(This is Part II of a 4-part series on the intersection of ethics and technology, specifically AI, in a practical sense. Part I can be found here)
Machine Learning (ML) complements and empowers workers to focus on their areas of experience and expertise while automating their mundane work, making them more effective and efficient at their job. The new hybrid form of augmented intelligence will provide an unprecedented ability to analyze and process large amounts of data, besting current human systems as well as strictly machine based artificial intelligence. Augmented intelligence relies on human expertise to make final decision, it depends on, and empowers the human in the decision loop. Artificial intelligence instead is a wholly mechanical process that is wholly dependent on the input data and makes decisions and recommendations based on logic derived solely from the data sources that trained the model.
Many ML systems output a ranked list of outcomes ordered by probability. These can be anything from the likelihood of a borrower to default to the chances that a student will drop out of school to whether a social media user is spreading jihadist propaganda. ML’s strength is at reducing a universe of thousands or millions to a segment that is manageable by humans who understand the intimate details of the problem and the effectiveness of the potential interventions. Augmented intelligence, by eliminating tedious data discovery work, allows humans to apply their unique knowledge and skill set more often and in places where it can be the most effective.
Truly achieving augmented intelligence involves fully incorporating ML with human expertise and skills. Organizations must identify problems, determine that ML is the most effective tool and then train the ML and the humans using it to effectively interpret and use the results. Proper training and education of subject matter experts in the basics of ML and the undergirding statistics allows for a more powerful ML system and more effective use of the results. For example, a counterintelligence professional working on identifying propaganda targeting the 2020 presidential campaign can use historical examples from the 2016 campaign to help train a system to identify social media content from state actors. The system trained with this sorts through, in real time, prominent social media network to identify similar content.
We at Omelas use a similar methodology to help tune our models to identify important content and the characteristics of actors relevant to security. The end user then analyzes the filtered content based on his/her expertise. With a basic understand of why the model flagged the content (language, grammatical errors, certain types of rhetoric etc.) the same expert could identify potential state-sponsored covert influence campaigns and use their own expertise to draw larger conclusion about intent and potential remedies. Machine learning systems are inherently dependent, and limited by, existing data they has already seen. The end user can fold in recent news, data from other sources and detect small changes or new approaches that a machine might identify but not flag as of primary importance. Omelas helps train end users to interpret and use the results that we return based on our analysis.
The augmented intelligence combination of the ML model and the expert requires both components for optimal performance. The AI needs an expert training it with appropriate data and assistance on edge cases while the expert will need to understand some of the statistical assumptions that AI is making in its final analysis. The dual process of training the model and training the expert is not overly complex, but it is essential. Allowing subject matter experts to help train the model drastically improves the final results. With a small brush up on introductory college-level statistics the expert can process the output of an ML system and make an authoritative assessment of potential threats and nascent information campaigns with more, and better data, than they could ever hope to process on their own.
The true power of machine learning is most effectively unlocked with an broad commitment to designing a human process to create a system that offloads simple, data processing tasks that used to be done by an expert, to a machine. We have seen this first hand at Omelas, where augmented intelligence has led to the discovery of otherwise-missed security risks and adversarial information operation campaigns. An organization that commits to full augmented intelligence through identifying problems and engaging its experts will reap massive benefits in faster and more effective solutions to its most vexing problems.
