What’s AI?
Humans have long been enamored with the idea of artificial intelligence – creating machines that can “think” like us. The foundations of today’s AI go back decades, at least to the dawn of the computing age in the 1950s. We’ve come a long way, but to understand what AI is, let’s quickly look at what it’s not.
We tend to think of AI as being mostly like us, and popular culture has done much to reinforce that stereotype. But artificial general intelligence is still very much the stuff of science fiction. A handful of futurists think that we may achieve machine self-awareness by 2030, But that seems ambitious since it requires an understanding of the human brain and its neural networks that we still don’t fully have a grasp on.
Computers don’t have consciousness, and don’t inhabit the world in a way that gives them context. So they’re really only capable of the tasks we set them to. That can sometimes be quite a lot because of the massive amounts of data such algorithms can process. In recent years we’ve managed to create systems that can accomplish some impressive tasks, like recognizing different types of objects. That’s driven real-world innovations in everything from spam filtering to driverless cars.
But this sort of “narrow AI” isn’t the stuff of humanoid robots. Simple tasks like walking or picking up an object are still incredibly difficult for machines. And they must typically be trained by teams of researchers, whereas humans can learn things on their own, even as babies. Even setting asides the complexities of controlling a body and the distinction between AI and robotics, we are very far from creating Jarvis or Mother (or Skynet, for that matter). In general, AI is far weaker and narrower than we imagine.
Narrow but Powerful
So, it turns out we have a long way to go before we create the AI of our popular imagination. But narrow, or “weak,” artificial intelligence is getting more sophisticated as we continue to develop new techniques. Machine learning, for example, is a broad subset of AI that allows algorithms to process more data than a human ever could and to make predictions based on that data. It was breakthroughs in machine learning techniques not that many years ago that catapulted AI, for years a relatively obscure corner of computer science, into one of the hottest fields of research.
Machine learning, especially when combined with the massive sets of data collected on the internet, provides methods for us to train algorithms to match certain criteria. That’s been incredibly useful for applications of computer vision, for instance, without which self-driving cars would not be possible. It’s also driven medical advances, allowing for more accurate diagnosis of certain conditions and diseases. Many fields and applications require recognition of commonly repeating patterns, and that’s a task well suited to powerful machine learning algorithms.
More recently still, AI researchers have learned to layer their algorithms in ways that more closely mimic the structure of the human brain. The idea here is to build algorithms based on human neural networks. These deep learning techniques allow systems to evaluate and process data in somewhat more organic ways. So rather than needing to train an algorithm towards a specific task, such systems are able to classify and cluster inputs in a rudimentary form of learning. It’s not “the singularity,” but it’s a lot more sophisticated than the AI of even just a few years ago.
Natural Language Processing
One of the most fertile grounds for such techniques has been in natural language processing (NLP). This would (among other things) give machines advanced language processing abilities, such as learning to parse text, evaluate meanings, analyze sentiment, translate between languages, and even generate comprehensible paragraphs of text on its own.
Language is difficult enough for humans, so it’s no small challenge to replicate linguistic skills in machines. Machines don’t “understand” language in the way that we do. Rather, they reference a previously compiled dataset in order to parse it and respond to it in some way. Previous generations of AI have been capable enough to answer question from a customer and even to write an original song, but they were still very recognizable as not quite human.
The GPT-3 language model is currently at the pinnacle of deep learning and hopefully is a precursor to even more intelligent applications. It can produce human-like text incredibly well and it’s available to anyone since it was developed by Open AI. Using multiple layers of neural network architecture, GPT-3 learns contextual relationships between words. In fact, it has been able to produce memes, short stories, music, press releases, technical manuals, text in the style of specific writers, and even computer code. It can carry on some interesting conversations.
Another deep leaning algorithm created a podcast. MIT used AI to co-create horror stories. Alibaba Group’s machine-learning technology was reportedly better at reading comprehension than humans, as measured in accordance with the Microsoft Machine Reading Comprehension dataset. SEO writers are already concerned that this type of system could easily take their jobs (though this is a constant trope with practical technology of any kind).
Facebook is working on an AI system that can translate between computer languages, potentially taking some of the pressure off legacy systems running COBOL and freeing up programmers to concentrate more on system design than syntax. Natural language processing can even translate normal human language into computer code, allowing anyone to create apps without even learning a programming language at all.
Some researchers do find even our newest NLP capabilities overrated. And they do sometimes produce non-sensical replies. Perhaps more importantly, it replicates some of our dangerous biases (a recurring issue with AI). But it’s undoubtedly a leap forward. And such capabilities should get us thinking about new and creative ways that natural language tools can be used for better, more immersive, and more efficient brand interactions.
Useful Tools
IBM, Microsoft, Google, Facebook, Amazon and others are all developing and improving their own NLP platforms. These tools will be trained and hosted in the cloud and made accessible through application programming interfaces (APIs), which allow for sharing and training the algorithms on a massive scale.
With the development of cloud-powered deep learning algorithms and language processing, we’ll see an evolution in the applications and uses for this cutting-edge technology for brands. One obvious application of sophisticated language processing would be for chatbots. Tomorrow’s chatbots will go well beyond the “how can I help you today?” decision-tree models that people tend to get frustrated with. In fact, in some cases, it’s already possible to build a chatbot so good that most people won’t know they’re interacting with a non-human entity.
Pretty much all of the existing AI platforms have strong chatbot capabilities. Google’s language processing service is called Dialogflow (previously known as API.ai). Amazon has its own platform named Lex and IBM provides the Watson Conversation Service. Microsoft calls theirs Azure Bot Service. Then there’s wit.ai which is focused on language processing for the internet of things (IoT). And there’s even an open source platform called RASA.
These tools can all help process massive amount of verbal or written language to determine customer intent, understand slang, and translate between languages. Which platform is right for which project depends on the particulars of what must be accomplished and also on what other capabilities and system need to be integrated or considered. (But brands can always hire capable creative digital agencies to help with questions like that.)
Language Processing in Action
Chatbots (like most things) exist on a spectrum. Some only need to handle very simple requests. Chatbots like that can be rule-driven and give customers only what they ask for, or follow a decision tree based on a narrow set of parameters. But others aspire to be more sophisticated, and using NLP engines like the ones above can add greater capabilities. And with greater capabilities also come more opportunities for creative innovation.
One particularly cool feature is their ability to converse in the linguistic style of a fictional character. Themed interactions like these can be a big boost to customer engagement. For example, Disney created a chatbot on Facebook Messenger that allowed fans of the film Zootopia to have an interactive dialogue with some of the film’s characters. Similarly, Marvel allowed people to chat with super heroes via Facebook and Twitter DMs. These branded bots can be great way to create a connection with fans, especially as they become more adept at human language.
Language processing has also been used efficiently in museums for years and has even been combined with robotics to create automated docents, with many adopting IBM’s famous Watson system to help answer questions of visitors. The Smithsonian’s deployment of Softbank’s “Pepper” robot has been a hit with visitors, giving them a more joyful museum experience with the goal of making them feel more present and encouraging them to immerse themselves in the art and architecture around them rather than checking a phone app for more information. Other AI chatbots are standalone features that engage visitors passing by, hoping to spark conversation about the collections.
On our phones, we can download chatbot guides to accompany us through museums and give us both educational material and a sense of purpose, instead of leaving us to wander around. The Akron Art Museum has the Dot chatbot both as a phone app and in kiosk form (hoping to engage passersby and spark conversations). Dot can even connect you via Facebook messenger to other people at the museum if you want to talk about what you’re seeing. Meanwhile, the Museum of Tomorrow in Rio de Janeiro collaborated with IBM to develop IRIS+, a chatbot that uses Watson’s advanced language skills to guide visitors, connect them to others, and help them follow up later with any ideas their visit may have sparked (such as getting involved with an environmental cause).
Conversational UI
The chatbots we’ve come to know have traditionally been built around text-based interactions. But as language processing improves, these systems can engage in conversation on more human terms, at least theoretically. As a language processing bot becomes more capable, it eventually becomes a full-on conversational user interface (UI).
With voice recognition and text-to-speech capabilities, conversational UIs can leave the text world behind and carry on spoken conversations. That’s quite compelling, as it’s how humans generally prefer to communicate with one another. Of course, anyone who’s ever tried to have a conversation with Siri or Alexa knows that there are serious limits to conversational UI. At times these interfaces fail us or annoy us by misunderstanding our requests. But the more data they collect, the better they can be trained to respond more naturally in the future.
Text models have been used during the pandemic to help quell loneliness, and to alleviate anxiety about medical issues for those with no one to talk to. And as the baby boomers age, there’s a good chance these conversational interfaces will be useful for older adults to engage with as well, both for companionship and a way to augment health care. At the current state of NLP, these solutions still fall short, since the goal is to make people feel understood and display some type of genuine empathy. An algorithm can only mimic that – but it’s still better than nothing.
Conversational UIs have great potential to be used for more entertaining purposes as well. Video games are a great example. People can interact with characters in role-playing games with more authenticity if they can do it conversationally. Games that use VR and AR are already incredibly immersive, but they can become even more so when combined with conversational UIs. Even for basic navigation the possibility of using voice interactions really takes things to the next level. We would no longer have to rely on dropdown menus and dialog boxes to make things happen or engage with other players. All of that would be built in seamlessly.
Similarly, any type of activation or application can make use of voice input. A wide range of consumer devices already work well using voice commands. In many cases the use of full conversational UI will tend to make such interactions more difficult rather than easier because of their current limitations. But if kept simple, voice interactions is already viable today, and at the pace of innovation we’ve seen in NLP it probably won’t be long before more sophisticated conversational UIs can be leveraged for a broad range of purposes.
What’s Next?
It’s clear that natural language processing holds enormous possibilities for customer engagement. Of course, most brand experiences will still depend on human interactions. The goal isn’t to automate away those interactions, but rather to make machines more pleasant to interact with so that they can augment human capabilities and make human interactions more meaningful.
Chatting with machines sounds simple, but the underlying technology (like all AI) is more complex than it sounds. It will still be a few years before we can create seamless systems that can engage with customers on a more authentic level. But we can already accomplish amazing things with the current generation of platforms, especially if we us it in creative ways. And there’s no reason we shouldn’t experiment with these technologies, especially in ways that are circumscribed and monitored. In fact, we stand to gain a treasure trove of data and customer feedback from these experiences, paving the way to even better customer engagement down the road.
Featured Image: Photo by Nick Fewings via Unsplash