Like all financial companies PayPal worries about fraudsters, armed with stolen credentials, logging into a legitimate customer's account and using a credit card linked to it.
"We want to stop these guys at our door," said Hui Wang, senior director of global risk and data sciences at PayPal.
The company has more reason than most to worry, given its high visibility and massive payment volumes. It generates $10,900 in payments every second, and it handled 4.9 billion payments in 2015 for 188 million customers in 202 countries.
To detect suspicious activity, and more importantly to separate false alarms from true fraud, PayPal uses a homegrown artificial intelligence engine built with open-source tools. The effort has forged a bond between its human fraud analysts and its AI program.
The human detectives train themselves to think like fraudsters and like law-abiding citizens as they examine real-life cases that have triggered fraud alerts. They develop scenarios for good and bad user behavior that they then feed into the artificial intelligence program to "put this human intelligence into production," Wang said.
"It's a unique combination of human story-based intelligence, data mining, and machine learning," she said.
In so doing — or at least in being willing to talk about the use of artificial intelligence in fraud investigations — PayPal is ahead of many big banks. Arin Ray, a Celent analyst who has studied banks' use of artificial intelligence in anti-money-laundering and anti-fraud work, said that he has not seen banks apply artificial intelligence to transaction monitoring or alert-investigation, although some may be doing this quietly. "I know for sure that many are actively considering it," he said.
PayPal has been using machine learning technology for 10 years. Wang's team of computer scientists and mathematicians build predictive models for a variety of uses around the company; a primary one is fraud detection.
In the beginning, the group used simple, linear models. Over the years it has migrated toward machine-learning technologies such as deep learning — software that uses so-called neural networks to mimic the workings of the human brain and recognize patterns in digital representations of sounds, images and other data.
"What we enjoy from more modern, advanced machine learning is its ability to consume a lot more data, handle layers and layers of abstraction and be able to 'see' things that a simpler technology would not be able to see, even human beings might not be able to see," Wang said.
A linear model can consume 20-30 variables whereas deep-learning technology can command thousands of data points, she said.
"There's a magnitude of difference — you'll be able to analyze a lot more information and identify patterns that are a lot more sophisticated," Wang said.
A key goal for the fraud team is to reduce the number of false alarms that can overwhelm a system so that the company focuses on real financial crimes. Another, of course, is to avoid blocking the accounts of legitimate customers.
Wang offers an example. If traditional analytics software sees a pattern of the same account being accessed by five internet protocol addresses within five days, it will flag that as suspicious. Yet PayPal's AI system can look at each situation more closely and see, for instance, that the user is a pilot buying gifts for his family while on the job.
"In order for us to identify that, we need to feed this information to the artificial intelligence machine, so the machine will be able to say you know what, there is a good side of a seemingly bad story," Wang said. "Then the machine will be able to provide a good assessment."
The system might be able to recognize that some of the IP addresses are coming from airport Wi-Fi networks, for instance, and use device geolocation to show the user was in those airports.
As a result of this collaboration of humans and AI software, Wang said PayPal has cut its false-alarm rate in half, Wang said.
When legitimate customers do get caught in the fraud detection net, Wang said the company will send them questions to prove their identity. It will then feed information about the incident back into the AI engine to help it learn the difference between true fraud and misleading situations.
PayPal has not reduced its human workforce in fraud and risk analytics as a result of the computers doing more of the work, Wang said. The machines have freed the human detectives up to identify new types of fraud patterns, which they can then inform the AI machine about, she said.
Yet banks have been slow to adopt artificial intelligence in anti-fraud operations for a few reasons.
An innovation chief told Celent's Ray that the technology was too new and fast-changing to bet on now.
"One concern is, how much change will you need to make to your existing infrastructure to accommodate the new technology," Ray said. "If it's large-scale change, human nature would prevent you from getting onboard immediately because there are large costs involved and you're not sure of how it will play out."
Another concern is making sure regulators are comfortable with the new technology. Bankers "won't want to answer extra questions later like, 'Why did you use this type of solution here?' " Ray said.
Where he does see large banks going ahead with artificial intelligence is in the analysis of unstructured information, such as checking customer names against different lists and monitoring social media.
For instance, if a customer sends a certain amount of money every month to a recipient who is on a watch list, that bank might need to investigate the situation and potentially report it to authorities.
"Once you have these abilities to look through large volumes of data such as at Facebook, LinkedIn or other large provided lists, that expedites the process and reduces manual efforts," Ray said. "The final call needs to be made by the analyst who is on the job, who will have to follow whatever policies and guidelines have been laid out by the bank."
Editor at Large Penny Crosman welcomes feedback at firstname.lastname@example.org.