How to prepare your company for incidents with artificial intelligence
If there is one constant in the world of technology, it is that the more a given technology is adopted, the more it fails.
Because responding to AI incidents is different from traditional response to other situations in several ways, efforts to resolve problems with artificial intelligence require their own policies and procedures to guide companies and personnel involved in responding to incidents.
Also read: This CEO fired almost 80% of his team because they didn’t adopt AI quickly
AI incident policies should address the following points:
— Create a definition of AI.
— Identify the most relevant damages.
— Designate those responsible for responding to incidents.
— Develop a short-term containment plan.
What to do after identifying an AI incident: containment, eradication and recovery
Once incidents are identified, the next step is to execute a longer-term containment strategy to prevent the damage from spreading further. Here are the critical questions this assessment must answer:
— Who is being harmed?
— What are the options for modifying the behavior of the AI system?
—What is causing the damage?
— Can existing damage be addressed or corrected in any way?
After the incident occurs, it is important to not only understand who was harmed, but what companies can do about it. A common type of incident arises when preferential services, such as product discounts, are only offered to specific demographic groups.
Once containment plans have been put in place and executed, the next step is to try to completely eliminate the cause of the incident.
Some AI systems may be amenable to eradication efforts, but other systems may only allow partial eradication, or the source of the incident may not be removable at all.
At a broader level, there are three main ways to address or fix problematic model behavior, all of which have long been used to debug machine learning models:
— Preprocessing: concerns the actions that can be taken before the model absorbs or is trained with input data. In some cases, unrepresentative training data can be the source of problematic model behavior. In this case, the solution is to retrain the model with more representative training data.
— Internal processing: This type of correction involves changing the weights themselves or the model architecture. Sometimes these updates are relatively simple, but in most cases they require a significant amount of time-consuming development. I’ve rarely seen this approach work in practice, mainly because it often involves building an entirely new model.
— Post-processing: This is the most straightforward option of all and has been widely used to solve AI problems for decades. It involves changing the behavior of the model after the system has already made its predictions. Output filters, for example, can simply prevent certain behaviors or prevent specific predictions from being generated. Typically this takes the form of rules that can be added to the model.
Lessons learned
After response activities are completed, it is critical that companies conduct a post-incident review to learn and improve from each incident.
This involves taking a step back and evaluating the successes and failures of how the incident was handled.
Preventing and responding to risks must be an ongoing exercise.
