2015 beneficial-AI open letter

In January 2015, an open letter hosted by the Future of Life Institute called for research that would make increasingly capable AI robust and beneficial to society. It endorsed an interdisciplinary research agenda, extending the question from what AI can do to whether its behavior serves human purposes.^[1]

FLI’s year-end account traces the initiative to its January 2–5 Puerto Rico conference, which brought together AI researchers, industry leaders and experts in economics, law and ethics. The account identifies the letter and research-priorities document as products of that meeting. The initiative dates to January; the October 28, 2015 date on the current letter webpage is a later webpage date.^[2][1]

What the letter endorsed

The letter recommends expanding research beyond making AI more capable to maximizing its societal benefit. Its central formulation is “our AI systems must do what we want them to do.” It connects this task to economics, law and philosophy as well as computer security, formal methods and AI research.^[1]

The current original lists Stuart Russell, Geoffrey Hinton, Yoshua Bengio and Yann LeCun among prominent signatories. Signing endorses the letter’s research recommendation; it does not establish shared forecasts, risk probabilities or support for later letters. The page’s running signature totals are not a count of support at the January launch.^[1]

From an appeal to research questions

The companion research agenda, by Stuart Russell, Daniel Dewey and Max Tegmark, turns the broad appeal into questions about economic effects, law, ethics and technical robustness. Its later Winter 2015 journal version separates four technical tasks: verifying that a system satisfies formal requirements, choosing requirements that produce desirable behavior, securing the system against unauthorized manipulation, and retaining human control after it starts operating (pp.106–109). These are proposed research priorities, rather than safeguards already demonstrated to work.^[3]

That distinction matters for AI Alignment: a system can meet a specification while the specification itself fails to capture the intended outcome. The agenda treats selecting the right requirements as a problem alongside implementing them correctly (p.108). The journal article is a later version of the agenda associated with the January initiative; its Winter 2015 publication does not date the letter’s launch.^[3][2]

Research support and public discussion

FLI reports that Elon Musk announced a $10 million donation for an AI safety research-grants program at the conference’s end. It subsequently announced 37 teams on July 1, 2015, awarded $7 million in a first round funded by Musk and the Open Philanthropy Project. These are the organizer’s reported funding activities, not evidence that the projects solved the research problems.^[2]

The discussion also reached public broadcasting: Science Friday’s April 10, 2015 segment on AI’s future lists Russell, Eric Horvitz and Max Tegmark as guests.^[4] The letter’s central recommendation was a research direction: making desirable behavior and societal benefit explicit subjects of AI research. Its signatures and associated funding document support for that work, rather than proof of safe systems or an implemented regulatory policy.^[1][2]