Chatbots, Conversational AI, and Replicating Human Interactions
Updated: Mar 28
After encountering a chatbot on an e-commerce website, you may have wondered if it's possible that one day these virtual assistants will become advanced enough to be indistinguishable from humans. If you’re familiar with natural language processing (NLP) you may know that the field is critical in advancing this area. As observable in many sci-fi shows, we as humans seem particularly interested in technology that replicates us: our appearance, our abilities, and our speech. Making chatbots and conversational AI more human is a part of this desire.
Chatbots vs Conversational AI
Chatbots and conversational AI are terms often used interchangeably, but there is some nuance between them. The term chatbot is more often used for what you are likely to encounter on a company’s website, politely asking you to select from a range of questions you may be interested in. While chatbots can be based on AI, learning from the information you give it, they can also be rule-based, sometimes known as scripted chatbots, with only a certain set of responses triggered by very specific words. Meanwhile, AI-driven chatbots, also known as conversational AI, have broader abilities and are meant to interact in a way that replicates a human conversation.
While the applications of conversational AI are virtually endless, naturally the use-cases that fit into company and customer needs most immediately, and that are most achievable, have received the most attention, research, and use. Still, if you are interested in testing the skills of a conversational AI purely to see how well it utilizes natural language processing, there are several conversational AI out there designed to showcase AI skills, so they can give you an idea as to the current capabilities without having to ask about a specific product.
There are a variety of speech properties that are non-literal, and so are difficult for even very advanced AI to understand. In addition, an AI carrying out this kind of conversational task may have trouble following general conversation norms. One perspective on conversational norms was described by philosopher Paul Grice, who believed that conversations generally followed four rules, which he called maxims:
The Maxim of Quantity: Humans generally respond to questions with the requested information, and while they may add some anecdote or additional bits of trivia, there’s usually an understanding as to how much information was needed and how long the response should be.
The Maxim of Quality: Unless deliberate deception is involved, we will usually report in a conversation as fact only what we believe to be true. With the vast spread of misinformation via the internet and social media, if an AI is learning based off of what people tell them, they may unintentionally spread misinformation, and it may be harder to program an AI capable of fact-checking itself like a human could.
The Maxim Relation: We usually understand that when speaking we should not suddenly change the subject. While this could be expected as a very easy task for AI, they can only stay on subject if they can correctly understand what the subject is, which could be difficult. Chatbots designed to help customers often make mistakes in this regard because they may misdiagnose a problem based on an incorrect understanding of the key words the customer used.
The Maxim of Manner: We generally try to be clear about what we are talking about. This includes avoiding ambiguity. An AI may not be able to correctly sense when it needs to ask for clarification, and may not capture nuances properly to remain clear.
You may be thinking that you know many people in your life who break these rules on a regular basis, but when humans break these rules they usually break them in certain contexts and ways that are difficult for AI to replicate. There may also be self-awareness on the part of the speaker that they are not operating within normal parameters, such as acknowledging that they went on a tangent. Additionally, these conversational guidelines are not consistent across all cultures, and an AI may not recognize every unique conversational norm.
Breakthrough or Modest Achievement?
Efforts to market products and generate interest in AI may set unrealistic expectations as to their current abilities. When AI that could make phone calls to book appointments first went public, many initial reactions were impressed and amazed as to the human-like quality of the voices. It also brought up some reactions of fear as to how many human jobs could be replaced, with counter arguments that technological advancement should not be stopped; however, as time passed it has been revealed that many of these fears and arguments are premature.
While the human-like quality of some of these products is impressive, the comprehension of the AI is still not nearly as impressive and reliable: New York Times journalists Brian X Chen and Cade Metz found that 25% of calls through the lead technology in this use-case start with a human caller rather than AI, and 15% of the calls that did start with AI required human intervention at some point. Some may feel disappointed and others relieved at this reality, but regardless, it demonstrated the current limitations of NLP AI. This doesn’t mean that the technology isn’t impressive: rather, it requires us to re-evaluate and shift our thinking as to what advancements are impressive.
Real achievements in this area include natural linguistic imperfections speaking AI show in mimicking the tiny inconsistencies in human speech, in addition to being able to handle some level of uncertainty. These attributes are worthy of attention even if they don’t negate the need for human intervention. Small things like being able to comprehend an idiom or understanding a small contextual clue seem so simple to us because our brains understand NLP so casually it’s hard to appreciate how much work and programming it takes to enable an AI to get it right.
Ultimately, any improvement to conversational AI takes a thorough understanding of both NLP programming and language itself. Linguistics and computer programming are both relatively new fields, and exciting future discoveries could very well one day enable AI to be masters of NLP. Until then, AI are still fantastic supports in many different areas—For example, At Kai Analytics we often create Microsoft PowerBI Dashboards, which come with a built-in chatbot called Q&A that can quickly translate questions into an SQL language that the database can understand. While there are many tasks AI cannot do as well as humans, they can improve and enhance various areas of work so that both humans and AI can accomplish more together than they could alone.