Waymo's Blueprint for AI Safety: Lessons from Autonomous Vehicles

As artificial intelligence (AI) technologies become increasingly prevalent across various industries, the stakes involved in deploying these systems are higher than ever. Few companies illustrate this point better than Waymo, the autonomous vehicle pioneer spun out from Google. With a mission to create a safe and efficient self-driving experience, Waymo’s approach to AI evaluation and safety sets a powerful precedent for enterprises across sectors. In a recent discussion at the VB Transform 2026 conference, Manasi Joshi, Waymo's Director of Engineering for Systems Intelligence and Machine Learning, detailed the rigorous methodologies that underpin their AI systems, particularly focusing on continuous evaluation and human oversight.

Waymo has already logged over 220 million miles in fully autonomous driving, boasting a safety record with 17 times fewer serious crash injuries compared to human drivers over equivalent distances. This remarkable achievement is not just a stroke of luck; it is a result of what Joshi refers to as “eval-forced development,” where evaluation is integrated into the engineering process from the beginning rather than serving as a final check before deployment. This philosophy can serve as a vital lesson for any organization looking to harness AI technology responsibly.

Understanding Eval-Centric Development

The concept of eval-centric development emphasizes that the evaluation of AI systems should not be an afterthought but rather an integral component of the development lifecycle. Joshi explains that at Waymo, the maturity of a project can be gauged by the robustness of its evaluation processes. This approach extends far beyond simply ensuring that a model performs well; it involves establishing rigorous metrics and benchmarks that reflect real-world performance and safety.

For businesses looking to implement AI, the implications are clear: If a company cannot reliably measure an AI system's performance, it is likely unprepared to deploy it. This framework should prompt organizations to ask critical questions about their own AI systems, such as:

• What evaluation metrics are in place?
• How frequently are these evaluations conducted?
• Do the evaluations correlate with real-world outcomes?

Continuous Evaluation as a Core Principle

Waymo’s commitment to continuous evaluation means that assessment occurs throughout the lifecycle of a project. Joshi notes that evaluations are not limited to pre-launch testing; they encompass a wide variety of scenarios, including simulations and real-time operational assessments. This method recognizes that AI systems are not static; they evolve as they interact with new data, user behaviors, and changing environments.

For enterprises, this means that ongoing evaluation is essential even after deployment. It is crucial to monitor how AI systems adapt to real-world challenges and to measure their impact on business objectives. Organizations should prioritize the development of infrastructures capable of conducting these evaluations efficiently, ensuring they remain agile in a rapidly changing technological landscape.

Safety First: Testing Under Realistic Conditions

At the heart of Waymo's evaluation strategy is an unwavering focus on safety. The company utilizes a diverse array of data sources, including first-party driving logs and third-party datasets, to expose their systems to a comprehensive set of driving scenarios. These include complex environments such as railroad crossings and construction zones where traditional metrics may fall short.

Waymo's strategy highlights the importance of testing not just for routine situations but also for rare and potentially dangerous scenarios. This principle is universally applicable across industries. For example, a financial AI system must be tested against fraud detection in unusual patterns, whereas healthcare AI must account for rare medical conditions that could lead to severe repercussions if misdiagnosed.

Real-World Applications and Implications

For businesses, the takeaway is clear: simply achieving high performance in common situations is insufficient. Rigorous testing must also account for edge cases that could result in significant financial, legal, or reputational fallout. This requires a blend of technology and human oversight. Joshi emphasizes that while automation plays a role, human judgment is irreplaceable when lives are at stake.

Balancing Efficiency with Reliability

As demand for AI capabilities surges, enterprises often face challenges related to resource allocation. Waymo, like many organizations, must balance the need for computational power with the requirement for efficiency. Joshi points out that data efficiency is critical—selecting the most informative training examples can yield better results than simply increasing data volume.

Since adopting transformer models in 2017, Waymo has explored various AI architectures, including large language models and vision-language-action models. This ongoing innovation is essential, not only for improving AI capabilities but also for ensuring that the systems remain scalable and efficient. Businesses should similarly focus on optimizing their AI processes, recognizing that efficiency should never come at the expense of reliability.

AI Agents: Internal Evaluation and Accountability

Interestingly, Waymo also deploys AI agents internally to streamline engineering tasks. These agents assist in data analysis and problem triage, providing valuable insights that help engineers focus on deeper technical challenges. However, just as external AI systems are evaluated, these internal agents undergo scrutiny to ensure they deliver trustworthy results.

The crucial lesson for organizations is that deploying AI agents requires more than simply selecting advanced models. A clearly articulated objective, robust evaluation data, continuous testing, and designated human accountability are all essential to fostering trust in AI deployments. As Joshi aptly states, “Earning trust is supremely important.”

Key Takeaways

• Eval-centric development: Make evaluation a core part of AI engineering from the outset.
• Continuous evaluation: Implement ongoing testing to adapt to changing conditions and maintain performance.
• Safety first: Test AI systems under realistic conditions, including edge cases and rare scenarios.
• Efficiency vs. reliability: Balance resource demands with the need for dependable AI performance.
• Accountability: Ensure human oversight in AI deployment and evaluation processes.

Frequently Asked Questions

What is eval-centric development?

Eval-centric development is an approach where evaluation is integrated into the AI development lifecycle from the beginning. This methodology ensures that the performance and safety of AI systems are continually assessed rather than being evaluated only at the end of the development process. Organizations adopting this approach are better equipped to identify potential issues early on and make necessary adjustments before deployment.

How does continuous evaluation improve AI systems?

Continuous evaluation allows organizations to monitor AI systems in real-time as they operate in dynamic environments. By regularly assessing performance against real-world metrics, companies can ensure that their AI technologies remain effective and safe over time. This ongoing scrutiny helps detect and address issues that may arise due to changes in user behavior, data input, or external conditions.

Why is safety a priority in AI evaluation?

Safety is paramount in AI evaluation, particularly for systems that impact human lives, such as autonomous vehicles. Rigorous testing under realistic scenarios helps identify potential risks and ensures that AI systems can handle dangerous or unexpected situations. Prioritizing safety not only protects users but also builds trust in AI technologies, which is essential for broader adoption.

What role does human oversight play in AI systems?

Human oversight is critical in AI systems to ensure accountability and ethical decision-making. While AI can automate many processes, the potential for error remains. Human judgment is essential in reviewing AI-generated outcomes, making deployment decisions, and addressing any ethical concerns. This oversight is particularly important in high-stakes situations where the consequences of mistakes can be severe.