In the era of artificial intelligence, how do we align AI systems to human values? Whose values do we align them to? Moreover, how do we avoid AI models turning out downright misanthropic and suggesting truly destructive ideas?
Courtesy of Patrick Beaudouin
Industry leaders and academics convened at Stanford Graduate School of Business (GSB) in early April to discuss one of the most important matters in using AI: Teaching it to be helpful and human-centered. Speakers at the Bridging Humans and Machines: Advancing Alignment in AI conference agreed on the technology’s power and potential, but because the risk of unintended consequences is so great, innovative technologists and ethicists are actively working to build AI that aligns with human values.
More than 100 industry professionals, academics, and students attended the gathering, presented by the GSB’s Business, Government & Society Initiative. GSB professors Susan Athey, Mohsen Bayati, Andy Hall, and Gabriel Weintraub organized the event.
Which Values and Whose?
Courtesy of Patrick Beaudouin
Speakers generally agreed that well-aligned AI reflects human values and minds, though deciding whose values and minds should underlie AI is complicated. “We share fewer universal human values than we’d like,” said the GSB’s Andy Hall. It can be difficult to “align to what people want because different people want different things.”
Sometimes the task of alignment is relatively simple. A company might use AI to manage internal queries by employees and respond based on the firm’s archive of emails, chats, and business records. In that case, it’s the company’s values and code of conduct, including its compliance and governance requirements, that clearly preside, said one speaker.
In other instances, it’s much harder. A company building foundational AI models has to build it for use across countries, languages, and cultural contexts. Who should decide which values the model adopts and which it refuses? In a morning panel, participants discussed how the tools of democracy might help to resolve these issues, allowing companies to avoid becoming the “arbiters of truth” and building fairer, more decentralized processes in which users, rather than elites, determine the values underpinning the AI models they will rely on.
Polling the Public
Companies that offer AI-driven platforms to the public might benefit from considering a wide range of public opinions when designing their systems. Meta, for one, in 2023 and 2024 collaborated with Stanford’s Deliberative Democracy Lab, which researches democracy and public opinion, to examine the public’s views on the use of AI. Over the two-year study, the Democracy Lab gathered over 2,400 people from nine countries, including the U.S., asking them questions such as, “Should users be allowed to use AI chatbots for romantic relationships?” More participants, after deliberation, rejected that idea than approved of it.
Deliberative Polling results “serve as a data point to form policy and products,” said Alice Siu, the Lab’s associate director. “It’s useful for regulators to know what the public is thinking and what they think technology should look like.”
Reasoning Required
Courtesy of Patrick Beaudouin
In addition to reflecting a range of public opinions, systems are more aligned when they incorporate more contextual information. Why a system does something can be as important as what it does. While recommendation systems are helpful, they aren’t as precise as they could be because they often lack the reasoning ability to determine why they’re suggesting an idea.
AI developers are trying to incorporate reasoning, particularly in important areas such as self-driving cars for safety measures: For a car that was trained in mostly mild weather, an inability to determine “why” makes it less able to extrapolate to driving in extreme weather. Lacking reasoning ability also makes AI less likely to recognize situations when it can’t produce a legitimate response and needs to stop and tell the user so, said an industry participant.
Helpful to Humans
Courtesy of Patrick Beaudouin
How we view and use AI is based on whether we believe it is aligned. Some AI already appears sufficiently aligned for use in medicine, where chatbots interact with patients in areas including prenatal care and mental health, said an industry speaker. AI also helps save physicians’ time by highlighting information from a patient’s medical record, which can stretch for thousands of pages, and has the potential to simulate cell behavior as part of drug development, added other industry participants.
However, because people expect face-to-face interaction with health-care providers, AI shouldn’t be allowed to replace personal communication and should know when a human should take over completely, opined another participant. Ideally, the best AI would join the best human competencies and the combined result would be superior to both.
