All around the world, customs agencies and commercial suppliers are examining how artificial intelligence (AI) can be leveraged to assist, or replace, human labor. However, what is often missing or misunderstood in these conversations is the recognition that the decision to use AI is not a binary, yes-or-no choice. Rather, the most important consideration is how much autonomy should the AI be permitted to have.
How have other industries tackled AI and autonomy? And what lessons can be learned to help set customs industry-specific definitions for levels of autonomy?
Keep Autonomous Calm …
You are most likely familiar with media coverage regarding the development and deployment of autonomous vehicles. Around the world today, there are already fully autonomous wheeled robots roaming warehouse aisles moving supplies from one area to another, cleaning floors in homes and offices, and even delivering food from the kitchen to the table in restaurants.
In each of these examples, the servers, sensors, and onboard logic are sufficiently complex, fully integrated, and rigorously tested to allow us to trust these devices to supplement or even replace human labor. These devices autonomously navigate their routes while adjusting to both fixed and moving objects in their path in real-time. It is perfectly understandable then, as a gut reaction, to perceive them as threats to future employment prospects.
However, upon further reflection, we should instead applaud these applications for their utility in performing tasks that are too dull, dirty, or dangerous to be desirable. This frees us humans to be employed in more noble, intellectually satisfying pursuits such as product design, service improvement, and customer care.
We can unarguably call these robots fully autonomous in the sense that they require zero human interaction, other than perhaps plugging in a power source at the end of the day to recharge – although that too is becoming autonomous.
You might be surprised to learn that Unimation, the world’s first robotics company, sold its first factory robot to General Motors in 1963, more than sixty years ago. By the early 2000s, iRobot was selling the Roomba vacuum cleaner for home use, which, along with similar products from competitors have since become nearly ubiquitous in residential settings.
More recently, global attention has been drawn to efforts by Tesla, Waymo, Cruise, BYD, and others to automate passenger cars and light trucks on public roads in an effort to create what some have termed “robotaxis.” As a result of these efforts, SAE International, the standards development organization for the automotive industry, in 2014 released J3016 Taxonomy and Definitions for Terms Related to Driving Automation Systems for On-Road Motor Vehicles. Quite the mouthful title for sure, but also a seminal document that allows all stakeholders – manufacturers, regulators, and consumers – to have a common frame of reference when designing, developing, procuring, and deploying such vehicles.
… and Mind the Autonomous Gap
As a starting point then, we propose a similar, six-stage continuum as follows:
Level 0 – No Automation
Level 0 equates to all scanning and screening systems – and related screening data – being manually processed by a human analyst who is 100 percent in the decision loop. An operational example would include an analyst who views an X-ray image on screen without any system-generated markings, reads a paper or digital copy of the manifest, and manually inspects the cargo.
Level 1 – Data Interpretation Assistance
At Level 1, data from one or more scanner subsystems automatically evaluates without cross correlation. The human analyst makes the adjudication with AI assistance in the interpretation of individual data sources.
In practice, automated optical character recognition would read a vehicle license plate at a border crossing and compare it against a database of pre-registered vehicles. The analyst provides confirmation of the match for 100 percent of the vehicles.
Level 2 – Adjudication Assistance
Level 2 would encompass data from one or more scanner subsystems being automatically evaluated with cross correlation. The human analyst makes the adjudication with AI assistance in the adjudication decision.
In the real world, an X-ray system algorithm would be working to detect anomalous signals compared to the digital manifest description, or a license plate reader would be matching against known registered operators. From there, the automated system provides recommendations to analysts for clearance to proceed (or secondary screening). An analyst would monitor the entire workflow process and make the final decision.
Level 3 – Partial Adjudication Automation
For Level 3, all scanner data is automatically analyzed with cross correlation – the same as Level 2 but at Level 3 an automated system clears low-risk conveyance without any analyst needing to intervene. An analyst can override the automated system at any time. Medium- and high-risk conveyance would require specific analyst intervention (to let pass or to send to a secondary inspection).
An operational example of Level 3 could include trucks with containers that are declared as empty being screened to verify emptiness, such that AI makes the decision if the manifest, license plate, operator’s biometrics, and X-ray image are all properly documented and aligned. One analyst can then monitor multiple lanes and rarely intervene in the adjudication process unless there’s an anomaly.
Level 4 – High Adjudication Automation
At Level 4, there is a full correlation of both scanning and other data to make fully autonomous decisions in specifically restricted conditions. The automated system adjudicates 100 percent of the scans, and analyst interventions would be required only for anomalous conditions above a certain threshold.
In a real-world scenario, primary inspection lanes would be entirely unmanned, and the decision to allow the passing or diversion to secondary screening of a conveyance is made autonomously with only occasional monitoring by an analyst during heightened states of alert triggered by specific threat indications.
Level 5 – Full Adjudication Automation
At Level 5, algorithms would fully correlate all data and autonomous systems without any human oversight under all conditions. Analysts would only monitor performance and trend data over time, either locally or remotely.
This would entail the establishment of a fully automated border crossing staffed by mostly maintenance and repair personnel as required. Inspection analysts would be on-call to assist shippers and operators if there are compliance issues.
He Said, She Said, We All Said
As you have probably determined by now, neither the technology nor the resources for Level 5 are on the near-term horizon. However, setting a common lexicon now for what that and all the other levels might look like helps all customs-related parties better communicate their challenges and proposed solutions. Additionally, standardized language helps leaders in both the public and private sectors focus on first principles when determining how best to support adoption roadmaps with appropriate financial and intellectual resources.
Learn more about how customs and border workforces are making screenings more efficient while reducing the cognitive load on personnel, click here.
The author, Jeffrey Goldfinger is Director of Capture at S2 Global and his co-author, Dennis Dujmic, Ph.D., is a data scientist and the Vice President of Algorithm Research at S2 Global.