LLMs fail at conditional logic

 

Anti Money Laundering Using Active Structure

 

The term "money laundering" refers to the processes by which the illegal origins of funds are concealed, usually through interactions with banks and businesses. An estimated 2.7% of the world's global GDP is laundered annually, equating to huge government losses on a global scale. These are the funds that allow criminal and terrorist organizations to continue their operations financially unhindered. The massive problem posed by money laundering has resulted in a huge market for anti-money laundering software. Despite increasing investments in software development, an estimated 90% of money laundering still goes undetected. So what's been going wrong? Why has the existing software been performing so poorly?

 

Let's explore some of the approaches being used. One can take an algorithmic approach and write a program to deal with the issue. Such a program can even be quite complex and take into account a variety of contingencies, but there are limitations to this approach. First off, the program can only be written to address known problems, meaning its horizons will be limited to only those issues its creator has thought up. Secondly, the utility of the program ends as soon as the situation changes. In the case of money laundering, once criminals begin to figure out what the parameters of your algorithm are, they can find a way around it and suddenly your program is worthless.

 

So what about machine learning approaches? These have some advantages over the algorithmic approach in that the machine can, given the right training data set and parameter specifications, potentially come to address problems that the creator did not even recognize existed. The problem with this approach is that the machine's competence will always be constrained by the quality of the training data set it has been given. And no matter how good your training data set is, the machine will never be able to anticipate entirely new problems from it. At best, it can extend the logic of the patterns it has extracted to new problems that arise. This may work at first¾to a point. But once the rules change too drastically, the machine is going to find itself completely helpless - it can't make new connections.

 

So why all the fuss over anticipating future problems? When it comes to money laundering, we're not dealing with a stable problem. We're competing against human intelligences, with all of the mental flexibility they represent. Every time an optimal solution is found, it won't be long before the competition finds a new evasion tactic and the process starts all over again. That is the nature of competing with truly flexible intelligences.

 

Where does that leave us? The situation I've laid out points to one obvious solution: if we hope to achieve any consistent level of automation in the detection of money laundering, the machine behind that detection needs to be able to recognize when the situation has changed, plot a new course to counter it, and modify itself to implement the new strategy. The alternative approaches will always be playing catch-up.

 

What we're proposing is a dynamic machine that is not limited by quantifiable data being fed to it. Rather than working with 1s and 0s, the medium it works with will be the English language itself. Because of this, it will have the ability to seek out its own data and find solutions to problems it identifies without supervision. It can actively search out problems and self-modify its structure to better counter them. Once properly implemented, the machine we're proposing would be able to anticipate and address entirely new problems.

 

And whatever parameters we start with, yes, criminals can still find a way to avoid it. Say we begin with a simple tactic like aggregating small transactions over a period of time to detect structuring: criminals will soon learn what that time period is and counter by spreading their transactions over a greater span of time. There's no getting around that. Trying to catch money launderers is always going to be a back-and-forth arms race. Our salvation here, and why our approach can be better than that of others, is that our machine will modify itself as the situation changes. If we get this right, every time criminals find a way around a new rule our machine will adapt to the new situation and set up another wall to block them. In a setting where the rules keep changing, only an adaptive strategy – an intelligent strategy not based on data - can keep the advantage for any amount of time, and that's what Active Structure has to offer.