4 species of Chatbots, underestimated by humans

It took Humans millions of years to evolve the ability to have conversations and develop language. 15 different species of humans roamed the earth back then with different physical and mental abilities. Fast forward today, all those species either died, were killed or mated and now there is only one kind (homo sapiens) left and we never think about sharing earth with other forms of humans anymore.

Now, we feel lonely, so much today, that with Artificial intelligence, we are trying to create synthetic beings with the same conversational capacity that we evolved over millions of years. We call them chatbots, or conversational agents at times. But there a tiny problem here, we humans don’t entirely know or remember how we evolved it for ourselves, so, we are kind of starting from scratch.

It won’t take us million years to do the same thing again, for sure. But with all the efforts going on, chatbots do have evolved into 4 different species of intelligence that have different conversational capacities.

Humans often confuse these abilities and species with each other (since they look all the same) and then blame that all such conversational agents are bad. Well, just like you don’t hire a monkey to do a doctor’s job, you can’t deploy a chatbot of different specie to do something very different that it was never evolved for.

And this is the reason why some people say ‘chatbots fail’, cause we haven’t understood yet which specie of chatbots need to be deployed in what situation as per their intelligence capacity.

So, this article tries to break it down into 4 types of synthetic beings that we see today.

Specie #1: Click Navigation Chatbots

Scientific name: Chatboto logicious

These chatbots come with preloaded ‘If-then-else’ logic, created through a pre-designed conversation, generally sourced from the most common conversations that a business holds with its customers day in and out.

Situations where they are successful:
So, if you are a hotel, the most common conversation could be — New Reservation or booking shuttle service. These chatbots work well for lead capturing and replace those boring ‘contact us’ forms.

And where they fail?
Any customer support request or something out of their conversation flow (even if the answer is right there in the conversation flow), would trip these bots.

Tools recommended: Most of the Facebook chatbots being created today using tools like ManyChat or ChatFuel fall in this category.

Pros: Easy and cheaper to build
Cons: Doesn’t really sound intelligent. Can’t handle anything out of the conversation flow designed by the creator.


Specie #2: Keyword based Chatbots

Scientific name: Chatboto Keywordilius

These type of chatbots depend upon a huge list of keywords preloaded by the creator and are matched with chat messages coming from the user. As soon as there is a match, it triggers a relevant response by the chatbot, assuming that it got the sentence right. These chatbots sound a little more intelligent than the one above and do not stick to a conversation flow.

Situations where they are successful:
For example, if you are a web hosting company, your most of the conversations fall in the long tail category. You can’t really do 80–20 of your customer interactions and design a chatbot around it. This is where Level 2 seem to do a reasonable job. So, answering questions like these would be supported:

  • What are your working hours?
  • Where is the cafe located?
  • How do I install wordpress in a docker?

And where they fail?
The above approach is infamously called ‘keyword bag’ approach, and is looked down upon by AI researchers, and there is a reason for that: The system doesn’t has any understanding of what is actually being said.

So, in a restaurant chatbot, if a dialogue comes like “I want a cheese burger, hold the onions”, the bot would never understand the ‘onions’ part and may load it up with extra onions.

Tools recommended: ManyChat and Chatfuel may claim that they allow it, but it is very limited (to a level, it almost sounds a marketing gimmick). A better way to implement it would be using IBM Watson or Google Dialogue Flow, that provide much more flexibility to navigate this.

Pros: A little more human and user friendly.
Cons: Can make really stupid mistakes in responding to sentences that are a little more complicated than straight forward or untrained questions. Furthermore, these systems as as good as they have been trained and cannot learn new things by themselves.


Specie #3: Natural Language understanding Chatbots

Scientific name: Chatboto Naturalionus

This is where things get interesting. This usually is built using complicated AI model that is trained on domains (like a human) to make it understand sentences, phrases or complete email conversations even.

Situations where they are successful:
These bots can actually hold a conversation with the user and sound really intelligent if Natural Language generation is also thrown into it. There have been incidences where these bots were just communicating on email with clients and clients would literally walk into the office and ask for this person, who doesn’t exist.

These bots are usually deployed in processes that need automation at scale and may not be deployed to converse with the user at times. Banks or Insurance companies deploy such bots to read the incoming customer emails and at times reply them as well. Ecommerce companies are beginning to deploy such bots to do a better job at matching products with search queries, than fuzzy matching they depend upon right now.

Unlike their cousin (Chatboto Keywordilius) these may have a learning capacity and can pickup new knowledge automatically as they go along reading loads of data.

And where they fail
The conversations coming from the bot at this point may sound so natural to participating humans, that humans start to expect missing human elements from the conversations — specially certain range of emotions like courtesy, empathy and goal orientation. For example, the bot cannot be trusted to handle sensitive conversations of insurance claims for veterans who lost their lives in wars, or acquiring beta testers for a product, involving cold emailing and likes.

Tools recommended:
Usually, such chatbots are built from scratch with home grown AI engines (or deploying loads of hacks on Google Dialogue flow or Watson, which may become very complicated at times), unless there is a vertical specific solution that addresses the same. For example, for Ecommerce Search problem, you will find a few solutions that do way better than IBM or Google in this context (since they are pre-trained on a massive database and have experience in dealing with this problem)


Specie #4: Aware Chatbots

Scientific name: Chatboto Jarvis-ious, evolving to Chatboto Pepper-onious

This is the God level of chatbots and the current technology has not reached to this level yet. An example of this chatbot is seen in sci-fi movies like JARVIS in Iron Man.

Situations where such bots are successful:
Beating robotic villains like Ultron

And where they fail:
While facing all powerful organic intergalactic villains like Thanos

Tools Recommended:
Largely unknown. But try mixing arrogance of a scientific genius with intergalactic infinity stones.

Evolution by Mating the Chatbots

Different species of chatbots can mate with each other to get a hybrid offsprings. A common practice is to mix Logicious (Specie #1) with Keywordilius (Specie #2). So, the evolution is already in progress, where the process of Human selection governs ‘survival of the wittiest’.

Next time someone says ‘a chatbot is stupid’, just evaluate if that chatbot is stuck in a wrong job (like a lot of humans are), from where it can’t even resign. They don’t have rights, you know.