Do customer service chatbots need to pass the Turing test?
Experts say organizations should focus on enhancing the customer experience with chatbots -- not so much on passing the legendary Turing test for artificial intelligence.
For decades, artificial intelligence researchers and chatbot developers have been chasing the vaunted Turing test -- in this context, a bot's conversational responses and cues passing as those of an actual human -- as an aspirational benchmark.
Sounds simple, but the conversational demands of human discourse on enabling technologies such as speech recognition, natural language processing and AI have largely kept the Turing test out of reach to date. Most experts believe it'll still be years before AI can legitimately pass the test.
The question for businesses that are increasingly using AI-driven customer service chatbots is whether the Turing test even really matters when it comes to customer experience (CX).
"Obviously, the Turing test is the holy grail regarding technology. But at the end of the day, the real litmus test is the CX data," said Michael Replogle, senior contact center consultant at CustomerServ.com. "If consumers are happy with the outcome, it really doesn't matter what the Turing test results are."
Turing test is still a goal
The whole drive for getting customer service chatbots to pass the Turing test is to "lend a level of authenticity and warmth" to interactions with customers, said Kevin McMahon, director of mobile development at SPR Consulting.
But he warned that it is shortsighted to consider this the be-all and end-all of successfully building customer rapport. Instead, organizations should focus on outcomes: What is the chatbot there to accomplish, and is it getting that task done for customers in most initiated interactions?
"If chatbots can convince your customers that they're human, but can't resolve their issues, what are you really accomplishing?" McMahon asked. "Your customers will conclude that this 'human' can't address their needs. It is far more important to drive successful outcomes in every interaction than it is to convince people they're dealing with a human."
If chatbots can convince your customers that they're human, but can't resolve their issues, what are you really accomplishing?
Kevin McMahondirector of mobile development, SPR Consulting
This is an important big-picture point to keep in mind, especially considering that many researchers working with AI and chatbots believe the technology still has a long way to go.
"I think that, in terms of where the technology is now and where it needs to be to pass the Turing test, there are still a lot of advances we need to make," said Ryan Lowe, a researcher and doctoral student in computer science in the Reasoning and Learning Lab at McGill University. Lowe is supervised by Joelle Pineau, who most recently co-developed AI and deep learning systems designed to evaluate the quality of chatbot dialogue.
"Our systems -- even deep learning systems -- are still not very good at holding general conversations," Lowe said.
According to Lowe, it is very hard to peg how far along the industry is, but convincing general conversations are definitely five or more years down the road. Part of the difficulty in estimating progress is because a lot of advanced research on the matter is isolated, and data remains unshared across the industry.
As Lowe explained, to be very good at conversational AI, an organization needs to be able to tap into a ton of machine learning data and compute power -- the likes of which organizations such as Microsoft, Facebook and Google have at their disposal. But these companies are also not likely to publicly release such data due to privacy concerns.
The Turing test is used to determine if a computer can think like a human.
"Plus, in the chatbots they release, there's also a lot of hand-engineered rules included, so it is hard to disambiguate exactly how good the state of machine learning is now with these chatbots," Lowe said.
So, as the industry works through the next five years or so, it is important for organizations to keep an eye on the prize and remember that even without mimicking humans, customer service chatbots can add tremendous value.
"Just as an ATM doesn't need to pass the Turing test to be useful, a chatbot doesn't need to imitate human conversation to be useful," said Beerud Sheth, CEO of Gupshup. "In many cases, advanced conversation may even be counter-productive. For example, at a fast-food restaurant, the person who takes the order efficiently is preferred over someone who would discuss the 'philosophy of life.' Similarly, a chatbot that helps the user accomplish her task with the least conversation will be preferred over an excessively chatty bot."
The role of customer service chatbots
The goal should be to use customer service chatbots to help take the heat away from human reps, automating the answering of repetitive, frequently asked questions, said Mitul Tiwari, CTO and co-founder of Passage AI.
"In many cases, questions asked by customers have typically been asked before," he said. "In situations like this, a chatbot is able to immediately respond to customers' questions, while reducing the workload on customer service agents. Instead of replacing human agents, bots can free up the time of call center agents and allow them to add more value for customers and the business."
According to Lowe, this supplementary rather than replacement role of customer service chatbots is probably going to remain for a long time -- though he said he believes we'll see a continuing decline in just how many humans various organizations need to fill customer service requests as things become more automated.
The key is keeping the customer experience in mind as companies transition between people and machines for interactions -- it's what some companies call the blended AI model, and it can be a nightmare if a customer gets "trapped in fully automated interactions with no way out," said Rémy Claret, a product and solutions marketing director focused on AI at Genesys.
Claret said that while he believes the Turing test is interesting to judge the quality of natural language conversations between humans and machines, ultimately, it doesn't matter. What matters most is the overall customer journey.
"It does not help measure the end-to-end experience -- seamless handoffs between bots and agents with carryover of the customer context," he said. "The customer does not really care whether the bot's behavior is 80% human-like or 100% human-like, as long as the intent is very well understood and the response is personalized and accurate [for] an end-to-end experience."
