The Necessity of Classifying Risks in AI

Deel dit bericht

The rise of big data and the generalized production and consumption of information changed the face of our society but also transformed the awareness about the impact of its use.

AI has been the main beneficiary of this change so far and it is progressively becoming the key benefactor of global innovation by:
• Drastically boosting the intellectual workforce of our society
• Breaking through major bottlenecks in every scientific domain

Now, as for every industry, AI is about to enter the next phase of democratization where it reinforces the promise of being the pillar of tomorrow’s civilization — all while its risks are yet to be understood by everyone.

Making Risk Assessment a Vector of Innovation
In general, statements about risks are driven by fear and it’s fairly easy to understand uneasiness created by the uncertainty of a world where AI will be present in every major aspect of our society, should we embrace it.  

If we were to compare it to previous major digital innovations, such as search engines and  technologies behind internet indexation and the democratization of online resource accessibility, we would certainly agree with Claude Shannon’s statement that says, “Artificial intelligence would be the ultimate version of Google.”

The challenge with the current state of our technology is that Shannon’s belief seems to be an understatement. We can see AI acting as an additional instinct, an alternative to many of our senses and many other cognitive capabilities, or even virtualizing coworkers or business partners.   

The gap in trust between services powered by search engines and the ones powered by AI is probably due to the fact that interacting with online services gives a sense of control from the user perspective. There were many unwritten rules and guidelines that were inherited from the very same service before their digitalization (i.e., being wary of a seller and checking his reputation or cross-validating information from multiple sources). Transposing similar guidelines to AI is definitely more difficult (even when you are not a data professional) due to:
• The critical role of automation in AI implementation and lifecycles and their (very large) scale
• The complexity and the diversity of machine learning models

Before giving access to AI to the general population, we need to build and share a simple way for everyone to identify the weaknesses and side effects, their origin, and their impact for each context — regardless of the business context or the professional background.  

At the moment, many doors remain closed, slowing down innovation in many places. However, these doors could be open with such a framework in the hands of:
• The regulators in governments or in specific domains
• The chain of command in every organization as they would clearly see what’s at stake
• The general population of voters as they would finally know what they are exposed to and what the limits of what is, sometimes, seen as a forceful technological invasion are.

People embracing the changes in society brought by AI need to be aware of the benefits and the risks associated with it. As we reach the era of Everyday AI, consumers will continue to become contributors — so the more people are informed, the better their ability to autonomously make the right decisions. In the upcoming blogs in this series, we'll unpack identifying risks and impacts at various stages such as the source data, AI models and service implementation, and adoption.

Joel Belafa is Director of Engineering at Dataiku.