AI that can pass the Turing test has been something of a holy grail for some computer scientists since mathematician Alan Turing developed the test to determine whether a computer uses language well enough to be mistaken for a human. In 2018, Google showed that its Duplex AI-powered assistant can book salon appointments and restaurant reservations over the phone with real people. The public response ranged from impressed to skeptical to nearly panicked, as industry watchers and scientists speculated about possible applications for voice AI that sounds truly human. One possible use is for better fraud prevention. But the same technology could also make fraud attacks worse. Here's what small business owners should know about what AI can do now for you, what new generations of AI may be capable of and how to safely use AI for business.
There's some debate in the data science community about whether Google Duplex opens the door to more complex and wide-ranging AI applications that pass the Turing test and whether those advances might happen anytime soon. Regardless, even applications like Duplex that are narrow in scope could eventually change the way online payment fraud is managed. The defining element probably won't be the technology itself but the intent of the people using it.
How AI that passes the Turing test might help prevent fraud
People can already chat, text and talk on the phone with robots, but the conversations, like Google Duplex's, are limited in scope. It's also easy to throw today's customer service robots off track with an unexpected question or heavy accent. But if customer service AI – especially voice AI – eventually passes the Turing test beyond the very narrow scope that Google has done, that technology could make it easier for companies to stop credit card fraud by taking over some challenging but crucial phone conversations.
Right now, online credit card fraud (card-not-present, or CNP, fraud) is a growing problem for businesses of all sizes, in part because so many criminals have access to breached consumer data. There are two ways for companies to handle online orders that – after analysis of a number of factors like shipping address, IP location and purchase history – seem like they might be fraudulent but might not be.
The first option is to simply reject these iffy orders. The problem with this approach is that if the order wasn't fraud, the rejected customer may be so frustrated, embarrassed or angry that they never shop with the company again. This is a common and expensive problem. In 2016, merchants lost $2 billion more to these false declines of good orders than they did to actual CNP fraud.
The second option is to manually review each suspicious order. In some cases, fraud reviewers have to call the customer to make sure they really did make the purchase. These voice conversations require skill to collect information without seeming accusatory, and they require company resources to make calls at a convenient time for customers no matter what part of the world they're in.
Because manual review is resource-intensive, not all companies can afford to do it in-house. Others try, but find that process stalls the order approval process during the holidays and other busy seasons. Voice AI tools that pass the Turing test could take on some of that workload and, with enough data fed into the system, perhaps become better than humans at spotting subtle verbal clues that indicate fraud. In the meantime, outsourcing to a third-party manual-screening provider can reduce the burden on businesses.
How fraudsters might exploit AI that passes the Turing test
Of course, such advanced voice AI could also be exploited by criminals who want to commit fraud more efficiently. As cybersecurity expert David Gerwitz wrote shortly after Google's voice AI demonstration, more advanced versions of this technology could be used to scam people over the phone. Voice AI that passes the Turing test could, in theory, impersonate consumers calling in orders, customer service reps calling with questions or specific individuals like politicians appearing to call for donations.
AI systems can also be duped by the inclusion of visual or audio messages that make the neural network think it's dealing with the opposite of what it sees or hears, even while things appear normal to humans. Researchers have already shown that they can embed instructions to voice assistants in audio files that humans can't detect. In one case, an audio news story contained "silent" instructions to disable the target's alarm system and unlock the front door. Imagine that sort of attack aimed at e-commerce customer service AI to override fraud controls and deployed at scale.
It all comes back to humans
These scenarios are far-fetched for now, but it's a given in cybersecurity and fraud prevention that each new advance brings new risks. There's no reason to believe voice AI systems that pass the Turing test will be any different. That means that when companies can finally use Turing test-passing voice AI to protect themselves from fraud, they'll also need to develop safeguards to thwart criminals who want to use that advanced technology against them.
What can small businesses do to get the benefits of today’s AI without incurring needless risk?
First, understand what current AI tools can do. Voice assistant devices, customer service chatbots and fraud-prevention algorithms are powerful tools that can save your business time and money, as long as you combine them with human oversight to prevent breaches, poor customer service or false declines of good orders. Next, be cautious about using AI-powered personal assistants to handle critical business tasks like alarm system management and banking. Finally, keep up with the news about the kinds of AI your business uses to stay informed about new features, functions and security issues as AI gets more sophisticated.