The truth about AI and due diligence

There has been a lot of discussion lately about the possible uses of artificial intelligence in anti-bribery compliance. Most commentators recognize that AI (and its more unassuming cousin, machine learning) won’t do away with current practices altogether. But while the significance of these technologies is recognized, their practical implementation often receives only cursory attention.

AI is full of potential to improve the efficiency and effectiveness of due diligence programs over the coming years. What follows are some specific suggestions of mature technologies that can enhance a team’s capabilities today.

To better grasp this potential, we need to understand the technology’s limits. Artificially intelligent systems like DeepMind’s AlphaGo and AlphaZero have already outclassed humans at complex games like Chess and Go. But playing games is all they can do. In that sense, they can be called narrow. An AI with wider applications that can engage the full complexity of human intelligence remains beyond our reach.

Anti-bribery compliance isn’t narrow. Although we cannot yet use computers to replicate the entire scope of a compliance team’s work, certain areas of AI/machine learning research have matured to the point of offering substantial assistance.

The first of these is optical character recognition (OCR), a set of technologies developed in connection with computer vision research. OCR can convert a picture of a document into machine-encoded, searchable text. Enterprise implementations of OCR are now commonplace, and readers likely already have this technology available in their offices.

The power of OCR becomes evident in the next area of research: natural language processing (NLP). This term encompasses a wide range of capabilities, including document comparison, machine translation and named-entity recognition. By automating the process of extracting meaningful information from a mass of paper documents, OCR and NLP can help move past time-intensive manual review to focus on more substantive issues.

Data gleaned from NLP can, in combination with other information sources, open up a new avenue for analysis in the form of quantitative risk assessment. In general, this process involves using data about an individual, company or organization to assess its specific risk level.

For instance, a lender might use a person’s credit score, loan history and salary information to assess the risk of default — whether in binary fashion using machine learning classification techniques (e.g. “will pay back loan” vs. “will not pay back loan”) or along a numeric scale using regression-based methods (e.g. “75 percent probability of paying back loan”).

In principle, machine learning algorithms can be applied to information about third-party agents, intermediaries and suppliers to assess the associated degree of corruption risk. The outcomes of these automated assessments could then be used alongside human review to augment a compliance program with a more macroscopic viewpoint.

The tools for conducting this sort of automated analysis are well-developed, but their application to anti-corruption due diligence is in its infancy at best. Most of the products in this area that are currently marketed as AI solutions are actually some combination of OCR and NLP, adding value and aiding review but not independently generating intelligent insight into the applicable risks.

There are significant hurdles to overcome before machine learning can yield this sort of insight about bribery risk. Limited data is one. Machine learning typically requires a sizeable set of examples of the thing being modeled for its algorithms to train upon. But while instances of bribery may or may not be as common as loan defaults, they usually do not become a matter of record, making it hard for a machine to identify the characteristics they may share.

Another issue regards interpretability and fairness. Board game AI has become so powerful, it’s a challenge to even figure out how it wins. That’s fine for games, but what about when it affects human lives and livelihoods?

Criminal sentencing is one area where the use of AI technology has come under serious criticism for its possible inclusion of systemic biases, hidden under a veneer of algorithmic objectivity. Any application of machine learning to anti-bribery risk assessment will need to be mindful of similar possibilities.

We should let this sort of danger serve as a reminder that while machine learning in its current state is well-suited for discrete problems that can be narrowly defined, it is not a complete solution to the broad scope of societal ills. Nevertheless, by making large quantities of information more manageable and intelligible, it can already help compliance teams better engage their own critical faculties in a more informed decision-making process.

_____

M. Merritt Smith, pictured above, is a Research Associate at TRACE International, where he specializes in the application of advanced statistical and machine learning techniques to corruption-related problems.