Beyond Code Review: Building Trust in AI-Assisted Development with Engineering Evidence - Interview with Dr. Martin Becker, MathWorks

Posted on August 04, 2026

In this interview, Dr. Martin Becker, Principal Application Engineer at MathWorks, explores how AI-assisted development requires a shift away from traditional code review towards higher-level evidence. He explains why AI-generated code cannot be treated in the same way as handwritten code, discusses the importance of trust and assurance, and describes how AI tools can be leveraged to produce trustworthy engineering evidence.

Dr. Becker will be speaking at HISC 2026, presenting the talk “Beyond Plausible Diffs: Assurance Evidence for AI-Assisted Code Changes”. Register now to secure your place.

Why is code review no longer enough when assessing AI-generated code?

AI can produce vast amounts of code in almost zero time. No human can keep up with that speed. Additionally, the code often looks plausible but can be functionally flawed and unnecessarily complex. Finding issues only at the code-review stage is too late, and humans are not especially good at identifying these issues manually. If we want to benefit from AI technology, we must replace manual code inspection with automated tools that provide deterministic corrections and hard evidence at a comparable speed. The engineering mindset should, in fact, shift away from code review entirely.

What shift in mindset is needed when moving to AI-assisted development?

The required shift is from seeing code as the main artifact, towards a structured workflow where assumptions, design decisions, and identified risks become the main artifacts for review. Engineers define intent, then AI generates and iterates solutions. During this, it continually verifies its own work and produces documentation and evidence that enables humans to understand its interpretation and decisions. The human responsibility can then move toward validation, i.e., answering the question “are we doing the right thing?”, instead of running verification tasks or reviewing implementation details such as code.

How can we trust the evidence that AI is generating?

This is the most important point in AI workflows: We can’t. We should treat outputs as unverified until proven otherwise. Since AI systems are probabilistic by nature and do not have a notion of uncertainty, they cannot verify themselves reliably. They can be “confidently wrong”, which is why engineering evidence must come from external, trustworthy sources such as static analysis, executed test cases, and formal verification. These methods are not biased by subjective interpretation, and they can provide a ground truth. In other words, the evidence must not come from the AI itself, but externally from trusted verification tools integrated into the workflow.

How should organizations approach evidence and assurance in AI-driven workflows?

This requires clear acceptance criteria, processes, and validation of evidence. Evidence becomes the foundation for decision-making, not intuition or plausibility. To make it concrete: You start with integrating verification tools in the agentic loop, and you enforce rules like “zero MISRA violations” and “100% MC/DC coverage”. You instruct your agent to follow skills, document decisions, assumptions and open risks, and you put strong guardrails in place (e.g. hooks or a post-commit safety net). Finally, you need to capture and version-control the evidence and implement clear hand-offs to human reviewers. I will provide examples during my talk.

What are the biggest challenges in that evidence-based workflow?

The key challenge is the subjective interpretation of natural language inputs, and its translation to an unambiguous specification and vice versa. One example is the translation of prose requirements to executable tests. AI can fill that gap, but this is precisely where humans must validate the interpretation. On the other hand, it is not a challenge to ensure complete end-to-end traceability from requirements to code, the robustness of the implementation, compliance to standards, and meeting all measurable requirements. There are trusted tools and platforms like Polyspace that can be integrated directly into the agentic loop and provide corrective signals as well as end-to-end traceability.

Can AI-assisted workflows improve software quality - not just speed?

In my experience, yes. AI can make rigorous practices more accessible by automating test generation, enabling continuous verification loops, and increasing coverage. When integrated into structured workflows with strong guardrails, AI can help apply verification processes more consistently and documents implicit information that is often lost in human-driven workflows.

One example is an AI-assisted workflow equipped with static code analysis. It can iterate until defined quality criteria are met and the risk of run-time failures is reduced or eliminated. This goes beyond the typical human approach of “best effort”.

What does success look like for teams adopting an evidence-based approach?

I am optimistic: Success means that every code change is backed by hard engineering evidence —tests that pass, requirements that are met, static analysis results that are “green”, and traceability between all artifacts. Successful teams encode this into their process and apply it consistently across their tools, including AI. They use trustworthy tools to automate verification, so that team members can focus on higher-level validation work. They spend significantly less time inspecting code, understanding the decisions that were made and the assumptions behind them, and remain accountable for the outcome. Currently, only few teams are operating at this level, and it remains to be seen whether “validation” becomes the new bottleneck.