“AI is great at taking over tasks, but not jobs.”
Vinod Khosla, VC investor
Here is an analogy I like. Experts expect that in the USA, the first AI doctor to be approved in 2028. This event will reduce the number of doctors, but not doctors as a job. For example, let's say you have a weird skin rash. You consult the AI and might even show some pictures. The AI might tell you to put some cream on your skin and consult 24 hours later. Or, it might refer you to a human doctor for further investigations. Ultimately, we'll need fewer doctors, but not remove the job of a doctor.
As such, AI is a tool for doctors, lawyers, teachers, programmers, auditors ...
AI is loosely based on human brains.
Neural networks represent a fundamental technology for AI. These neural networks are modeled on our biological brain's information processing units and storage mechanisms.
Image of a neural network loosely representing the human brain (source: AuditedAI)
A large number of simple processor elements, called neurons, are connected to a large number of neighboring neurons via so-called synapses. The network collects information and calculates its results by straightforward computational steps.
However, the power of this approach is not in the computation of individual neurons but in the parallel data processing of several million or even billions of individual elements. In our case, we’ve developed a dataset of 7 billion accounting data transactions, probably the world's largest accounting dataset for training AI.
Although the information processing of a single neuron is simple - in the simplest case, a multiplication, addition, and threshold comparison - enormous performance can be achieved by the high interconnectedness of those neurons.
In other words, an artificial neuron is a mathematical function created by a highly simplified abstraction of the biological model.
Teaching an AI to detect accounting fraud
AI is loosely based on human brains. So how do we humans learn? In general, we learn through trial and error. For example, we often make mistakes when we learn to pronounce a new word. With feedback, we get better. It's the same for an AI learning to detect accounting fraud. Technically, this is called “supervised learning.” However, not all AIs learn the same way though.
In accounting fraud, by definition, a fraudster has to invent numbers to hide profits or losses. However, we don't know how a fraudster invents numbers. But that's not important to our AI. We need to understand what legitimate accounting transactions look like.
What do legitimate accounting transactions look like? Fraud experts have used Benford's Law for a long time. According to Benford's Law, legitimate accounting transactions display a particular pattern. Namely, low digits appear more frequently than high digits. For large accounting transaction samples (i.e., over 3,000 transactions at a 100% confidence interval), Benford’s Law manifests universally with accounting data, no matter the currency.
Based on Benford's Law, we taught our AI to detect accounting fraud. We've developed probably the world's largest accounting dataset for AI to train our AI (7 billion transactions). As mentioned before, the approach is "supervised," which means in our case that each transaction had to be manually labeled as fraud "yes" or "no."
Probably the world’s largest accounting dataset for AI (7 billion transactions)
Traditionally, an AI is trained on individual transactions, such as sales numbers and the weight of patients. However, in our case, we trained the AI based on Benford's Law, and thus, the AI had to "see" the data like Benford's Law does: accounting data based on the first digit 10 – 99.
Now we feed the labeled accounting data to the AI. Based on the labels (fraud yes/no), the AI learns to predict the patterns corresponding to accounting fraud.
In the first few iterations, the accuracy starts poor …
… however, the AI learns by doing many iterations and adjusting its prediction.
In other words, the AI learns the multidimensional patterns associated with accounting fraud.
After many iterations …
… the accuracy of the AI improves.
Similarly to how humans acquire knowledge, the AI learns through an iterative process, improving its prediction power with each iteration.
Due to the large accounting dataset (7 billion transactions), the AI learns to detect complicated patterns.
The power of AI lies in its ability to learn complex patterns from large datasets, even hidden from the best traditional algorithms (e.g., traditional Machine Learning or Benford's Law).
June 9, 2023 Franco Arda