From Validation to AI Assurance in Automotive
Automotive AI Series Blog 4
Turning Connected Evidence into Justified Trust
Validation Leaves Us With an Important Question
The first three blogs (Blog 1, Blog 2, Blog 3) in this Automotive AI Blog series established an important foundation: Validation activities generate valuable evidence about AI-enabled automotive systems. Yet evidence alone does not automatically explain whether confidence in that system is justified.
Engineering teams may have access to test results, simulation outcomes, robustness analyses, explainability reports, operational observations, and many other artifacts. Each of these can provide insight into a particular aspect of system behavior. However, the existence of evidence does not automatically explain what that evidence means, whether it remains relevant, or how it supports confidence in the overall system.
This distinction becomes increasingly important as AI-enabled systems grow more complex. Confidence depends not only on individual validation results, but also on understanding how those results relate to risks, assumptions, intended operating conditions, and the evolving system lifecycle.
This is where the discussion begins to shift from AI validation to AI assurance. Validation activities generate evidence. AI assurance provides the structure needed to understand what that evidence means, how it relates to engineering decisions, and whether it collectively supports confidence in the system [1].
Validation Generates Evidence. AI Assurance Explains Confidence.
Validation and AI assurance are closely related, but they answer different questions.
Validation activities generate evidence. They help engineering teams understand how an AI-enabled system behaves under specific conditions. Testing, simulation, robustness assessments, explainability analyses, operational monitoring, and other validation activities all contribute valuable insights. Together, they help build an evidence base for understanding system behavior.
AI assurance begins where individual validation activities end: Its purpose is not to generate more evidence, but to determine what the available evidence means, how it relates to the intended application, and whether it collectively supports confidence in the system.
Figure 1. Testing, Validation, and Assurance Answer Different Questions
This distinction becomes increasingly important as AI-enabled systems become more complex. Engineering teams rarely rely on a single result when making decisions. Confidence is typically built from multiple sources of evidence, accumulated across different lifecycle stages, generated by different teams, and interpreted within a specific operational context.
AI assurance therefore focuses on maintaining the relationship between evidence and the claims supported by that evidence.
It connects validation results with assumptions, operating conditions, known limitations, and engineering decisions. Rather than treating individual validation activities as isolated assessments, assurance seeks to explain what the collective body of evidence says about the behavior of the overall AI-enabled system.
Viewed this way, AI assurance does not replace testing or validation. It provides the broader structure needed to interpret their results, maintain their relevance as systems evolve, and understand whether confidence in the system remains justified over time.
From Evidence to Justified Trust
Validation activities generate large amounts of information about AI-enabled systems. Test results, simulation outcomes, robustness assessments, explainability analyses, operational observations, and monitoring data can all provide valuable evidence. Yet evidence alone does not automatically explain why confidence in a system is warranted.
Figure 2. From Validation Evidence to Justified Trust [2]
AI assurance begins with the recognition that evidence must be interpreted within its intended context. Validation results are only meaningful when considered alongside the risks, assumptions, requirements, system boundaries, and operating conditions under which they were generated. A robustness result, for example, may be highly relevant in one operational context while offering little confidence in another.
This is where the concept of an assurance argument becomes important. [3] The role of the assurance argument is not to generate additional evidence, but to explain why the available evidence supports a particular claim. It provides the logical connection between what has been observed, the conditions under which it was observed, and the confidence that can reasonably be derived from those observations.
Viewed together, three elements are required:
- Evidence provides the observable results generated throughout the AI lifecycle.
- Context explains the assumptions, risks, requirements, operating conditions, and system characteristics that give those results meaning.
- Assurance arguments explain why the available evidence supports a particular confidence or safety claim.
In this context, justified trust should not be understood as blind trust. Rather, it is trust grounded in evidence, context, and a defensible assurance argument. Only when these elements are combined can organizations move beyond isolated validation results and toward a defensible understanding of AI behavior.
This distinction is important. AI assurance does not claim that a single test, metric, or analysis can prove an AI system safe or trustworthy. Instead, it seeks to answer a different question:
Given the available evidence, operating context, assumptions, and known limitations, why is confidence in this AI-enabled system justified?
In other words, validation generates evidence. AI assurance explains what that evidence means and why it supports confidence in the system.
That confidence cannot remain static. As systems evolve, assumptions change, software is updated, and new operational evidence emerges, the relationships between evidence, context, and assurance arguments must evolve as well. This is why AI assurance must extend beyond individual validation activities and span the entire AI lifecycle.
AI Assurance Must Span the Lifecycle
AI assurance is not a one-time activity performed at the end of development. It is a lifecycle discipline that evolves alongside the AI-enabled system itself. As discussed throughout this series, systems change, software is updated, operating conditions evolve, and new evidence continuously emerges during real-world operation. Confidence must therefore be maintained, not assumed.
Figure 3 reflects the lifecycle-driven AI validation framework discussed in our white paper Rethinking AI Validation. A more detailed discussion of the underlying methodology and lifecycle stages can be found there. It reflects five connected stages, from data and problem analysis through feature engineering, model development and evaluation, to continuous inferencing during real-world operation. These stages are not isolated activities. Together, these stages generate the evidence that AI assurance seeks to connect, interpret, and maintain throughout the lifecycle
The lifecycle perspective recognizes that confidence is built from evidence generated across all stages of development and operation, not from any single validation activity.
A model limitation discovered during evaluation may have roots in dataset composition. An operational anomaly may reveal an assumption that was never challenged during development. A software update may affect evidence that previously supported confidence in the system.
AI assurance therefore requires more than collecting validation results. It requires maintaining the relationships between evidence, assumptions, system changes, and operational experience throughout the lifecycle. Evidence generated during data analysis, model development, evaluation, and continuous inferencing must remain connected if organizations want to understand how confidence in the system is affected as conditions evolve.
This is also why the figure should be understood as a continuous assurance loop rather than a linear development process. Real-world operation continuously produces new evidence that may trigger additional analysis, re-evaluation, or updates to earlier assumptions. Assurance is therefore not a final project milestone. It is an ongoing process of maintaining confidence as systems and environments change over time.
When viewed this way, AI assurance provides the structure needed to interpret lifecycle evidence and maintain confidence in an evolving AI-enabled system. Validation generates the evidence. Assurance helps explain what that evidence means in context.
Turning AI Assurance Into Engineering Practice
The concepts discussed so far may appear straightforward in principle: gather evidence, interpret it in context, and maintain confidence throughout the lifecycle. In practice, however, this requires engineering teams to manage large amounts of evidence generated across datasets, models, systems, and real-world operations.
For lifecycle-driven AI assurance approaches, the intended contribution is therefore not simply the execution of individual validation activities. It is supporting the generation, organization, traceability, analysis, and interpretation of evidence across the AI lifecycle.
Solutions such as AX1000A AI Software Integrity Builder are intended to help engineering teams maintain visibility across datasets, models, and operational behavior while preserving the relationships between validation activities, evidence, and engineering context.
This enables teams to move from isolated validation artifacts toward a more structured understanding of system behavior and confidence throughout the lifecycle.
From Validation to AI Assurance
Throughout this series, we explored why performance alone cannot establish confidence in AI-enabled systems. Validation generates valuable evidence, but that evidence can lose relevance as systems evolve, become fragmented across lifecycle stages, or remain disconnected from the assumptions and context needed to interpret it.
AI assurance provides the structure needed to understand what the available evidence collectively says about the behavior of an AI-enabled system. It preserves the relationships between evidence, risks, assumptions, operating conditions, system changes, and real-world behavior, helping organizations recognize where limitations exist and when earlier conclusions need to be re-evaluated.
AI validation generates evidence. AI assurance connects evidence, context, and argument to enable justified trust.
Organizations beginning this journey do not need to start from scratch. A practical first step is to identify where evidence is generated today, where lifecycle gaps exist, and where critical assumptions, validation results, and operational insights become disconnected.
That journey from validation to AI assurance is only beginning, and it is likely to become one of the defining engineering challenges for the next generation of AI-enabled automotive systems. Learn how Keysight’s AX1000A AI Software Integrity Builder can help engineering teams connect validation evidence across the AI lifecycle and build a structured foundation for justified trust.
References:
[1] Centre for Assuring Autonomy, University of York: Body of Knowledge Definitions. https://www.york.ac.uk/assuring-autonomy/about/aaip/body-of-knowledge/definitions/
[2] Concept inspired by assurance terminology used within the Assuring Autonomy International Programme (AAIP), University of York, https://www.york.ac.uk/assuring-autonomy/about/aaip/body-of-knowledge/definitions/
[3] Centre for Assuring Autonomy, University of York: Body of Knowledge Definitions. https://www.york.ac.uk/assuring-autonomy/about/aaip/body-of-knowledge/definitions/