Chatbots Explored
Written by Sue Gee   
Sunday, 26 February 2017

One of the areas of AI that is already having an impact on a day-to-day basis is that of the chatbot. More and more often often when you ask questions on a shopping or reservation booking site it's a bot that answers rather than a human. We asked Justine Baron of Recast.AI to tell us more about the current state of chatbot development.


I-Programmer: Recast.AI bills itself as a "Collaborative bot platform" which can help you build a conversational bot. Can you tell me what you mean by a "chatbot" and provide a few examples in different areas. 

Justine Baron: Recast.AI is indeed a bot creation platform. Chatbots are a sub-category of bots: they are automated programs capable of understanding, responding to human language and trigger actions. Whenever there is conversation, you can add a bot! Here are a few examples we love: 

For more examples check out's curated  lists of favourite French and English bots.

Much attention is currently being paid to Amazon's Alexa, which has voice recognition - would you consider Alexa a chatbot?

Absolutely. By definition, Alexa is a program that understands humans, and triggers actions from their commands. That fits into the definition of a chatbot, a very complex one of that.

What elements of AI are used to create a chatbot?

AI is a vast field, ranging from computer vision (what autonomous cars use to find their way on the road) to spam detection (Google inbox is a good example), through decision making (high-frequency traders make more than 10,000 req/s).

To build a bot, you'll need parts of AI, but not all of it.

The key task is natural language processing (NLP): it is a research domain aiming to make computers understand natural language. It is by detecting words, understanding their relationships in a sentence and detecting their grammatical connections that AI helps us create bots.

But problems occur when you need to process 1000 sentences per second. That's why we need automation, and that’s where machine learning comes in. The point is to go from bits of information (datasets) and end up with a mathematical model able to predict/classify a new, unseen information.

So to build an agent, you need two things: to process and understand natural language, and automation with machine learning algorithms. And that’s what we provide.

What programming skills are needed to use the Recast.AI platform?

Today, developing a bot requires developing skills. At Recast.AI, we want to make the experience hassle free. Any developer comfortable enough in one programming language, knowing how to use an API, can build an efficient bot.

We are all familiar with horror stories where chatbots go wrong - Microsoft's Tay for example. How can you avoid repeating these mistakes?

Microsoft Tay was a fun experience. While it failed to distinguish right from wrong, it was very successful in learning from user interactions. After all, it’s the people interacting with it that fed it the inappropriate content.

That however showed the limits of the approach Microsoft used, where the bot is learning as the conversation goes without any human supervision. Even though this approach is very interesting in other branches of machine learning applications, bots are not ready to be based on this technology. We believe they should be trained with semi-supervised or supervised training to avoid such events.

Also, bots today should be created to solve or enhance one precise use case. That way, they can be extremely good at what they do. Building general AI is a very difficult task and we’re not quite there yet.

Can you share your wish list for chatbot features you'd like to improve on/include in future?

Here are things we’re excited about: 

  • bringing true context awareness to bots

  • automatize training in a sensible way, while still giving humans full supervision

  • and be able to broadly use bots in groups on messaging apps

Which areas of AI are going to be incorporated into the Recast.AI platform

We’re looking into: 

The Turing Test is seen as the key test that a chatbot needs to pass.  What do you think it will take for humans to be convinced that a chatbot has real intelligence and is this desirable.

As mentioned above, we believe general AIs aren’t our best bet, as the technology isn’t ready yet. However, building specific bots already provides very good results today. With an enhanced context management, we’ll get great bots!

Do you see any dangers if we start to consider chatbots to be human?

When building bots, the question of the name, gender and personality is extremely important. We often advise our clients to let the users know the bot is, well, a bot. Pretending the bot is human only raises expectations so it's better disappoint them sooner rather than later, because, well… a bot isn’t a good as a human today.

Further along, we can raise questions about the humanization of AIs and everything that goes with it. While that might be an issue, it’s still very far away. Before that, we’ll get to see augmented relationships, augmented industries, even augmented humans! That’s much more exciting.




Recast.AI is one of the Media Partners for AI With The Best an online conference which has one of its four tracks devoted to NLP/Chatbots.

Tickets for this event, which let you attend the event and have access to  replays for 15 days, booked before April 16th normally cost $60 ($20 for students).

But if you book via this link using the code IPROGRAMMER you can take advantage of a discount - 50% until March 14th and 25% after that date. 


More Information


AI With The Best

Related Articles

AI With The Best Online With Geoffrey Hinton

What Makes A Bot?

Passing The Turing Test Brings It Into Disrepute

Loebner Prize Judges Could Easily Identify Chatbots

New Amazon Alexa Skills Contest

Turing's Test, the Loebner Prize and Chatterbots


To be informed about new articles on I Programmer, sign up for our weekly newsletter, subscribe to the RSS feed and follow us on Twitter, Facebook or Linkedin.




Amazon Bedrock Adds Support For Anthropic's Claude3 Opus

Bedrock, Amazon's fully managed service for building generative AI applications, has been enhanced with support for Anthropic's Claude 3 Opus Foundation Model.

Apple's Walled Garden Is Crumbling

Or is it? Depends on which side of the wall you are on. Apple is doing all it can to keep control and more importantly keep the revenuse from its App Store, but from the outside it seems to be losing  [ ... ]

More News

raspberry pi books



or email your comment to:

Last Updated ( Monday, 27 February 2017 )