AI 造成傷害時誰應負責？Deb Raji 談問責與政策執法

Who is Responsible When AI Causes Harm?

Date

October 14, 2026

Topics

It's one of AI policy's hardest unsolved problems. Expert Deb Raji joins Stanford HAI to find answers that move beyond principles and into practice.

Deb Raji (Inioluwa Deborah Raji) focuses on a deceptively simple set of questions: When an AI system causes harm, who is responsible, and how do you prove it?

Raji joins the Stanford Institute for Human-Centered AI as a Visiting Fellow this fall and as center fellow and assistant professor in computer science in fall 2027, following a fellowship year at the Institute for Advanced Study and the completion of her PhD in computer science at UC Berkeley. Her research on bias in commercial facial recognition, conducted with scholar Joy Buolamwini and the Algorithmic Justice League, helped push IBM and Amazon to stop selling the technology to police departments. While at Google, she helped create Model Cards, now a widely used standard for disclosing what a model can and cannot do. Her work has been recognized by Time, MIT Technology Review, and the Electronic Frontier Foundation.

In this conversation, Raji shares what she’s working on now, why enforcement has become AI policy's hardest unsolved problem, and why worrying about catastrophic risk overlooks the mundane disasters.

At the highest level, what are you currently exploring?

AI accountability infrastructure. We've done a lot of thinking about how to build new AI policy and legislation, and now that we've made some progress, there's this big question of enforcement. How do we actually get companies to change their behaviors? How do we build the government capacity to enforce what's on the books?

A lot of that turns out to require technical infrastructure. If a bill has a transparency requirement, what does that documentation template actually look like? If we want incident monitoring, what does the database look like — what constitutes a valid incident, and what's the appropriate response from the government and from the company? We've been talking about AI auditing for five years now, mostly in a high-level, almost speculative way. I think we're finally arriving at the point where there's enough will from enough actors to work through the pragmatics: what tools are necessary, what policies, what institutions we need to build. I'm excited to have my hands on the ground for that.

Is there a particular project you'd highlight?

I've become very interested in AI use in government. I've been in conversation with national AI safety institutes, civil society groups, and government organizations that want to use these systems and want to do it responsibly — they're thinking about procurement, safeguards, contestation, all the accountability measures you'd want given the stakes. But they’re worried because existing benchmarking practices and off-the-shelf guardrails don't meet government standards. Public agencies have IT laws, accessibility requirements, privacy rules, performance expectations. It's a much stricter deployment environment than a typical firm.

I've been reading a lot of Herbert Simon on administrative and organizational behavior, trying to work out how much of those methodological tools help us understand AI deployment in regulated, public-interest settings. Government is one. Education and healthcare are others.

Why do those settings matter so much to you?

Because you can't opt out of them. When people talk about the risks of AI, especially for young people, there's often this response of "just turn it off" — the same thing we heard about social media. Just get off Facebook.

But in these organizational settings, the users are government workers and the affected population is us. We don't have the option. In many cases we don't even know that AI shaped an outcome that ripples down into our lives, and without a policy-mandated notification, we're certainly not in a position to contest it or file an incident report. I don't think people realize that this is disproportionately how most people encounter AI — not by choosing to log onto a platform, but by living in a society where institutions use these tools in ways that affect them by default.

There's a capacity problem underneath it, too. A lot of agencies, especially in certain states, are operating on a shoestring. The idea of an everything machine that handles your most labor-intensive administrative work is incredibly seductive. But they also don't have the resources to do due diligence on the purchase, so they end up accepting the vendor's narrative because they can't afford to challenge it.

"Human-centered" gets defined pretty loosely. What does it mean to you?

When people talk about AI, we tend to dehumanize the actual humans — the humans in the data, the humans making choices about the model, the humans affected by it — and then we anthropomorphize the model itself. It's a strange inversion. People forget where the humans are.

That matters for my work specifically, because the legal and organizational levers I push on all depend on being able to ask: who is the stakeholder, and what did they do or fail to do that produced this result? If you obscure the relationship between someone's actions and an outcome, that's how accountability disappears. The default when something goes wrong is to put your hands up and blame cosmic forces. It reminds me of a toddler swearing and the parent saying, "I have no idea where they learned that." You did, actually. You made choices, and they produced outcomes.

What's the most important issue in AI policy right now?

Enforcement, and I don't think it's close.

Around 2017 the conversation was all high-level principles. From roughly 2020 to 2023 we got serious legislative proposals. After ChatGPT, we actually started getting laws — New York City's Local Law 144, which requires bias audits of AI hiring tools, is one I think about a lot. SB 53 (California’s landmark AI bill) is another.

But I think we badly underestimated the enforcement challenge. Do existing agencies have the capacity to scrutinize these systems under the rules already on the books? Are the enforcement tools we use in other domains even applicable here? Is the ecosystem of third-party actors as mature as it is in other fields? It doesn't seem like it.

There were clues. Look at how litigation over terms of service played out — knowing intuitively that something shouldn't be happening is very different from collecting the evidence and actually getting the expectation enforced. The nitty-gritty of applying a law is completely non-trivial, and it's where I want to spend my time.

A lot of policy attention has shifted toward catastrophic risk. What do you make of that?

My honest judgment is that the bar is set too high. The definitions in SB 53 and in the Frontier AI Act (currently being debated in Congress) are oriented around extreme physical consequences — death, injury, property damage, dollar thresholds in the billions. Almost nothing meets that bar. Things I would consider genuinely significant harms do not meet that bar.

And a lot of the significant harms already happening fall disproportionately on marginalized people, which makes the evidence even harder to collect. So you get this situation where a great deal flies under the radar precisely because the threshold for accountability is so high — and I don't think it needs to be. We can hold people accountable for the smaller harms we can see today.

The other thing I'd say is that the framing imagines one explosive event. From what I've seen in mental health cases and elsewhere, the effects are often pernicious and cumulative — they build over months, years, maybe decades. It might not be a bomb going off. It might be waking up years later and realizing this isn't the world we wanted with AI.

What brought you to HAI?

What drew me here is that HAI has taken policy seriously since its origin. So much of what human-centered AI means in practice is setting guardrails and creating incentives that counteract the commercial pull these companies are under, so the technology contributes to human flourishing instead of disrupting people's lives.

I got pulled into policy work early, before my PhD, because of the kind of research I do — evaluation, third-party auditing. And it escalated to the point of being almost unsustainable. HAI had already built the internal infrastructure to make that kind of engagement sustainable alongside a research agenda. I felt like I'd have the support I needed to do the work I want to do.

Any advice for students entering this field?

I meet so many students who want to use their technical training for a public-interest purpose, and policy implementation is close to a no-brainer for that — it's a place where a computer scientist can contribute directly to accountability outcomes.

Practically, they need to work with people already operating at that intersection, and take advantage of the programs built to bridge the gap, whether that's a Hill internship, a DC-facing fellowship, or a research Institute like HAI. A lot of civil society groups could use student interns and don't advertise it well. I keep track of those opportunities and I'd like to find a better way to get them in front of students.

And for policymakers?

Two things. First, reach out directly to researchers whose work you find useful. I used to keep a standing block on my calendar as policy office hours, and I'm hoping to restart it.

Second, seek out independent technical perspectives. I get some inbound about expert testimony, and I've come to appreciate how much it matters — especially to civil society groups — when an academic shows up. It shouldn't just be a panel of CEOs. Companies are structurally limited in how candid they can be about how well a system is or isn't working. Academics and civil society technologists are coming from a more grounded place.

AI policy is a genuinely wicked problem, and it needs many disciplines in the room. Coming from computer science, it's been hard to find spaces that are organically open to that. HAI is one, and building those interdisciplinary spaces is something I think about constantly in my research and my teaching.

Deb Raji is a Visiting Fellow at Stanford HAI and will join as a center fellow and assistant professor in computer science in fall 2027. She serves on the advisory boards of the Center for Democracy and Technology AI Governance Lab, Health AI Partnership, TeachAI, REALML, and the Center for Civil Rights and Technology. She has been named to Forbes 30 Under 30, MIT Technology Review’s 35 Innovators Under 35, the Time 100 Most Influential in AI, the Tech for Humanity Prize (2024), the Mozilla Rise 25 (2024), and the EFF Pioneer Barlow Award, with Joy Buolamwini and Timnit Gebru.

Link copied to clipboard!

Authors

###### Shana Lynch

Inioluwa Deborah Raji

Visiting Fellow, Stanford HAI

Related News

Can AI Be Slowed Down? Stanford HAI Experts Weigh the Risks, Rules and Race Ahead

Shana Lynch

Sep 22, 2026

News

In a new series called Prompt Response, Stanford’s Surya Ganguli, Diyi Yang and Rob Reich examined emergent agent behavior, recursive self-improvement, independent evaluation and whether a kill switch can make advanced AI safer.

News