanding and natural language generation. In the section below, different types conversational platforms are discussed with a view on ethical considerations.

#### **4.1 Commercial virtual assistant platforms**

Most commercial virtual assistants use pattern matching and natural language understanding AI models. The primary task of the AI model in this case is to classify intent of the question for pre-defined set of answers. The assistants can also understand specific details in the text like country name or time and more. For example if asked "What is the weather in Singapore?" assistant will classify this as the request to find weather information and also extract Singapore country name. This information will be passed to backend API to retrieve the temperature and presented back as the answer. Example of these virtual assistants used by business are IBM Watson Assistant, Microsoft Bot framework, Amazon Lex, Google Dialog flow and more. Learning on these platform is generally supervised and the knowledge corpus is limited to the business use case. Sometimes, extension of these platforms is done where a large document corpus is ingested and most relevant document is brought forward to the user based on search and retrieval techniques.

In these platforms, it is the role of conversation designer and developer to ensure that the virtual assistant adheres with the ethical principles of Transparency, Justice & fairness, Non-maleficence, Responsibly and Privacy. Further, it is a good practice that document corpus is screened before being ingested into these virtual assistants to ensure relevant and proper responses.

#### **4.2 Mass market virtual assistants**

Siri, Alexa and Hey Google are examples of mass market, virtual assistants. These virtual assistants are pre-trained from a large language corpus and have the ability to retrieve personal information from calendar, phonebook, music, credit card and more. The organization developing these Virtual assistants publish their terms of service, privacy policy [27] publicly and it is consumers decision to understand and then interact with them.

The ubiquitous nature of these Virtual assistants poses a bigger question to society on how they should respond to different types of talk ranging from Rude talk, Abusive talk, Romantic talk or Suicidal talk. We discuss below two cases in detail, rude and romantic talk.

Rude Talk – the virtual assistants tend to respond back positively with information without prompts for polite or rude requests. This has an influence on manners especially in case of younger consumers [28]. For example "Alexa can you please tell me the weather forecast for today" or "Alexa weather forecast today" – the answer would be the same. These assistants should try to add nicer words like "Thank you" when consumers say "please".

Romantic Talk and Gender – when asked on gender, the virtual assistants tend respond on gender neutrality. However, by default they respond in a female voice. In the article by Jessi Hempel [29] in Wired she explains that people tend to perceive female voices as helping us to solve our problems. This also opens the door to romantic talk [30] for female persona based virtual assistants. Most of assistant are trained to handle this conversations by evading, or positively responding to consumers, but they rarely respond negatively [31]. This does extend in some cases to general acceptance of sexual harassment for assistants.

#### **4.3 Niche virtual assistants [open domain]**

A special mention here to Virtual assistants who can talk about anything in the open domain. These assistants are trained using sophisticated deep learning AI models (un-supervised learning), have billions of parameters and are closest to how a human would sensibly and specifically answer questions. Many gigabytes of training data (dialog response) is ingested in these AI models and it generates the answers (natural language generation) based on learning. Example of these assistants are


Other than Mitsuku which uses supervised learning, for other virtual assistants, it is difficult predict responses since they are learning from dialog-response corpus. In these cases, it would be beneficial to have a language filter that checks for ethical considerations like abuse words and more before presenting the answers to consumers.
