Before You Use The Word ‘Deep,’ Show Me The Learning

Winter is coming — an AI winter, that is.

Artificial intelligence (AI) winter, or a period of reduced interest and investment spending in the field, is not imminent, per se, but Brighterion founder and CEO Akli Adjaoute does think it could happen if people keep hyping AI the way they have been.

In a recent podcast, Adjaoute and Karen Webster agreed the terms “AI” and “machine learning” have been thrown around so much that they’ve nearly lost all meaning. As a professor in the field for 20 years, Adjaoute feels the buzzword burnout more acutely than most.

The CEO believes many companies who paste terms like “AI” and “machine learning” into their product descriptions are just telling customers what they want to hear. In such an oversaturated market, wondered Webster, who can even identify the difference between a good AI and a bad one?

“I would not even use the words ‘good AI,’” Adjaoute said. “Few topics have generated more hype than machine learning and artificial intelligence. I would not be surprised to hear about a dog food created with artificial intelligence. It’s just become nonsense.”

This numbness to the concept of AI could not come at a worse time, however. Adjaoute remarked that AI — true AI that is personalized, adaptable and self-learning or, in other words, actually intelligent — is one of the few tools that will remain effective against fraudsters, who are themselves quite adaptable.

It is hardly news that legacy fraud solutions just aren’t cutting it anymore, although Adjaoute has his own take on why. But, according to unpopular opinion alert, he doesn’t think AI and machine learning solutions are cutting it, either — at least, not the way they are now.

Adjaoute and Webster have spoken before about the surprising age of AI — it was born in the 1950s — and the minimal change it has undergone since then. Technology’s biggest fans tout it as if it were cutting-edge, but that is no longer the case. Data mining, neural networks and deep learning have been around since the early years of AI. Criminals have had 50 years to learn how to trick them.

“There is real value in machine learning and artificial intelligence, but you really need to use the right one,” said Adjaoute. “Before I use the word ‘deep,’ show me the learning.”

The Problem with Legacy

The problem is not, as one may expect, that legacy systems are only optimized for a single channel in an omnichannel world. According to Adjaoute, these systems aren’t doing the right job even for a single channel. He said that’s because legacy fraud tools are built on past observation, protecting against fraud attacks that have already happened and not anticipating what could come next.

“Rules in legacy systems are outdated the same day you put them on the system,” Adjaoute said. “It’s like saying, ‘I have an antivirus system, but I can’t update it.’”

It is true, however, that the lack of omnichannel visibility can be an issue for some legacy systems. When solutions are focused on only one channel, or cover multiple channels but lack communication from one to the next, it creates an easy in for fraudsters, who are always seeking the weakest link.

Adjaoute said legacy’s other limitation is that, whatever rules it has been fed based on past scenarios, it applies them equally to all customers and transactions across the board. He surmised that’s the reason why merchants are losing $100 billion per year in false declines — and who knows how many customers are falling by the wayside in the process?

The Next Generation of AI

There are three things AI must be if it is going to succeed in a cross-channel environment, Adjaoute said.

First, it must be adaptive. Fraud changes every day, so a one-and-done type of solution will not serve a merchant very well for very long. Just like those legacy tools, in the face of millions of new malware being created every day, a one-and-done solution will soon become obsolete.

Second, it must be self-learning. It is impossible, said Adjaoute, to teach an AI everything it needs to know before launching it. Again, by the time those training the AI know what rules to feed it, the attack has already happened. They are reacting. By contrast, self-learning AI extrapolates. It is proactive. This is what merchants need, he explained, to stop fraud before it happens.

Finally, AI must be personalized. Every customer, merchant and device is different, so what good is a fraud tool that treats them all the same? Consumer behavior profiles are becoming more thorough and complex, said Adjaoute. Soon, AI may be able to identify when a “cardholder” is not acting like himself and flag his activity as suspicious. That occurs across channels, so it factors in activities like issuing checks, visiting an ATM or taking out loans in addition to commerce and eCommerce activity.

In fact, that’s a concept Adjaoute said occurred to Brighterion around 2014, something the tech company has since been working to bring to fruition. Its new iDetect product is a dedicated tool for data breach defense and mitigation. It identifies cards that are acting abnormally, then shares that information across the system and traces it back to its source.

Like a home security system that allows homeowners to know when a criminal has broken in, iDetect alerts issuers if Brighterion detects abnormal behavior so that the source of the breach can be traced — all within 24 hours, Adjaoute said.

By way of example, he explained, such next-generation AI could have substantially reduced the harm done by a massive data breach like the one at Equifax.