What Companies Can Learn From the U.S. Government’s Use of Artificial Intelligence

November 5, 2019, 3:16 PM UTC

This is the web version of Eye on A.I., Fortune’s weekly newsletter covering artificial intelligence and business. To get it delivered weekly to your in-box, sign up here.

The federal government, like the private sector, is turning to artificial intelligence to improve operations and reduce busywork plaguing its sometimes overworked staff.

And like most businesses, federal agencies face obstacles along the way, including a lack of software tools and data infrastructure required for cutting-edge data crunching.

At an A.I. conference last week at Stanford University, Stanford law professor David Engstrom discussed the challenges facing the federal government’s foray into machine learning. In many ways, the challenges mirror those facing corporations, like having employees how know how to operate sophisticated machine-learning software. 

Engstrom shared some preliminary findings from the Stanford Policy Lab, which has analyzed technology use by the federal government. The team will eventually present the analysis to the Administrative Conference of the United States, a federal agency intended to improve government processes, to create guidelines for agencies using machine learning. 

For the project, Engstrom said members of the policy lab analyzed about 150 Federal departments and agencies to find how they are using machine learning. Ultimately, the researchers identified 171 different uses.

Two of the leading agencies were the Securities and Exchange Commission and the Social Security Administration, he said. 

The SEC, for instance, currently uses machine learning to help identify scammers who may engage in insider trading. The SSA, meanwhile, uses machine learning to catch possible errors in draft decisions that spell out who receives payouts on claims.  

The SEC became savvy in machine learning because it had to keep up with the rapidly changing financial sector. Engstrom described the commission as having a “strong innovation culture” that is “way ahead of most other agencies in terms of the development of these tools.” 

The SSA developed its machine-learning chops due to the “entrepreneurial efforts of a few employees” and a recently retired judge, Gerald Ray, who helped spearhead data-crunching projects, Engstrom said. The judge hired lawyers who had computer-programming skills and then let them work on data-crunching projects.

Businesses can learn from these two agencies and their machine-learning successes. For instance, the SEC’s tech-focused culture requires its technical and administrative staff to routinely “gather and compare notes,” ensuring that everyone agrees on which machine learning projects to pursue, Engstrom said.

Companies starting out with machine learning should identify a strong entrepreneurial tech advocate who can help the company build the necessary team to focus on data projects, like the judge who helped the SSA. Engstrom referred to the judge as a “one-man show” who helped the agency move forward despite some staff who were less enthusiastic about technology. 

In general, the findings contradict conventional wisdom that the government is slower to adopt technology than businesses. In fact, some agencies are doing just fine.

Jonathan Vanian 


Not sure this will fly. The Australian government is considering whether to use facial recognition technology to verify the ages of people who watch pornography online, The New York Times reported. Although the proposal is far from becoming law anytime soon, its mere suggestion has startled privacy advocates and some lawmakers including Australian senator Rex Patrick, who said, “I think people should be very concerned about any government department that’s seeking to store this kind of information.”

Microsoft and Amazon both agree. Microsoft and Amazon oppose an effort by Germany and France to develop a European-specific cloud-computing project intended to limit the influence of U.S. cloud giants in Europe, the Wall Street Journal reported. European government officials told the Journal that they hope the cloud project will “help European companies in industries such as manufacturing and health care in their efforts to develop algorithms and artificial intelligence without relying on foreign technology providers.”

Tell us about your facial-recognition projects. The American Civil Liberties Union sued the Department of Justice, the Drug Enforcement Administration, and the FBI to get them to reveal information about their projects that use facial-recognition technology. The ACLU said in the court filings that it wants to know “how face recognition and other biometric identification technologies are currently being used by the government, and what, if any, safeguards are currently in place to prevent their abuse and protect core constitutional rights.”

There are no right-or-wrong answers. Amazon appears to be poorly vetting user-submitted answers from people who participate in Alexa Answers, a crowdsourced project in which people can submit responses to questions that the Alexa digital assistant doesn’t know. Tech publication VentureBeat analyzed some of the crowd-sourced answers and found several problems, including inaccurate responses, sponsored content, and trolling by Internet pranksters.   


White House CTO Michael Kratsios commented on the competition between China and the U.S. in artificial intelligence during a Stanford University A.I. conference last week. He mentioned computer vision technology as an area in which China is increasingly challenging the U.S. “For us that’s one place where we probably see they're re moving ahead on a little further," Kratsios said.


Nyla Technology Solutions, a technology company that serves the federal government, hired Stephanie Beben to be its chief data scientist. Beben was previously a chief data scientist for Booz Allen Hamilton’s strategic innovation and national security group.

The U.S. Navy hired Tom Sasala to be its chief data officer and Jane Rathbun to be chief technology officer, government-focused trade publication FedScoop recently reported. The Navy’s CIO Aaron Weis, who joined the military branch in September after serving as a senior advisor to the Department of Defense’s Office of the CIO, picked the new officials as part of a broader restructuring, the report said.



Everyone is talking about language these days. Researchers from Google, Facebook, and Salesforce all recently published papers detailing progress in natural language processing (NLP), in which computers learn to understand language. The papers underscore a current boom in NLP, driven in part by Google’s BERT (Bidirectional Encoder Representations from Transformers) A.I. language model, which other researchers have either improved upon or used to better their own A.I. language software.

Facebook’s research, for instance, details the social networking giant’s BART technology, and its ability to translate languages and improve how computers automatically generate answers to questions. The paper also describes progress in teaching computers to more accurately summarize news articles.

Salesforce’s research is focused on using A.I. to more precisely summarize articles. The company's researchers, who used Google’s BERT technology, said they hope their paper “will encourage more research efforts in the important task of verifying and improving the factual consistency of abstractive summarization models.”

Google’s research paper examines how NLP can benefit from the A.I. technique of transfer learning, which could help researchers more easily train A.I. systems. The idea behind transfer learning is that pre-trained A.I. systems will have an easier time learning from other datasets that are similar to the ones used during the original data-training process.


Booz Allen Creates ‘App Store’ for A.I.–By Jeremy Kahn

3 Key Takeaways from Google’s Fitbit Acquisition–By Sy Mukherjee

A Better Picture of Your Muscles, Thanks to ‘Deep Learning’ A.I.– By Andrew Nusca

HPE’s CEO: China ‘Needs the West to Continue to Teach Them’ –By David Z. Morris

The U.S. Interior Department Has Grounded Its Fleet of 800 Drones, Fearing Chinese Surveillance– By Eamon Barrett


It isn’t just fun and games. Fortune’s Jeremy Kahn reported on a milestone research paper by Google’s DeepMind A.I. unit that described the AlphaStar A.I. software that can beat many of the best human players in the computer strategy game Starcraft II. The paper, which was published in the journal Nature, is noteworthy for a number of reasons, one of which being that it shows A.I. software can handle multiple tasks at once while planning for the long term. As A.I. expert Dan Klien of University of California at Berkeley (Klein was not part of the DeepMind team who built AlphaStar) told Fortune about the paper’s significance, "Starcraft, like life, is complex in many ways all at once." It will still take years for this type of advanced A.I. software to move out of the research lab and into business, but the paper shows that A.I. continues to improve. 

Read More

CEO DailyCFO DailyBroadsheetData SheetTerm Sheet