This article is based on a keynote I delivered at the United Nations Headquarters during the ICPD30 Global Dialogue on Technology in June 2024. The event brought together governments, technology companies, researchers, civil society organizations, and international institutions to discuss a critical question: How can technology create a more equitable future?

The central message of that presentation has only become more relevant as AI capabilities continue to advance.

Most conversations about AI safety focus on what the model does wrong: hallucinations, harmful content, bias, and other technical failures. These are real problems. But they are often downstream of a more fundamental one:

What happens when safety is not built into a product from the beginning?

I’ve spent much of my career working at the intersection of technology, public policy, behavioral science, and Trust & Safety. At the World Bank, I saw how technology could improve access and economic opportunity at scale. At Uber, I saw how product design decisions could directly influence real-world safety outcomes. At Match Group, I saw how scammers, stalkers, and other bad actors exploited product features in ways that extended beyond their original design assumptions.

Across these very different environments, I kept encountering the same lesson: the most dangerous failures aren’t AI failures. They’re design failures.

AI Doesn't Create Harm. It Scales It.

Technology-facilitated gender-based violence is one example. But the same pattern appears across online scams, harassment, child exploitation, privacy violations, misinformation, and financial fraud.

These harms are not new. What is new is the scale, speed, and reach with which technology can amplify them.

A scammer who once targeted dozens of victims can now target thousands. A manipulator can leverage algorithms and AI to influence millions. A stalker can access information that once required physical proximity.

The internet did not invent manipulation; it made manipulation scalable. AI did not invent bias; it can make bias faster, cheaper, and more persuasive.

This distinction matters because it changes how we think about AI safety. The central challenge is often not the technology itself, but how technology interacts with existing human vulnerabilities.

What makes harm scale is rarely malicious AI. More often, it is the absence of intentional design—products built for speed, engagement, and growth without sufficient consideration of how they might be misused, exploited, or weaponized.

The “Move Fast and Break Things” Era Left a Large Attack Surface

For years, the technology industry embraced the philosophy of "move fast and break things." It accelerated innovation, but it also left products vulnerable to predictable forms of misuse.

Across the organizations where I've worked, I observed the same pattern. The problem was rarely a lack of intelligence, resources, or good intentions. More often, it was that safety entered the conversation too late.

Product teams focused on functionality. Growth teams focused on adoption. Engineering teams focused on performance. Each priority was reasonable on its own.

Yet one critical question was often missing:

How could this feature be misused?

That question lies at the heart of Safety by Design.

Safety by Design begins with a simple assumption: misuse is not an edge case—it is an expected part of the product environment. Rather than reacting to incidents after they occur, it integrates harm prevention into every stage of product development, anticipates foreseeable misuse, and designs with vulnerable users in mind from the very beginning.

Consider seatbelts and anti-lock braking systems in cars. No one argues that these safety features should be optional or added only after accidents occur. They are built into the product because predictable risks deserve proactive protection.

The digital world—and increasingly, AI systems—needs the same mindset.