In Stanley Kubrick's cinema masterpiece "2001: A Space Odyssey, an advanced onboard AI computer named HAL provides a spaceship crew with critical updates and essential information, ultimately shaping the mission's trajectory as it journeys to Jupiter.
HAL is no ordinary computer. Indeed, the 1968 film introduces us to a fully sentient AI system that provides real-time updates to the crew while monitoring and maintaining the spacecraft's systems, reporting system anomalies and possible malfunctions. HAL informs the crew in one scene that a radio communication device will soon fail. When asked for more information, HAL responds with complete clarity - and authority.
Why Am I Seeing This? Explain Yourself, Please
Today's AI systems are far less sentient than HAL. But today, AI is powerful and capable enough to monitor and assess enterprise activities to help detect financial fraud.
For example, employees tasked with detecting fraud increasingly rely on AI and machine learning to detect anomalies in email correspondence.
But the fraud alert in and of itself isn’t enough. When a reviewer is presented with recommendations, she needs to understand how the AI came to that conclusion. In short, she needs to have complete assurance that the AI is accurate and correct.
"The next revolution in machine learning and AI will be about explainability," said Chris Stapenhurst, Senior Principal Product Manager at Veritas. "Reviewers need to understand why and how the AI engine arrives at its decisions," he said.
When a reviewer faces a list of red flags in dozens of emails the AI has flagged, the reviewer needs to know why. It needs to provide transparency into its reasons, Stapenhurst explained.
The AI detects certain words in the subject line or in the content of the emails that lead towards relevancy. But, more than just keywords, it also factors in metadata attributes such as message direction and participants as well as receives contributions from elements based upon Natural Language Processing such as sentiment analysis.
"It could be designed with a classification engine that provides out of the box expert pre-trained models that can identify items like money laundering,” he said.
Stapenhurst continued: “It places a tag which is then factored as a contributing element by a machine learning engine to assist with its predictions supporting human input. This type of system is often considered augmented AI, which is AI, plus human interaction.”
“It learns to predict which items may be relevant for the reviewer with specific tags like suspected money laundering and can even accept submissions of key words and phrases directly by reviewers to augment its learning as part of the item review process.”
Transparency Creates Trust
When Sherlock Holmes famously says, "Elementary my dear Watson," he finally establishes the source of his conclusions that solved a crime. The reader feels satisfied. Watson is relieved of doubt. Trust in Holmes is reinforced once again.
Holmes was an uncanny genius and, by dramatic flourishes, kept the reader in the dark until the end of the story. Not that he was opaque. Far from it. He was the very portrait of transparency but only makes us wait until the end of the story to understand his reasoning.
