How an AI Safety Researcher Is Teaching the Next Generation of Engineers

Amanda Askell, a researcher at Anthropic, has joined Stanford University's CS 153 course as a guest lecturer. The course meets Wednesday afternoons from 12:30 to 2:20 PM, bringing her industry work directly into the classroom.

Askell is a research scientist at Anthropic, a company focused specifically on AI safety. Her job involves making sure AI systems—especially large language models like Claude—behave in ways that are helpful, harmless, and honest. Her work centers on two key techniques: constitutional AI (a method of training models to follow a set of principles, similar to a rulebook) and human feedback systems (where people review and correct AI outputs to improve its behavior).

Bridging Industry and Academia

This appointment reflects a broader pattern: Stanford brings working technologists into the classroom to teach students what's actually happening in industry right now, not just what textbooks say.

Askell's background is unusual in AI. She has a PhD in philosophy from New York University, where she studied decision theory and philosophy of mind. This matters more than it might sound. As AI systems have become more powerful, questions about how to align them with human values—how to make sure they do what we actually want—have become technical problems that benefit from philosophical thinking. Askell's transition from philosophy to AI safety research reflects a broader trend: philosophers and ethicists are increasingly moving into hands-on AI development roles.

What Makes Anthropic Different

Anthropic positions itself as an AI safety company. The core difference is in how it trains models. Rather than relying only on human feedback to shape AI behavior, Anthropic uses what it calls "constitutional AI"—training the model against a set of explicit principles. Think of it as the difference between teaching someone by example versus teaching them a clear set of rules to internalize. The goal is to make AI behavior more predictable and controllable as the models become larger and more capable.

The Claude model family is where these techniques show up in practice. Claude is trained using three complementary approaches: supervised fine-tuning (teaching it patterns from human examples), reinforcement learning from human feedback, or RLHF (rewarding good outputs and penalizing bad ones), and constitutional AI training (applying explicit principles).

Worth flagging: Anthropic's emphasis on safety research comes as the broader AI industry faces increasing scrutiny over how and where AI systems are being deployed, and what risks they might pose.

What Students and Industry Get Out of This

Guest lecturers like Askell serve a practical purpose on both sides. Students encounter real problems that aren't yet in textbooks—challenges that matter to companies shipping AI systems today. Meanwhile, industry practitioners benefit from spending time in academic settings, where people think systematically about fundamental problems without the time pressure of shipping products.

The timing matters too. Universities are now adding AI safety and alignment into their core courses, alongside the traditional material on optimization and neural networks. This shift reflects something the field has come to accept: safety and capability have to advance together. You can't bolt safety on afterward.

A Pattern We've Seen Before

We have seen this before. In the 1990s, when the commercial internet was taking off, companies like Netscape, Yahoo, and early Google placed engineers in university classrooms to teach web technologies and protocols. Those partnerships worked: students learned bleeding-edge techniques, and companies got access to academic research and a pipeline of talent.

The current AI-academia collaboration follows similar logic, but the stakes are higher. Web technologies changed how we access information and do business. AI systems increasingly make decisions in healthcare, transportation, finance, and criminal justice—domains where errors or biases carry real human cost.

Location and Timing

Stanford's computer science department has long had close ties to the tech industry—its proximity to Silicon Valley makes that natural. Guest lectures, joint research projects, and internships are standard. That advantage becomes particularly relevant as AI companies concentrate in the Bay Area.

The Wednesday afternoon timing is practical: it lets an industry professional with a full-time job still commit to teaching substantive material without logistics becoming a burden.

Analysis: This appointment signals that AI safety research, once considered a specialty concern for a small group of researchers, now has institutional weight in mainstream computer science education. As AI capabilities advance, universities are recognizing that graduates need to understand not just how to build AI systems, but how to build them responsibly.

What This Means Going Forward

Askell's role as a guest lecturer is one sign among many that AI safety is moving from the academic margins into the center of how the industry and universities work together. Her philosophical training and industry experience position her to ask questions that purely technical approaches might miss: How should AI systems interact with humans? What are the broader social implications of deploying these tools?

It also reflects something real about how AI development has changed. Companies like Anthropic and OpenAI now have dedicated safety teams whose work shapes actual product development, not just academic papers. That's a shift from even five years ago.

For Stanford students, hearing directly from Askell provides context for technical skills they learn elsewhere. Constitutional AI methods, alignment techniques, human feedback systems—these are no longer theoretical. They're embedded in systems millions of people use. As those students go on to careers where they may build, deploy, or help regulate AI systems, that direct exposure to current safety research becomes genuinely useful.

This appointment underscores how academic computer science and the AI industry have become tightly intertwined. Theory and practice increasingly inform each other in real time, rather than waiting for research to percolate into industry practice years later.