BITES THE DUST
Panel members in U.K. chafed at lack of information access, authority and independence
Google disbands another panel overlooking its AI research,
Google is disbanding a panel here to review its artificial-intelligence work in health care, people familiar with the matter say, as disagreements about its effectiveness dogged one of the tech industry’s highest-profile efforts to govern itself.
The Alphabet Inc. unit is struggling with how best to set guidelines for its sometimes-sensitive work in AI—the ability for computers to replicate tasks that only humans could do in the past.
It also highlights the challenges Silicon Valley faces in setting up self-governance systems as governments around the world scrutinize issues ranging from privacy and consent to the growing influence of social media and screen addiction among children.
AI has recently become a target in that stepped-up push for oversight as some sensitive decision-making—including employee recruitment, healthcare diagnoses and law-enforcement profiling—is increasingly being outsourced to algorithms. The European Commission is proposing a set of AI ethical guidelines and researchers have urged companies to adopt similar rules. But industry efforts to conduct such oversight in-house have been mixed.
Earlier this month, for instance, Google unveiled a highprofile, global, independent ethics council to guide it on the responsible development of all of its AI-related research and products. A week later, it disbanded the council following an outpouring of protests about the panel’s makeup.
In a post earlier this month, Kent Walker, Google’s senior vice president for global affairs, said the company “was going back to the drawing board.”
Months earlier, in late 2018, Google began to wind down another independent panel set up to do the same thing—this time for a division in the U.K. doing AI work related to health care. At the time, Google said it was rethinking that board because of a reorganization in its healthcare-focused businesses.
But the move also came amid disagreements between panel members and DeepMind, Google’s U.K.-based AI research unit, according to people familiar with the matter. Those differences centered on the review panel’s ability to access information about research and products, the binding power of their recommendations and the amount of independence that DeepMind could maintain from Google, according to these people. A spokeswoman for DeepMind’s health-care unit in the U.K. declined to comment specifically about the board’s deliberations. After the reorganization, the company found that the board, called the Independent Review Panel, was “unlikely to be the right structure in the future.”
Google bought DeepMind in 2014, promising it a degree of autonomy to pursue its research on artificial intelligence. In 2016, DeepMind set up a special unit called DeepMind Health to focus on health-carerelated opportunities. At the same time, DeepMind cofounder Mustafa Suleyman unveiled a board of nine veterans of government and industry, drawn from the arts, sciences and technology sectors, to meet once a quarter and scrutinize its work with the U.K.’s publicly funded health service. Among its tasks, the group had to produce a public annual report.
Two months after its founding, DeepMind Health admitted missteps in accessing records of 1.6 million patients in the U.K. without proper disclosure from the partnering hospital to patients. The board was praised for going on to publicly raise concerns in its annual report about the lack of clarity around that data-sharing agreement, and for recommending that DeepMind Health publish all future contracts with public health providers, a recommendation the division complied with.
But internally, some board members chafed at their inability to review the full extent of the group’s AI research and the unit’s strategic plans, according to the people familiar with the matter. Members of the board weren’t asked to sign nondisclosure agreements about the information they received. Some directors felt that limited the amount of information the company shared with them and thus the board’s effectiveness, according to one person.
Not all board members felt cut out. “We were told a huge amount of stuff and trusted to be responsible with it,” said Julian Hubbert, a former British lawmaker who was on the review board. Tensions increased when DeepMind Health told the independent board that Google was taking a more hands-on approach to running the unit, according to these people. In November, DeepMind Health announced that one of its bestknown projects, an app called Streams that helps hospital doctors in the U.K. detect kidney disease, would be run from Google’s Mountain View, Calif., headquarters. That transition is still in progress.
Google said it would build the app into an “AI-powered assistant for nurses and doctors everywhere.” That caused concern in public health and privacy circles because of previous assurances from Google and DeepMind that the two wouldn’t share health records.
DeepMind Health is due to be renamed Google Health. Alphabet also owns health-tracking platform Google Fit and Verily, a life-sciences research unit.
Inside the review board, many directors felt blindsided, according to people familiar with the matter. Some directors complained they could have played a helpful role in explaining the change of control of the Streams app to the public if given earlier insight.
The review panel still plans to publish a final “lessons learned” report, according to a person familiar with the matter, which will make recommendations about how better to set up such boards in the future.