A flurry of language-processing products has created the false impression of major advancements in ‘conversational’ AI.
When the MIT Technology Review asked Bill Gates to choose 10 tech developments which would change the world in 2019, ‘smooth-talking AI assistants’ made the cut.
Siri, Alexa and the more utilitarian-branded ‘Assistant’ from Google all combine machine learning with voice synthesis technology to generate ‘conversational’ interactions with users, but these interactions still run along predefined scripts.
The enthusiasm that deep-learning innovations like sentiment analysis will enable customer service chatbots to have emotionally-intelligent conversations overlooks a key limitation: that the same statistics-based approaches are still being honed.
‘Hey Google’ is still a nod to early ideas about AI
Machine learning teaches AI assistants to predict what should come next in an exchange, mostly based on a statistical analysis of webpage-content. Buzzwords such as ‘neural’ belie the total disparity of this approach with the way humans learn to label their world and converse with one another. Until AI breaks free of machine learning, having novel conversations with Siri will be like teaching an old, big-data-driven dog new tricks – however seamlessly it might perform menial tasks in customers’ homes such as turning on an oven.
Even such targeted innovations as Seamless Speech Recognition from Mitsubishi Electric, which translates in noisy multilingual environments like airports, are basically developments arising from early incarnations of statistically-trained technology.
Machine learning will nonetheless shape the debate around human-AI interaction
The increasing capabilities and penetration of conversational AI still raise important issues. Last month, the pending use of voice assistants in criminal investigations in Germany made international headlines.
The OpenAI GPT-2 language generator also showed how a minor development in machine learning can have significant repercussions. The non-profit decided against an open release, fearing ‘malicious applications’ such as fake news.
Meanwhile, healthcare chatbots are using machine learning to diagnose patients based on symptoms communicated verbally. Healthcare professionals have expressed concern that information prompted by tactful human interaction will be missed.
However, hopes for a leap in the linguistic sophistication of artificial intelligence are being constrained by machine learning. Nonetheless, the current level of sophistication should give consumers and tech companies alike enough to think about regarding the long-term future of the technology.