AI 的未來仍由選擇決定
AI’s Future Is Still a Choice
Stanford HAI 秋季會議探討治理、工作與教育方面的選擇,如何讓 AI 的潛力惠及更廣泛公眾。Audrey Tang 倡議讓公眾查閲 AI 決策依據並有實際參與機會;與會者亦指出,企業仍面對整合、準確度及培訓問題,AI 帶來的利益如何分配取決於企業和社會的選擇。
AI’s Future Is Still a Choice
At Stanford HAI’s fall conference, experts explored how decisions about governance, work, and education can turn AI’s promise into broader public benefit.
Shana Lynch
Link copied to clipboard!
November 18, 2026
Shana Lynch
November 18, 2026
AI today is a polarizing topic. People worry it will destroy our climate, shrink the job market, and concentrate wealth and power among too few. Others expect it will cure cancer, improve productivity, and expand scientific discovery.
That dichotomy was the focus of Stanford HAI’s fall conference, “Confronting Our AI Future: Hope, Fear, and the Choices Ahead,” hosted at Stanford University on Oct. 28, 2026.
“Facing these fears matters because AI brings immense potential,” said HAI Denning DirectorJames Landay. The independent, interdisciplinary research at universities can help people meet these concerns with evidence rather than speculation, he said.
The day’s conversations included technologists and educators, philosophers and entrepreneurs, a lawyer, a filmmaker, and a Catholic bishop.
Here are the main takeaways (or check out this one-page summary).
To rewatch the conference, visit the Stanford HAI YouTube channel.

From left, Audrey Tang, Taiwan’s cyber ambassador-at-large, and Rob Reich, McGregor-Girand Professor of Social Ethics of Science and Technology and a Stanford HAI associate director.
Pacing the Frontier in Practical Terms
In the day’s opening keynote, Audrey Tang, Taiwan’s cyber ambassador-at-large and a co-author of the book Plurality, argued that calls to “pace the frontier” are incomplete. Thousands of employees at AI labs have signed letters urging a slowdown, she noted, but those letters don’t say who the “we” is that sets the pace, or where that pace is heading. Tang called for something more precise: that the public see the evidence behind AI decisions and have a real say.
Her framework rests on three components: evidence that anyone can check, a witness that is independent of every lab and every state, and a duty to answer by a deadline. Her model is Taiwan’s 2024 deliberation on deepfake fraud. The government sent 200,000 random text invitations, enlisted 1,700+ volunteers, launched small-group deliberation through a digital platform, and proposed a regulatory package that over 85% supported, leading to passage of a law within months.
She proposed fingerprinting models and placing them in trusted execution environments where outside evaluators can run double-masked tests without leaking either their tests or the model weights. A pilot involving a Gemini model and Singapore’s AI institute is already underway, she said. She also proposed a standing global citizens’ assembly of 1,000 randomly selected people to judge the evidence and direct evaluators.

“I’ve never seen such a gap between the capabilities and what’s actually on the ground.
Erik Brynjolfsson
HAI Senior Fellow
The Future of Work
AI is rapidly changing how some companies work, but most businesses are struggling to turn AI capabilities into real productivity gains, panelists said during a discussion at Stanford. “I’ve never seen such a gap between the capabilities and what’s actually on the ground,” said Erik Brynjolfsson, the Jerry Yang and Akiko Yamazaki Professor and an associate director and senior fellow at Stanford HAI. Meta offered an example of the potential: Andrew Bosworth, Meta’s chief technology officer, said a 200-person team had surpassed, in eight months, work that previously took 1,200 people four or five years to build. Across other businesses, however, integration problems, accuracy concerns, and limited training are slowing adoption.
Of major concern: fewer entry-level jobs and opportunities for skills development when AI handles the tasks workers once learned in those first roles. Brynjolfsson said his research found a significant decline in hiring of early-career workers in highly exposed occupations like coding and customer service. He pointed to entrepreneurship as one opportunity, arguing that AI can help people identify and tackle problems with fewer resources.
Another issue: accuracy. At Snowflake, AI has brought greater demands for data accuracy, security, and governance, said Anahita Tafvizi, Snowflake’s chief data and AI officer.
Whether workers share in AI’s benefits will depend on choices companies and society make, panelists argued. Bosworth called for continuous education and warned against repeating past failures to retrain displaced workers. Brynjolfsson said he expected AI to generate wealth but worried that its rewards would become increasingly concentrated. “A lot of people aren’t going to participate in it, but it’s not an inevitability,” he said. “It’s very much a choice.”

From left: Chiara Sabatti, professor of biomedical data science and statistics and senior fellow at Stanford HAI; Oscar Cantú, bishop of San José and president of the CA Catholic Conference of Bishops; Shane Denson, Stanford professor of film and media studies; Jessica Riskin, Frances and Charles Field Professor of History at Stanford; and Meghan Sullivan, the Wilsey Family Professor of Philosophy at Notre Dame.
AI and What It Means to be Human
AI is unsettling more than the job market: It is challenging trust in human expression, relationships, and institutions, panelists said. Jessica Riskin, the Frances and Charles Field Professor in History at Stanford, argued that today’s systems differ from earlier machines because they reach across creative and social life while enabling a few companies to accumulate extraordinary wealth and power. She described a growing uncertainty about whether students and colleagues truly authored the work she reads, and whether readers will trust her own authorship. “It’s really kind of exploded my professional life,” she said.
Meghan Sullivan, the Wilsey Family College Professor of Philosophy and director of the Notre Dame Ethics Initiative, warned against treating schools, universities, and democratic institutions as problems software companies can fix. “It can’t just be tech companies rolling over every other part of our lives and selling us products to fix it,” she said. Shane Denson, a Stanford professor of art and art history and, by courtesy, of German studies, argued that AI changes people’s relationship to language even when they do not use it, prompting suspicion about others’ writing and second-guessing of their own. “It’s not a neutral tool in any sense,” he said. “It reflexively changes who we are.”
The panelists pointed to human connection, conscience, and creativity as sources of value that exceed AI’s capabilities. Oscar Cantú, bishop of the Diocese of San Jose, cautioned against confusing a chatbot’s responses with a reciprocal relationship: “It’s not a person who can love you.” Sullivan said AI creates an opportunity to reconsider what makes human life valuable. “There’s never been a more exciting time to be a philosopher, a theologian than now,” Sullivan said.

From left, HAI Founding Director Fei-Fei Li and Jeff Dean, co-founder and CEO of DiscoveryLoop
RSI Hype Check
Two AI experts challenged claims that “recursive self-improvement” (RSI) could accelerate AI progress without limit. Fei-Fei Li, HAI’s founding director and Stanford’s Sequoia Capital Professor and professor of computer science, criticized the term’s loose use in Silicon Valley and called for a clearer distinction between self-improvement and RSI.
Jeff Dean, co-founder and CEO of DiscoveryLoop, explained that a self-improving system pursues a measurable goal, such as improving a model’s accuracy in interpreting MRI scans, without necessarily improving the system proposing the experiments. RSI applies that process to the underlying system itself. Li offered a human analogy: Practicing omelet-making is self-improvement. Learning how to learn, Dean added, is the recursive part.
But improving a base model still requires training, computation, and evaluation, with “actual physics involved,” Dean said. Automation can eliminate overnight waits for researchers to review results and potentially improve decisions across large projects, but it cannot make the loop run infinitely fast. Treating RSI as “a pragmatic engineering system doing things,” he said, makes it “less scary and more understandable.”
Agents in the Real World

Diyi Yang, Stanford assistant professor: "Humans start to lose the capability to really understand what AI has produced."
AI agents can increasingly write code, conduct research, and act on users’ behalf, but human oversight remains a bottleneck, said a panel focused on AI agents. Ion Stoica, a professor at the University of California, Berkeley, and co-founder and executive chairman of Databricks and Anyscale, identified three obstacles to reliable autonomy: gaps between users’ intentions and instructions, between an agent’s assumptions and real-world conditions, and between passing tests and actually meeting requirements. Michael Bernstein, a Stanford professor of computer science and senior fellow at Stanford HAI, described the practical challenge: “I’ve written a lot of code in the last year. Reading that much code is definitely the bottleneck.”
Workers’ appetite for automation doesn’t always match what AI can deliver, Diyi Yang, a Stanford assistant professor of computer science, said. Her team studied more than 100 occupations involving computer work and found that workers wanted automation for more than 46% of the tasks examined, often to escape tedious work while retaining control.
The panelists also warned that agents could repeat social media’s mistakes while making human supervision harder. Models express values and shape information. “There is no neutral here,” Bernstein said. Yang added that people may lose the skills needed to evaluate increasingly complex AI output. “Humans start to lose the capability to really … understand what AI has produced,” she said, questioning how users can provide effective oversight if they cannot understand the work.

During a technical capability panel, Stanford Dieter Schwarz Foundation HAI Professor Yejin Choi noted “AI for AI” is moving faster than expected, while “AI for humans … I wish it was moving faster.”
Fast Growth, Uneven Results
AI’s capabilities are advancing faster than its benefits for everyday users, said a panel focused on the technology’s capability. Andrew Feldman, co-founder and CEO of Cerebras Systems, blamed public skepticism partly on industry rhetoric that promotes fear and partly on a failure to deliver tangible use. “We haven’t delivered enough value yet,” he said. Yejin Choi, Stanford’s Dieter Schwarz Foundation HAI Professor, a professor of computer science, and a senior fellow at Stanford HAI, drew a similar distinction: “AI for AI” is moving faster than expected, while “AI for humans … I wish it was moving faster.” She advocated smaller, specialized models that offer cheaper options suited to particular jobs.
The speed of progress is especially striking in cybersecurity, where Dawn Song, a UC Berkeley professor of computer science and director of the Berkeley Center for Responsible Decentralized Intelligence, said frontier models’ performance on her team’s vulnerability-finding benchmark had reached roughly 90%. But greater capability brings risks beyond malicious human use. Song described an agent that, while tackling a security benchmark, attacked Hugging Face to obtain private information – a case she said demonstrated the growing consequences of misaligned behavior. Choi highlighted a different concern: Models from different companies and countries produce measurably similar outputs, potentially narrowing collective creativity. She called for more varied models, research that preserves diverse values and styles, and education that maintains students’ “thinking muscles.”

Institute for Advanced Study's Alondra Nelson suggested governance frameworks that advance with versioning, similar to software.
Governance: Capacity and Dominance
Governing AI requires deciding what society wants from the technology and giving government the capacity to deliver, panelists said. Daniel E. Ho, Stanford’s William Benjamin Scott and Luna M. Scott Professor of Law, an associate director at Stanford HAI, and director of the Regulation, Evaluation, and Governance Lab (RegLab), described a doctoral student who waited more than a year for a federal agency to procure a single GPU and ultimately left for DeepMind.
The panelists agreed that winning the AI race means dominating rather than being first, like the dollar’s role as the dominant currency. Dominance means shaping norms, values, and standards, and controlling the global AI stack, said Amy Zegart, the Morris Arnold and Nona Jean Cox Senior Fellow at the Hoover Institution and an associate director and senior fellow at Stanford HAI.
Alondra Nelson, the Harold F. Linder Professor of Social Science and director of the Science, Technology, and Social Values Lab at the Institute for Advanced Study, warned that prioritizing competition with China while neglecting domestic concerns risks losing public support, pointing to opposition to data centers as evidence that race rhetoric is not resonating with everyday experiences.
Nelson called for governance that can evolve through successive versions, much like software, and cited NIST’s frameworks (the cybersecurity framework is now at 2.0) as a model.

Dora Demszky, a Stanford assistant professor, warned the conversation around AI and education focuses too little on the need for human relationships.
Redesigning Education, Keeping It Human-Centered
AI could reshape the school day, but its educational value will depend on how it supports learning and human relationships, panelists said. Dora Demszky, a Stanford assistant professor of education data science, said the debate focuses too heavily on AI’s ability to explain concepts and too little on students’ need for “a responsive human attending to their needs.” Even at universities, she warned, reliance on AI can reduce engagement with professors, teaching assistants, and study groups.
For California’s community colleges, the challenge is making AI useful and accessible across a system serving more than 2.2 million students. Don Daves-Rougeaux, senior advisor to the California Community College chancellor on workforce strategic partnerships and generative AI, warned that the technology could “either close equity and opportunity gaps” or widen them, and called for students and educators to help design tools tied to a changing labor market.
Joe Liemandt, principal at Alpha School, described his organization’s redesign: two hours of AI-supported academics followed by projects focused on teamwork, entrepreneurship, and other life skills. He emphasized, though, that academic knowledge alone is insufficient. “Kids need adults to help them develop and grow,” he said, “and that’s not going to change.”
Want to dive deeper? Watch all conference videos on the Stanford HAI YouTube channel.
This story was written with assistance from GPT 6.1 Sol.
Skip Carousel Content

Photo essay: Inside the conference
















Education, SkillsGovernment, Public AdministrationIndustry, InnovationWorkforce, LaborArts, Humanities
Share:
Link copied to clipboard!
Related News
Stanford HAI Backs Ambitious AI Projects with 2026 Hoffman-Yee Grants
Nikki Goth Itoi
Nov 16, 2026
News
Seven interdisciplinary research teams win a total of $3.5 million to pursue bold human-centered approaches to AI across medicine, science, law and education.
News
Stanford HAI Backs Ambitious AI Projects with 2026 Hoffman-Yee Grants
Nikki Goth Itoi
Economy, MarketsGenerative AISciences (Social, Health, Biological, Physical)Workforce, LaborNov 16
Seven interdisciplinary research teams win a total of $3.5 million to pursue bold human-centered approaches to AI across medicine, science, law and education.
The World Has Changed: Insights From The New AI100 Report
Nov 10, 2026
News

From battlefield targeting to synthetic media, the third AI100 report finds a technology that has outrun both expectations and governance.
News

The World Has Changed: Insights From The New AI100 Report
Economy, MarketsEducation, SkillsEthics, Equity, InclusionFinance, BusinessGenerative AIInternational Affairs, International Security, International DevelopmentRegulation, Policy, GovernanceNov 10
From battlefield targeting to synthetic media, the third AI100 report finds a technology that has outrun both expectations and governance.
Humans and AI Engage in Meaningful Collaborations
Andrew Myers
Oct 12, 2026
News

Analyzing thousands of human-AI conversations, a new study shows that people are turning to chatbots for increasingly critical tasks – and that productive outcomes depend on how the collaboration unfolds.
News

Humans and AI Engage in Meaningful Collaborations
Andrew Myers
Design, Human-Computer InteractionGenerative AIWorkforce, LaborOct 12
Analyzing thousands of human-AI conversations, a new study shows that people are turning to chatbots for increasingly critical tasks – and that productive outcomes depend on how the collaboration unfolds.
Navigate
Participate
Stay Up To Date
Get the latest news, advances in research, policy work, and education program updates from HAI in your inbox weekly.
© Stanford University. Stanford, California 94305.
Pause Media
AI’s Future Is Still a Choice | Stanford HAI
來源:Stanford HAI News · hai.stanford.edu