​SAS recommends government start small with AI and machine learning

The company's director of Global Government Practice has said it's best to start small and build confidence in-house before pushing a data analytics-based project out to the masses and hoping for the best.
Written by Asha Barbaschow, Contributor

Government problems are well-suited to be solved using artificial intelligence and machine learning, SAS director of Global Government Practice Steve Bennett has said, but the resulting project should be well thought through and championed from within before being pushed out to citizens.

Speaking with ZDNet following his visit to Australia where he met with government entities to discuss the role of data and analytics within the public sector, Bennett said there is a challenge in finding where it is best to use emerging technologies, noting that machine learning is good for finding difficult-to-detect patterns, connecting crimes, and determining tax or benefits fraud in a "very complex background of data".

"So from that perspective, the sorts of problems the government struggles with are very well-suited to use these advanced technologies," he said.

"But the challenge whenever you're going to use technology that's going to support a decision that could impact the citizen -- the level of accountability and transparency goes up, as it should with government."

Using the example of shopping online, Bennett said if an ecommerce provider is using a machine learning algorithm to recommend the next shirt a customer buys and it gets that wrong, there are little-to-no repercussions. However, that's not the case if the government flagged somebody as a potential terrorist threat or classes a benefits claim as fraudulent.

"We don't just shrug our shoulders and say, 'oh well' like we would if a shirt was the wrong recommendation," he continued.

"So government has a high level of accountability when it uses these technologies to be able to explain why you're taking a potentially adverse action against the citizen and then if that's going eventually end up in court, you've got to be able to unroll that in a way that stands up against evidentiary scrutiny."

Bennett wasn't too familiar with "robo-debt" when delivering this advice.

At the end of 2016, the Australian Department of Human Services kicked off a data-matching program of work that saw the automatic issuing of debt notices to those in receipt of welfare payments through the country's Centrelink scheme.

The program had automatically compared the income people declared to the Australian Taxation Office (ATO) against income declared to Centrelink, and the debt notice -- along with a 10 percent recovery fee -- was subsequently issued when a disparity in government data was detected.

One large error in the system was that it was incorrectly calculating a recipient's income, basing fortnightly pay on their annual salary rather than taking a cumulative 26-week snapshot of what an individual was paid.

"It's usually a big risk to change a lot of things all at once, particularly in a government agency," Bennett said. "Government, at least western democracies, are built to change very slowly and when you try to make a lot of change in a very big program all at once you introduce a lot of risk."

Instead, his advice is to pick something that can show a "quick win" using the technology.

"You could conduct a pilot or proof of concept ... choose something you can demonstrate quick value to test what you're going to try, rather than applying something all at once without being certain that you know how it's going to work," he explained.

"That does a couple of things: One it tests the technology, lets you kick the tyres a little and say, if we do something small and limited with a lot of oversight and we don't put it into production in front of citizens right away, we kind of run it off to the side to watch what it's going to do and we compare that, build some confidence that we understand what the technology's doing. It also builds confidence in the legislative oversight."

He said it also offers staff within the agency the ability to trust some of the technologies, which then allows the footprint to be expanded with confidence from the ones at the receiving end of citizen concerns.

"[Often], the way government is built to procure things is to do it all once and you write a really big contract that's going to make this big change," he explained. "But often we found that the better successes are incremental. You can show value or impact and you achieve not only the benefits from the technology but you get some of that cultural change as well coming along."

Discussing the similarities between Australian government entities and their global counterparts, Bennett said one common theme is challenges in policing.

"We are drowning in data -- our law enforcement agencies are overwhelmed with data -- and we have to do the best job we can to make sense of that data to keep people safe," he explained.

"Certainly in our case, in [the United States], we could point after the fact to a lot of missed connections of data after the events of September 11 ... and I think this is true in places like the UK and elsewhere."

Prior to joining SAS, Bennett spent 12 years with the US Department of Homeland Security. He said there's a lot less tolerance from the public of information failures, noting they expect that if the government has data that they are able to use it to find things before they happen and keep people safe.

"Everybody is interested in using data to help keep children safe, to identify children at risk, to help police agencies like Victoria Police do a better job with their data. So really the challenges were quite similar and I found that to be true just about everywhere," he said.


Editorial standards